betaFeedback

WorksMathematics

Iterated Steiner symmetrization of convex bodies always converges in shape

Abstract

We prove that every sequence of iterated Steiner symmetrals of a convex body converges in shape. That is, every such sequence can be corrected by a sequence of rotations to achieve convergence in Hausdorff distance. Moreover, for a fixed corresponding family of subspaces, the same rotations can be used for every initial convex body, and their successive increments can be chosen to be square-summable.

Introduction

The story of Steiner symmetrization started almost two centuries ago with the work of Jakob Steiner on the isoperimetric inequality for convex bodies (that is, non-empty compact and convex subsets of Rn\mathbb{R}^{n}), based on the following remarkable property: Steiner symmetrization preserves the Lebesgue measure of a convex body while decreasing its surface area. Moreover, given a convex body K⊂RnK \subset\mathbb{R}^{n}, it is always possible to find a sequence of hyperplanes such that the corresponding iterated Steiner symmetrals converge to a centred Euclidean ball in the Hausdorff metric. For more details, we refer the reader to the exposition by Bianchi and Gronchi [5].

The first general result in Rn\mathbb{R}^{n} is attributed to Gross [9], who proved that for every convex body KK there exists a symmetrization process converging to a centred Euclidean ball. Decades later, Mani-Levitska [11] proved a much stronger statement: Choosing the sequence of hyperplanes independently and uniformly at random, and applying the corresponding symmetrization process to a convex body with non-empty interior, the corresponding sequence of symmetrals will converge almost surely to a centred Euclidean ball. See Burchard and Fortier [7] and Coupier and Davydov [8] for further results on convergence of random Steiner symmetrizations. Even if these results do not leave too much space to non-converging symmetrization processes, a counterexample was eventually found in the work of Bianchi, Burchard, Gronchi, and Volčič [1]. In the same work, the notion of convergence in shape was introduced in order to fix the counterexample: A sequence (Lm)(L_m) of convex bodies is said to converge in shape if there exists a sequence of rotations (Rm)⊆SO(n)(R_m) \subseteq SO(n) such that the adjusted sequence (RmLm)(R_mL_m) converges in the Hausdorff metric. In [1 Theorem 2.2], it is then proved that a class of Steiner symmetrization processes which includes the aforementioned counterexample does indeed always converge in shape. In that case, the successive angles between the symmetrizing directions are square-summable, and the same rotations work for all initial compact sets.

It is natural to ask whether every sequence of Steiner symmetrals converges in shape, a problem posed in Question 6.2 of [1]. In this paper, we give a positive answer in the class of convex bodies. Additionally, inspired by a conjecture of Bianchi and Gronchi [4], we show that

MSC 2020 Classification: Primary 52A20. Secondary 51F15, 52A30. every non-convergent sequence nevertheless admits correcting rotations with square-summable successive increments, a property shared by the counterexample [1]. In fact, for a given sequence of hyperplanes, these rotations can be chosen once for all initial convex bodies.

This work is completed by some consequences and extensions to more general symmetrizations, a framework introduced by Bianchi, Gardner, and Gronchi [2, 3]. In particular, our results extend to Minkowski symmetrization which, in turn, allows us to analogously treat compact subsets of Rn\mathbb{R}^{n}, following a procedure from [14].

Main result. We start by introducing some notation. We denote by Kn\mathcal{K}^{n} the family of convex bodies in Rn\mathbb{R}^{n} and by Grn,n−1\mathrm{Gr}_{n,n-1} the linear hyperplanes in Rn\mathbb{R}^{n}. For H∈Grn,n−1H \in\mathrm{Gr}_{n,n-1} and K∈KnK \in\mathcal{K}^{n} the corresponding Steiner symmetral SHKS_{H}K (see § 2 for the precise definition of Steiner symmetrization). For a sequence (Hm)⊂Grn,n−1(H_{m}) \subset\mathrm{Gr}_{n,n-1}, the corresponding Steiner symmetrization process produces, for every K0∈KnK_{0} \in\mathcal{K}^{n}, the iterated sequence of symmetrals

Km=SHmKm−1,m∈N.K_{m}=S_{H_{m}}K_{m-1}, \qquad m \in\mathbb{N}.

Convergence of convex bodies is always understood with respect to the Hausdorff distance, denoted by dHd_{H}. Moreover, we denote by dSO(n)d_{SO(n)} the usual distance on SO(n)SO(n) induced by its standard Riemannian structure (see § 3 for the details). Our main result reads as follows.

Theorem 1.1. Consider a sequence of hyperplanes (Hm)⊂Grn,n−1(H_{m}) \subset\mathrm{Gr}_{n,n-1}. Then there exists a sequence of rotations (Rm)⊂SO(n)(R_{m}) \subset SO(n) such that, for every K0∈KnK_{0} \in\mathcal{K}^{n}, the sequence of symmetrals Km=SHmKm−1K_{m}=S_{H_{m}}K_{m-1}, m∈Nm \in\mathbb{N}, satisfies

lim⁡m→∞dH(RmKm,L)=0\lim_{m\to\infty} d_{H}(R_{m}K_{m},L)=0

for some L∈KnL \in\mathcal{K}^{n}. The rotations can be chosen with square-summable successive increments. That is,

∑m∈NdSO(n)(Rm,Rm+1)2<∞.\sum_{m\in\mathbb{N}}d_{SO(n)}(R_{m},R_{m+1})^{2}<\infty.

Additionally, if (Km)(K_{m}) does not converge for some K0∈KnK_{0} \in\mathcal{K}^{n}, then these rotations satisfy

∑m∈NdSO(n)(Rm,Rm+1)=∞,∑m∈NdSO(n)(Rm,Rm+1)2<∞.(1)\sum_{m\in\mathbb{N}}d_{SO(n)}(R_{m},R_{m+1})=\infty,\qquad \sum_{m\in\mathbb{N}}d_{SO(n)}(R_{m},R_{m+1})^{2}<\infty. \tag*{(1)}

Notice that if (Km)(K_{m}) converges for every K0∈KnK_{0} \in\mathcal{K}^{n}, one can trivially choose Rm=Id⁡R_{m}=\operatorname{Id}, the identity matrix, for every m∈Nm \in\mathbb{N}. Therefore, Theorem 1.1 can be understood as the following dichotomy: Either the symmetrization process converges for every initial body, or all its sequences of symmetrals converge in shape with the same rotations (Rm)(R_{m}) satisfying (1). While the proof of convergence in shape for a single body uses basic theory of convex bodies, the summability statement and the choice of a common sequence of rotations draw on additional elementary tools from the theory of Lie groups and spherical harmonics. For this reason, we correspondingly split the proof of Theorem 1.1 between § 2 and § 3.

Structure of the paper. In § 2, we introduce further notation and definitions, upon which we build the proof of convergence in shape for a single body. Moreover, we show how (1) adapts to the case of ellipses in the plane, providing a preliminary intuition of the convergence mechanism. § 3 is dedicated to proving that condition in complete generality and to choosing the rotations independently of the initial body. Finally, in § 4, we show how the core constructions apply to the more general framework of symmetrizations introduced in [2, 3].

Convergence in shape

We recall that by Kn\mathcal{K}^n we denote the family of non-empty compact and convex subsets of Rn\mathbb{R}^n, known as convex bodies. More generally, Cn\mathcal{C}^n denotes the non-empty compact subsets of Rn\mathbb{R}^n. For K∈KnK \in\mathcal{K}^n, its support function is hK(u)=sup⁡x∈K(x⋅u)h_K(u) = \sup_{x \in K}(x \cdot u), u∈Sn−1u \in S^{n-1}, where Sn−1S^{n-1} is the Euclidean unit sphere of Rn\mathbb{R}^n and x⋅yx \cdot y is the standard scalar product in Rn\mathbb{R}^n. Both Kn\mathcal{K}^n and Cn\mathcal{C}^n are endowed with the topology induced by the Hausdorff metric. See Schneider’s monograph [12] for further details on the theory of convex bodies.

Consider H∈Gr⁡n,n−1H \in\operatorname{Gr}_{n,n-1} and K∈KnK \in\mathcal{K}^n. The Steiner symmetral SHKS_HK is obtained by replacing every non-empty interval K∩(x+H⊥)K \cap(x + H^\perp), x∈Hx \in H, by the (possibly degenerate) interval of the same length centred at xx and orthogonal to HH. The resulting set is again compact and convex. We only use Steiner symmetrization on Kn\mathcal{K}^n below. The final compact-set extension concerns Minkowski symmetrization.

For C,D∈CnC,D \in\mathcal{C}^n, their Minkowski sum is the set C+D={x+y:x∈C, y∈D}C + D = \{x + y : x \in C,\ y \in D\}. Moreover, for H∈Gr⁡n,n−1H \in\operatorname{Gr}_{n,n-1} we denote by ρH\rho_H the linear reflection in Rn\mathbb{R}^n with respect to HH. A subset of Rn\mathbb{R}^n is HH-symmetric if it is invariant for ρH\rho_H. For C∈CnC \in\mathcal{C}^n and H∈Gr⁡n,n−1H \in\operatorname{Gr}_{n,n-1}, the Minkowski symmetral of CC with respect to HH is the compact set MHC=(C+ρHC)/2M_HC = (C + \rho_HC)/2. If, additionally, CC is convex, then so is MHCM_HC. When restricted to Kn\mathcal{K}^n, Steiner and Minkowski symmetrizations satisfy the inclusion

SHK⊆MHKfor every K∈Kn, H∈Gr⁡n,n−1.(2)S_HK \subseteq M_HK \qquad\text{for every } K \in\mathcal{K}^n,\ H \in\operatorname{Gr}_{n,n-1}. \tag*{(2)}

See, for example, [8 Lemma 2.1]. In Section 4, we will explain how the generalization of this inclusion in the framework of [2, 3] allows us to extend our results.

The aim of this section is to prove convergence in shape for a fixed initial body, which we state here for the convenience of the reader.

Theorem 2.1. Consider K∈KnK \in\mathcal{K}^n, a sequence (Hm)⊂Gr⁡n,n−1(H_m) \subset\operatorname{Gr}_{n,n-1}, and the corresponding sequence of Steiner symmetrals

K0=K,Km=SHmKm−1,m∈N.K_0 = K,\qquad K_m = S_{H_m}K_{m-1},\qquad m \in\mathbb{N}.

Then, there exists a sequence of rotations (Rm)⊂SO(n)(R_m) \subset SO(n) and L∈KnL \in\mathcal{K}^n (both allowed to depend on KK and (Hm)(H_m)) such that

lim⁡m→∞dH(RmKm,L)=0.\lim_{m \to\infty} d_H(R_mK_m,L) = 0.

The result above has the following immediate consequence.

Corollary 2.2. All subsequential limits of KmK_m are the same up to a rotation in SO(n)SO(n).

Proof of Theorem 2.1

We start with some preliminary observations. First, we recall that given K0∈KnK_0 \in\mathcal{K}^n and (Hm)⊂Gr⁡n,n−1(H_m) \subset\operatorname{Gr}_{n,n-1}, choosing a>0a > 0 large enough so that K0⊆aBnK_0 \subseteq aB^n, where BnB^n is the Euclidean unit ball, the sets Km=SHmKm−1K_m = S_{H_m}K_{m-1} satisfy Km⊆aBnK_m \subseteq aB^n for every m∈Nm \in\mathbb{N} since Steiner symmetrization preserves inclusions. For the remainder of this subsection the set K0K_0, the sequence (Hm)(H_m), and the corresponding symmetrization process will be fixed.

In terms of support functions, writing fm=hKm+af_m = h_{K_m} + a for m≥0m \geq0, we infer

0≤fm≤2a0 \leq f_m \leq2a

for every m≥0.Denotingbym \geq0. Denoting byO(n)$ the orthogonal linear group of Rn\mathbb{R}^n, for Q∈O(n)Q \in O(n) and f∈C(Sn−1)f \in C(S^{n-1}) we write UQf(u)=f(Q−1u)U_Qf(u) = f(Q^{-1}u). Consider now the transformations

Tm=12(Id⁡+UρHm).(3)T_m = \frac{1}{2}(\operatorname{Id} + U_{\rho_{H_m}}). \tag*{(3)}

They are self-adjoint orthogonal projections on L2(Sn−1)L^2(\mathbb{S}^{n-1}) and hence 1-Lipschitz. Moreover, (2) implies

0≤fm≤Tmfm−1.(4)0 \le f_m \le T_m f_{m-1}. \tag*{(4)}

Indeed, by the definition of Minkowski symmetrization,

Tmfm−1=hMHmKm−1+a.T_m f_{m-1}=h_{M_{H_m}K_{m-1}}+a.

Thus (3) follows from (2). Denote by ⟨⋅,⋅⟩2\langle\cdot,\cdot\rangle_2 the standard scalar product on L2(Sn−1)L^2(\mathbb{S}^{n-1}) and by ∥⋅∥2\|\cdot\|_2 the corresponding norm. From (3) and since TmT_m is 1-Lipschitz, we infer

0≤∥fm∥2≤∥fm−1∥2.0\leq\|f_m\|_2\leq\|f_{m-1}\|_2.

In particular, the sequence (∥fm∥2)(\|f_m\|_2) converges to a value s≥0s\geq0 since it is non-increasing.

Next, we introduce the following pseudo-distance, which will be the key ingredient of the proof of Theorem 2.1. For K,K′∈KnK,K'\in\mathcal{K}^n, consider

δ2(K,K′)=min⁡R∈SO(n)∥hRK−hK′∥2.\delta_2(K,K')=\min_{R\in SO(n)}\|h_{RK}-h_{K'}\|_2.

It is well defined since R↦∥hRK−hK′∥2R\mapsto\|h_{RK}-h_{K'}\|_2 is easily checked to be continuous with respect to the standard topology of SO(n)SO(n) (see § 3 for more details) and SO(n)SO(n) is compact. In particular, ∅≠argmin⁡{R↦∥hRK−hK′∥2}\varnothing\ne\operatorname{argmin}\{R\mapsto\|h_{RK}-h_{K'}\|_2\}. For the sequence of symmetrals (Km)(K_m), δ2\delta_2 satisfies the following property.

Lemma 2.3. Given the sequences (Km)⊂Kn(K_m)\subset\mathcal{K}^n and (fm)⊂C(Sn−1)(f_m)\subset C(\mathbb{S}^{n-1}) as above, for every $j>i\geq 1

δ2(Ki,Kj)2≤∥fi∥22−∥fj∥22.(5)\delta_2(K_i,K_j)^2\leq\|f_i\|_2^2-\|f_j\|_2^2. \tag*{(5)}

Proof. Fix j>i≥1j>i\geq1. Repeated use of (3) and preservation of pointwise order gives

0≤fj≤Tjfj−1≤Tj⋯Ti+1fi=g.0\leq f_j\leq T_jf_{j-1}\leq T_j\cdots T_{i+1}f_i=g.

Expanding the operators, we can write

g=∑k=12j−i12j−iUQkfi,g=\sum_{k=1}^{2^{j-i}}\frac{1}{2^{j-i}}U_{Q_k}f_i,

where the Qk∈O(n)Q_k\in O(n) are compositions of reflections. Hence

∑k=12j−i12j−i∥UQkfi−fj∥22=∥fi∥22+∥fj∥22−2⟨g,fj⟩2=∥fi∥22−∥fj∥22−2⟨g−fj,fj⟩2≤∥fi∥22−∥fj∥22.\begin{aligned} \sum_{k=1}^{2^{j-i}}\frac{1}{2^{j-i}}\|U_{Q_k}f_i-f_j\|_2^2 =\|f_i\|_2^2+\|f_j\|_2^2-2\langle g,f_j\rangle_2 \\ &=\|f_i\|_2^2-\|f_j\|_2^2-2\langle g-f_j,f_j\rangle_2 \\ &\leq\|f_i\|_2^2-\|f_j\|_2^2. \end{aligned}

Since this is an average of nonnegative numbers, at least one term satisfies the same upper bound. Thus, for some k0k_0,

∥UQk0fi−fj∥22≤∥fi∥22−∥fj∥22.\|U_{Q_{k_0}}f_i-f_j\|_2^2\leq\|f_i\|_2^2-\|f_j\|_2^2.

If det⁡Qk0=−1\det Q_{k_0}=-1, replace Qk0Q_{k_0} by Qk0ρHi∈SO(n)Q_{k_0}\rho_{H_i}\in SO(n). This does not change Qk0KiQ_{k_0}K_i because KiK_i is HiH_i-symmetric. The last display is therefore an admissible competitor for δ2(Ki,Kj)\delta_2(K_i,K_j) and proves (4). □

:::

Next, observe that for every A,A′,K,K′∈KnA,A',K,K' \in\mathcal{K}^n

∣δ2(A,A′)−δ2(K,K′)∣≤∥hA−hK∥2+∥hA′−hK′∥2.(6)\left|\delta_2(A,A')-\delta_2(K,K')\right| \leq\lVert h_A-h_K\rVert_2+\lVert h_{A'}-h_{K'}\rVert_2. \tag*{(6)}

Indeed, notice that for R∈SO(n)R \in SO(n)

∥hRA−hA′∥2≤∥hRA−hRK∥2+∥hRK−hK′∥2+∥hA′−hK′∥2.\lVert h_{RA}-h_{A'}\rVert_2 \leq\lVert h_{RA}-h_{RK}\rVert_2+\lVert h_{RK}-h_{K'}\rVert_2+\lVert h_{A'}-h_{K'}\rVert_2.

In particular,

δ2(A,A′)≤δ2(K,K′)+∥hA−hK∥2+∥hA′−hK′∥2,\delta_2(A,A') \leq\delta_2(K,K')+\lVert h_A-h_K\rVert_2+\lVert h_{A'}-h_{K'}\rVert_2,

and (5) is proved inverting the role of A,A′A,A' and K,K′K,K'.

To finish the proof of Theorem 2.1, by the Blaschke selection theorem [12 Theorem 1.8.7] there exists a subsequence (Kmj)(K_{m_j}) converging to some L∈KnL \in\mathcal{K}^n in the Hausdorff metric. For every m∈Nm \in\mathbb{N} apply Lemma 2.3 to the indices mj>mm_j>m and let j→∞j \to\infty to obtain

δ2(Km,L)2≤∥fm∥22−s2.\delta_2(K_m,L)^2 \leq\lVert f_m\rVert_2^2-s^2.

The right-hand side tends to zero. Now, for every mm, choose a rotation RmR_m minimizing δ2(Km,L)\delta_2(K_m,L). Then hRmKm→hLh_{R_mK_m} \to h_L in L2(Sn−1)L^2(\mathbb{S}^{n-1}). Since all these convex bodies lie in the fixed ball aBnaB^n, convergence in the L2L^2 support-function metric is equivalent to Hausdorff convergence, see [12 Notes for Section 1.8, Note 9]. Hence RmKm→LR_mK_m \to L.

A 2-dimensional example: Behaviour on ellipses

Before completing the proof of Theorem 1.1 in the next section, we present an easier proof restricted to centred ellipses in R2\mathbb{R}^2 to give a geometric intuition of (1), which we are going to prove in full generality in the next section. We recall that for every 1-dimensional subspace HH, the set SHES_HE is still an ellipse, and its axes are aligned along HH and H⊥H^\perp.

Theorem 2.4. Consider a sequence (Hm)⊂Gr⁡2,1(H_m) \subset\operatorname{Gr}_{2,1}. Then, for every centred ellipse E0⊂R2E_0 \subset\mathbb{R}^2 with semiaxes of lengths a0>b0>0a_0>b_0>0, the sequence of symmetrals

Em=SHmEm−1E_m=S_{H_m}E_{m-1}

has necessarily one of the following mutually exclusive behaviours:

  • The sequence EmE_m converges to some ellipse EE.

  • The sequence EmE_m does not converge, but converges in shape to some ellipse EE and, denoting by αm∈[0,π/2]\alpha_m \in[0,\pi/2] the smaller angle between the major axes of EmE_m and Em−1E_{m-1},

∑m∈Nαm=∞,∑m∈Nαm2<∞.\sum_{m\in\mathbb{N}}\alpha_m=\infty,\qquad\sum_{m\in\mathbb{N}}\alpha_m^2<\infty.

Proof. If EmE_m converges, then we are in the first case. Suppose otherwise. Theorem 2.1 implies convergence in shape. Write λm=am2\lambda_m=a_m^2 and μm=bm2\mu_m=b_m^2, where am≥bma_m\geq b_m are the semiaxis lengths. We represent a centred ellipse by a positive definite matrix AA through {x∈R2:xTA−1x≤1}\{x\in\mathbb{R}^2:x^{T}A^{-1}x\leq1\}. In coordinates in which HH is the first axis, an ellipse represented by A=(abbc)A=\begin{pmatrix}a&b\\b&c\end{pmatrix} is sent by SHS_H to the ellipse represented by diag⁡(d,det⁡A/d)\operatorname{diag}(d,\det A/d). Consequently, λm\lambda_m is nonincreasing, μm\mu_m is nondecreasing, and λmμm=λ0μ0\lambda_m\mu_m=\lambda_0\mu_0. Hence λm→λ∞\lambda_m\to\lambda_\infty and μm→μ∞\mu_m\to\mu_\infty, with λ∞≥μ∞\lambda_\infty\geq\mu_\infty. The inequality is strict, since otherwise μmB2⊆Em⊆λmB2\sqrt{\mu_m}B^2\subseteq E_m\subseteq\sqrt{\lambda_m}B^2 would force convergence to a disk. No symmetral $E_m can itself be a disk, since all subsequent symmetrals would then be the same. Thus, all the major axes are well defined. Finally,

∑m∈Nαm=∞,\sum_{m\in\mathbb{N}} \alpha_m = \infty,

because a finite sum would make the sequence of major-axis lines a Cauchy sequence in the angular metric. Together with convergence of the semiaxes, this would imply convergence of EmE_m.

Now, observe that

λm−1−λm≥(λm−1−μm−1)sin⁡2αm≥(λ∞−μ∞)sin⁡2αm.(7)\lambda_{m-1}-\lambda_m \ge(\lambda_{m-1}-\mu_{m-1})\sin^2\alpha_m \ge(\lambda_\infty-\mu_\infty)\sin^2\alpha_m. \tag*{(7)}

Indeed, for two successive symmetrals Em−1E_{m-1} and EmE_m, the major axis shifts in one of the two following ways: Either it aligns with HmH_m, in which case

λm=λm−1−(λm−1−μm−1)sin⁡2αm,\lambda_m=\lambda_{m-1}-(\lambda_{m-1}-\mu_{m-1})\sin^2\alpha_m,

or with Hm⊥H_m^\perp, and

λm=λm−1μm−1μm−1+(λm−1−μm−1)sin⁡2αm.\lambda_m=\frac{\lambda_{m-1}\mu_{m-1}}{\mu_{m-1}+(\lambda_{m-1}-\mu_{m-1})\sin^2\alpha_m}.

In both cases, we obtain (6). To conclude, the elementary inequality sin⁡α≥2α/π\sin\alpha\ge2\alpha/\pi for α∈[0,π/2]\alpha\in[0,\pi/2], together with (6), gives

αm2≤π24(λ∞−μ∞)(λm−1−λm).\alpha_m^2\le\frac{\pi^2}{4(\lambda_\infty-\mu_\infty)}(\lambda_{m-1}-\lambda_m).

Thus,

∑m∈Nαm2≤π24(λ∞−μ∞)∑m∈N(λm−1−λm)=π24(λ∞−μ∞)(λ0−λ∞)<∞.\sum_{m\in\mathbb{N}}\alpha_m^2\le\frac{\pi^2}{4(\lambda_\infty-\mu_\infty)}\sum_{m\in\mathbb{N}}(\lambda_{m-1}-\lambda_m)=\frac{\pi^2}{4(\lambda_\infty-\mu_\infty)}(\lambda_0-\lambda_\infty)<\infty.

Choosing the right rotations in Theorem 1.1

To obtain the full statement of Theorem 1.1, we first work with a single sequence of symmetrals. The main idea is the following: Working on the orbit of the possible limits of the sequence (Km)(K_m), one can iteratively choose the rotation Rm+1R_{m+1} from Theorem 2.1 in a suitably small neighbourhood of RmR_m inside SO(n)SO(n). To this end, we work on SO(n)SO(n) as a Lie group and a Riemannian manifold. We start by recalling the corresponding metric. In this section, for ease of notation we write G=SO(n)G=SO(n). Then, on GG we consider its standard Riemannian metric, which, at the identity, reads as

⟨X,Y⟩g=12tr⁡(XTY),X,Y∈g=so(n).\langle X,Y\rangle_g=\frac{1}{2}\operatorname{tr}(X^{T}Y),\qquad X,Y\in\mathfrak{g}=\mathfrak{so}(n).

Here, so(n)\mathfrak{so}(n) is the Lie algebra of SO(n)SO(n). See, for example, [6] for this and other basic definitions. As it is standard, the metric on the whole GG is defined by left translates. Recall that both left and right translates are isometries for this metric. In particular, for every A,B,Q∈SO(n)A,B,Q\in SO(n)

dG(AQ,BQ)=dG(QA,QB)=dG(A,B),dG(QAQ−1,I)=dG(A,I).(8)d_G(AQ,BQ)=d_G(QA,QB)=d_G(A,B),\qquad d_G(QAQ^{-1},I)=d_G(A,I). \tag*{(8)}

We will need the following lemma. Given a real finite-dimensional inner-product space VV, O(V)O(V) denotes its orthogonal group and ∥⋅∥V\lVert\cdot\rVert_V its norm.

Lemma 3.1. Consider an orthogonal smooth representation of GG on VV. That is, we are given a smooth group homomorphism

U:G⟶O(V)\mathcal{U}: G \longrightarrow O(V)
Q⟼UQ.Q \longmapsto\mathcal{U}_Q.

For a fixed v∈Vv \in V consider its orbit

M={UQv:Q∈G}.M = \{\mathcal{U}_Qv: Q \in G\}.

Then MM is a smooth embedded submanifold of VV. Moreover, there exist η>0\eta> 0, Cη>0C_\eta> 0, such that for every y,z∈My,z \in M with ∥y−z∥V<η\lVert y-z\rVert_V < \eta there exists U∈GU \in G satisfying

UUy=zanddG(U,I)≤Cη∥y−z∥V.\mathcal{U}_U y = z \qquad\text{and} \qquad d_G(U,I) \le C_\eta\lVert y-z\rVert_V.

Proof. First, suppose that M={v}M = \{v\}. Then the statement is trivial. Therefore, in the remainder of the proof we suppose that MM is not a singleton. Since GG is compact, its smooth action on VV is proper. See, e.g., [10 Corollary 21.6]. As orbits of proper Lie group actions are embedded submanifolds, so is MM, which is compact since it is the continuous image of GG. This proves the first claim.

To prove the second part observe that the orbit map

p:G⟶Mp: G \longrightarrow M
Q⟼UQvQ \longmapsto\mathcal{U}_Qv

is a smooth submersion. By the local section theorem [10 Theorem 4.26] there is a smooth local right inverse qq of pp with q(v)=Id⁡q(v) = \operatorname{Id}. Then, upon choosing η>0\eta> 0 sufficiently small, there exists a constant Cη>0C_\eta> 0 such that for every w∈Mw \in M with ∥v−w∥V<η\lVert v-w\rVert_V < \eta holds

dG(q(w),Id⁡)≤Cη∥w−v∥V.d_G(q(w),\operatorname{Id}) \le C_\eta\lVert w-v\rVert_V.

The conclusion now follows by a standard translation argument. For y,z∈My,z \in M with ∥y−z∥V<η\lVert y-z\rVert_V < \eta, choose Q∈GQ \in G with UQv=y\mathcal{U}_Qv = y and set w=UQ−1zw = \mathcal{U}_{Q^{-1}}z. Then U=Qq(w)Q−1U = Qq(w)Q^{-1} satisfies

UUy=z,dG(U,Id⁡)=dG(q(w),Id⁡)≤Cη∥y−z∥V,\mathcal{U}_U y = z,\qquad d_G(U,\operatorname{Id}) = d_G(q(w),\operatorname{Id}) \le C_\eta\lVert y-z\rVert_V,

where bi-invariance of dGd_G and orthogonality of the representation were used. □\square

Square-summable convergence in shape

We now prove that the rotations in Theorem 2.1 can be chosen to have successive square-summable distance in GG. This behaviour already appeared in [1], and Bianchi and Gronchi [5] suggested the author that this phenomenon should extend to all sequences of iterated Steiner symmetrals of convex bodies.

Theorem 3.2. Consider (Km)(K_m) as in Theorem 1.1. If (Km)(K_m) does not converge, then for any subsequential limit LL of (Km)(K_m) the rotations (Rm)⊂G(R_m) \subset G giving RmKm→LR_mK_m \to L can be chosen so that

∑m∈NdG(Rm,Rm−1)2<∞.(9)\sum_{m\in\mathbb{N}} d_G(R_m,R_{m-1})^2 < \infty. \tag*{(9)}

Proof. Retain the notation aa, fmf_m, TmT_m, ss from § 2, and fix a limit LL as in the proof of Theorem 2.1. We start by refining the main argument in the proof of Lemma 2.3. Since TmT_m is self-adjoint and Tmfm=fmT_m f_m=f_m,

⟨fm−1,fm⟩22=⟨Tmfm−1,fm⟩22≥∥fm∥22.\langle f_{m-1},f_m\rangle_2^2=\langle T_mf_{m-1},f_m\rangle_2^2\ge\lVert f_m\rVert_2^2.

Consequently,

∥fm−fm−1∥22≤∥fm−1∥22−∥fm∥22.\lVert f_m-f_{m-1}\rVert_2^2\le\lVert f_{m-1}\rVert_2^2-\lVert f_m\rVert_2^2.

and therefore

∑m∈N∥fm−fm−1∥22≤∥f0∥22−s2<∞.(10)\sum_{m\in\mathbb{N}}\lVert f_m-f_{m-1}\rVert_2^2\le\lVert f_0\rVert_2^2-s^2<\infty. \tag*{(10)}

Thus, ∥fm−fm−1∥2→0\lVert f_m-f_{m-1}\rVert_2\to0. Write f=hL+af=h_L+a and consider the compact orbit

O={UQf:Q∈G}⊂L2(Sn−1).\mathcal{O}=\{U_Qf:Q\in G\}\subset L^2(S^{n-1}).

By the considerations above, dL2(fm,O)2=δ2(Km,L)2→0d_{L^2}(f_m,\mathcal{O})^2=\delta_2(K_m,L)^2\to0. Although O\mathcal{O} is homeomorphic to the finite-dimensional homogeneous space G/Stab⁡G(f)G/\operatorname{Stab}_G(f), its linear span may be infinite-dimensional. In order to be able to apply Lemma 3.1, we now get to a reduction argument.

For j∈N∪{0}j\in\mathbb{N}\cup\{0\} let Hjn\mathcal{H}_j^n be the space of spherical harmonics of degree jj and set

VN=⨁j=0NHjn,PN:L2(Sn−1)→VN.V_N=\bigoplus_{j=0}^{N}\mathcal{H}_j^n,\qquad P_N:L^2(S^{n-1})\to V_N.

These orthogonal projections commute with rotations and satisfy PNf→fP_Nf\to f in L2L^2. Define

GN={Q∈G:UQPNf=PNf}.G_N=\{Q\in G:U_QP_Nf=P_Nf\}.

Each GNG_N is a closed subgroup of GG. Since the projections commute with rotations and PNPN+1=PNP_NP_{N+1}=P_N, these groups form a decreasing sequence. The descending-chain theorem for closed subgroups of a compact Lie group [6 p. 136] gives N0N_0 such that GN=GN0G_N=G_{N_0} for every N≥N0N\ge N_0. Moreover,

⋂N=0∞GN=Stab⁡G(f),\bigcap_{N=0}^{\infty}G_N=\operatorname{Stab}_G(f),

because for QQ in the intersection, ∥UQf−f∥2≤2∥f−PNf∥2→0\lVert U_Qf-f\rVert_2\le2\lVert f-P_Nf\rVert_2\to0. Therefore

GN0=Stab⁡G(f).(11)G_{N_0}=\operatorname{Stab}_G(f). \tag*{(11)}

Write P=PN0P=P_{N_0}, V=VN0V=V_{N_0}, and v=Pfv=Pf. Equality of stabilizers implies that PP is injective on the orbit O\mathcal{O}. Let

M={UQv:Q∈G}⊂V,xm=Pfm.M=\{U_Qv:Q\in G\}\subset V,\qquad x_m=Pf_m.

Then dV(xm,M)→0d_V(x_m,M)\to0 and ∥xm−xm−1∥V≤∥fm−fm−1∥2\lVert x_m-x_{m-1}\rVert_V\le\lVert f_m-f_{m-1}\rVert_2.

If MM is a singleton, injectivity of P∣OP|_{\mathcal{O}} makes O\mathcal{O} a singleton, and dL2(fm,O)→0d_{L^2}(f_m,\mathcal{O})\to0 would imply convergence of (Km)(K_m), contrary to the hypothesis. Thus MM is not a singleton. By Lemma 3.1, MM is a compact embedded submanifold of VV. By the tubular neighbourhood theorem [10 Theorem 6.24], nearest-point projection π\pi onto MM is smooth on a sufficiently small tube, and its differential is bounded on a smaller closed tube by compactness. Since dV(xm,M)→0d_V(x_m,M)\to0 and ∥xm−xm−1∥V→0\lVert x_m-x_{m-1}\rVert_V\to0, the segment from xm−1x_{m-1} to xmx_m lies in the smaller tube for all sufficiently large mm. In particular, the points ym=π(xm)y_m=\pi(x_m) satisfy

∥ym−ym−1∥V≤Cπ∥xm−xm−1∥V.\lVert y_m-y_{m-1}\rVert_V\le C_\pi\lVert x_m-x_{m-1}\rVert_V.

Choose m0m_0 sufficiently large that Lemma 3.1 gives Um∈GU_m \in G for m>m0m > m_0 such that

UUmym−1=ym,dG(Um,Id⁡)≤Cη∥ym−ym−1∥V.\mathcal{U}U_m y_{m-1}=y_m,\qquad d_G(U_m,\operatorname{Id})\le C_\eta\lVert y_m-y_{m-1}\rVert_V.

Additionally, consider Qm0∈GQ_{m_0}\in G with UQm0v=ym0\mathcal{U}Q_{m_0}v=y_{m_0}. Set Rm=Qm0−1R_m=Q_{m_0}^{-1} for 0≤m≤m00\le m\le m_0, and for m>m0m>m_0 define

Qm=UmQm−1,Rm=Qm−1.Q_m=U_mQ_{m-1},\qquad R_m=Q_m^{-1}.

Then URmym=v\mathcal{U}R_my_m=v, and

dG(Rm,Rm−1)2=dG(Um,Id⁡)2≤Cη2∥ym−ym−1∥V2≤Cη2Cπ2∥xm−xm−1∥V2≤Cη2Cπ2∥fm−fm−1∥22≤Cη2Cπ2(∥fm−1∥22−∥fm∥22).\begin{aligned} d_G(R_m,R_{m-1})^2 &=d_G(U_m,\operatorname{Id})^2 \le C_\eta^2\lVert y_m-y_{m-1}\rVert_V^2\\ &\le C_\eta^2C_\pi^2\lVert x_m-x_{m-1}\rVert_V^2\\ &\le C_\eta^2C_\pi^2\lVert f_m-f_{m-1}\rVert_2^2\\ &\le C_\eta^2C_\pi^2\left(\lVert f_{m-1}\rVert_2^2-\lVert f_m\rVert_2^2\right). \end{aligned}

Thus, (9) follows by (10).

Finally, we show that RmKm→LR_mK_m\to L. Put gm=URmfmg_m=\mathcal{U}R_mf_m. Since PP commutes with rotations,

Pgm−v=URm(xm−ym)⟶0.(12)Pg_m-v=\mathcal{U}R_m(x_m-y_m)\longrightarrow0. \tag*{(12)}

Also,

dL2(gm,O)=dL2(fm,O)⟶0.d_{L^2}(g_m,\mathcal{O})=d_{L^2}(f_m,\mathcal{O})\longrightarrow0.

Since O\mathcal{O} is compact, every subsequence of (gm)(g_m) has a further subsequence converging in L2L^2 to some g∈Og\in\mathcal{O}, and (11) implies Pg=PfPg=Pf. Since P∣OP|_{\mathcal{O}} is injective, g=fg=f. Hence gm→fg_m\to f in L2L^2.

Since

gm=hRmKm+a,f=hL+a,g_m=h_{R_mK_m}+a,\qquad f=h_L+a,

the equivalence of the L2L^2 support-function and Hausdorff topologies on uniformly bounded convex bodies [12 Notes for Section 1.8, Note 9] yields RmKm→LR_mK_m\to L.

Remark 3.3. We use a dimension reduction from O\mathcal{O} to MM, since even if O\mathcal{O} is finite dimensional, its linear span need not be. E.g. support functions of noncircular ellipses have infinitely many Fourier coefficients, so their rotation orbits have infinite-dimensional linear span.

Finding a common sequence of rotations

We are left to show that the rotations in Theorem 1.1 can be chosen independently of the initial body. We start by considering finitely many initial bodies and applying the preceding arguments to their symmetrals simultaneously. For a tuple K=(K1,…,Kr)\mathbf{K}=(K_1,\ldots,K_r) and Q∈GQ\in G, we write QK=(QK1,…,QKr)Q\mathbf{K}=(QK_1,\ldots,QK_r). Convergence of tuples is understood in the Hausdorff metric in each component.

Lemma 3.4. Consider a sequence (Hm)⊂Grn,n−1(H_m)\subset\mathrm{Gr}_{n,n-1} and a finite family K0(1),…,K0(r)∈KnK_0^{(1)},\ldots,K_0^{(r)}\in\mathcal{K}^n, with r≥1r\ge1. For each ℓ=1,…,r\ell=1,\ldots,r let Km(ℓ)=SHmKm−1(ℓ)K_m^{(\ell)}=S_{H_m}K_{m-1}^{(\ell)}, and write Km=(Km(1),…,Km(r))\mathbf{K}_m=(K_m^{(1)},\ldots,K_m^{(r)}). Then any two subsequential limits of Km\mathbf{K}_m differ by a common rotation. Moreover, for each subsequential limit LL there is a sequence (Rm)⊂G(R_m)\subset G satisfying (8) such that RmKm→LR_m\mathbf{K}_m\to L.

Proof. Choose a>0a>0 such that K0(ℓ)⊆aBnK_0^{(\ell)}\subseteq aB^n for every ℓ\ell. In H=(L2(Sn−1))r\mathcal{H}=(L^2(\mathbb{S}^{n-1}))^r, endowed with the product scalar product, put

fm=(hKm(1)+a,…,hKm(r)+a),Em=∥fm∥H2.f_m=\left(h_{K_m^{(1)}}+a,\ldots,h_{K_m^{(r)}}+a\right),\qquad E_m=\lVert f_m\rVert_{\mathcal{H}}^2.

We let UQ\mathcal{U}_Q and TmT_m act in each component as in § 2. By (4), 0≤fm≤Tmfm−10 \le f_m \le T_m f_{m-1} componentwise, so EmE_m decreases to a limit E∞E_\infty. Notice that in the expansion of Tj⋯Ti+1T_j \cdots T_{i+1} the same compositions of reflections appear in each component. Moreover, if det⁡Q=−1\det Q=-1, then QρHi∈GQ\rho_{H_i} \in G and UQρHifi=UQfi\mathcal{U}_{Q\rho_{H_i}}f_i=\mathcal{U}_Qf_i, since each Ki(ℓ)K_i^{(\ell)} is HiH_i-symmetric. Thus the proof of Lemma 2.3 gives, for j>i≥1j>i\ge1,

min⁡Q∈G∥UQfi−fj∥H2≤Ei−Ej.\min_{Q\in G}\|\mathcal{U}_Qf_i-f_j\|_H^2\le E_i-E_j.

Now let L=(L1,…,Lr)L=(L_1,\ldots,L_r) be a subsequential limit and put f=(hL1+a,…,hLr+a)f=(h_{L_1}+a,\ldots,h_{L_r}+a). Passing to the corresponding subsequence in the preceding inequality gives

dH(fi,O)2≤Ei−E∞,O={UQf:Q∈G}.d_H(f_i,\mathcal{O})^2\le E_i-E_\infty,\qquad\mathcal{O}=\{\mathcal{U}_Qf:Q\in G\}.

In particular, since O\mathcal{O} is compact, every subsequential limit of fmf_m lies in O\mathcal{O}. The first assertion follows since support functions determine their bodies.

If KmK_m converges, we can take Rm=IdR_m=\mathrm{Id}. Otherwise, notice that Tmfm=fmT_mf_m=f_m. As in the proof of Theorem 3.2, we then obtain

∥fm−fm−1∥H2≤Em−1−Em,∑m∈N∥fm−fm−1∥H2≤E0−E∞.\|f_m-f_{m-1}\|_H^2\le E_{m-1}-E_m,\qquad\sum_{m\in\mathbb{N}}\|f_m-f_{m-1}\|_H^2\le E_0-E_\infty.

Apply PNP_N in every component. The stabilizers of the resulting projections of ff form a decreasing sequence with intersection Stab⁡G(f)\operatorname{Stab}_G(f). By the descending-chain theorem, the projection onto VNrV_N^r is therefore injective on O\mathcal{O} for some NN. We can now repeat the nearest-point projection and lifting argument of Theorem 3.2 in VNrV_N^r. It gives rotations satisfying (8) and URmfm→f\mathcal{U}_{R_m}f_m\to f in H\mathcal{H}. By the equivalence of the support-function and Hausdorff topologies on uniformly bounded convex bodies, RmKm→LR_mK_m\to L. □\square

We now choose a finite family for which adjoining another sequence of symmetrals cannot reduce the stabilizer of the limit tuple. For L=(L1,…,Lr)L=(L_1,\ldots,L_r) its stabilizer is Stab⁡G(L)=⋂ℓ=1rStab⁡G(Lℓ)\operatorname{Stab}_G(L)=\bigcap_{\ell=1}^r\operatorname{Stab}_G(L_\ell).

Proposition 3.5. For every sequence (Hm)⊂Grn,n−1(H_m)\subset G_{r n,n-1} there exists a sequence (Rm)⊂G(R_m)\subset G satisfying (8) such that, for every K0∈KnK_0\in\mathcal{K}^n, the symmetrals Km=SHmKm−1K_m=S_{H_m}K_{m-1} satisfy RmKm→LR_mK_m\to L in the Hausdorff metric for some L∈KnL\in\mathcal{K}^n.

Proof. Fix (Hm)(H_m) and consider the collection D\mathcal{D} of the stabilizers of all subsequential limits of tuples as in Lemma 3.4. Since these are closed subgroups of GG, the descending-chain theorem gives a minimal member DD under inclusion. Choose a tuple of symmetrals KmK_m and a subsequential limit LL with Stab⁡G(L)=D\operatorname{Stab}_G(L)=D. By Lemma 3.4, we can fix rotations (Rm)(R_m) satisfying (8) such that RmKm→LR_mK_m\to L.

Consider now any other initial body K0∈KnK_0\in\mathcal{K}^n and let (Km)(K_m) be its sequence of symmetrals with respect to the fixed family (Hm)(H_m). By the Blaschke selection theorem, along a subsequence on which Km→LK_m\to L we may assume that Km→LK_m\to L for some L∈KnL\in\mathcal{K}^n. We observe that

Stab⁡G(L,L)=D∩Stab⁡G(L)∈D.\operatorname{Stab}_G(L,L)=D\cap\operatorname{Stab}_G(L)\in\mathcal{D}.

By minimality of DD, this intersection equals DD. In particular, every Q∈DQ\in D fixes LL.

Let L′L' be a limit of a subsequence of RmKmR_mK_m. By compactness we may pass to a further subsequence on which

Rm→R,(Km,Km)→(A,B)R_m\to R,\qquad(K_m,K_m)\to(A,B)

for some tuple AA of convex bodies and a further B∈KnB\in\mathcal{K}^n. Applying Lemma 3.4 to the enlarged tuple, we find Q∈GQ\in G such that (A,B)=Q(L,L)(A,B)=Q(L,L). Since RA=LRA=L, we have RQ∈DRQ\in D, and consequently L′=RB=RQL=LL^{\prime}=RB=RQL=L. The corrected symmetrals are uniformly bounded, so uniqueness of their subsequential limit gives RmKm→LR_mK_m\to L. This proves the result since the rotations were fixed before choosing K0K_0.

Completing the proof of Theorem 1.1. By Proposition 3.5, there is a common sequence (Rm)(R_m) satisfying (8). If the symmetrals converge for every initial body, we may take Rm=Id⁡R_m=\operatorname{Id}. Suppose instead that there is an initial body for which (Km)(K_m) fails to converge. If ∑mdG(Rm,Rm−1)<∞\sum_m d_G(R_m,R_{m-1})<\infty, the rotations are Cauchy and converge to some R∈GR\in G. But then RmKm→LR_mK_m\to L implies Km→R−1LK_m\to R^{-1}L, a contradiction. This completes the proof.

Final remarks and extensions

Apart from HH-symmetry of the symmetral, a careful inspection of the arguments shows that the key properties of Steiner symmetrization which are used to prove the results above are the following:

  • a) For every K∈KnK\in\mathcal{K}^n such that ρHK=K\rho_HK=K, SHK=KS_HK=K,

  • b) If K,L∈KnK,L\in\mathcal{K}^n satisfy K⊆LK\subseteq L then SHK⊆SHLS_HK\subseteq S_HL,

  • c) For every K∈KnK\in\mathcal{K}^n the inclusion SHK⊆MHKS_HK\subseteq M_HK holds.

In fact, (a) and (b) were used to keep the symmetrals in a fixed ball, and this also follows directly from (c) since MHM_H is monotone and fixes aBnaB^n. It follows that if we were to identify maps of Kn\mathcal{K}^n into itself producing HH-symmetric sets and satisfying (c), the proofs in the previous sections would apply to these maps as well. For example, Minkowski symmetrization satisfies the required assumptions. As a matter of fact, Bianchi, Gardner, and Gronchi [2 Corollary 7.3] proved that (a) and (b) give (c) under one additional condition. These maps are particular examples of HH-symmetrizations, as defined in [2, 3]. We report a version of [2 Corollary 7.3] adapted to our context. Another application of this result to convergence of symmetrization processes can be found in [13].

For a linear subspace HH of Rn\mathbb{R}^n, we use the same notation ρH\rho_H for the orthogonal map that fixes HH and acts by −Id⁡-\operatorname{Id} on H⊥H^\perp. The Minkowski symmetral MHCM_HC is defined by the same formula as in Section 2. We denote by FHKF_HK the fibre symmetral of KK, obtained by replacing each non-empty section K∩(x+H⊥)K\cap(x+H^\perp), x∈Hx\in H, by its Minkowski symmetral within that fibre with respect to reflection about xx. In particular, FHK=SHKF_HK=S_HK when HH is a hyperplane.

Proposition 4.1. Consider a map

⋄:(⋃i=0n−1Gr⁡n,i)×Kn⟶Kn\diamond:\left(\bigcup_{i=0}^{n-1}\operatorname{Gr}_{n,i}\right)\times\mathcal{K}^n\longrightarrow\mathcal{K}^n

such that for every HH and KK one has ρH⋄HK=⋄HK\rho_H\diamond_HK=\diamond_HK. Suppose, moreover, that the analogues of (a) and (b) hold with SHS_H replaced by ⋄H\diamond_H, and that

⋄H(K+x)=K\diamond_H(K+x)=K

whenever KK is HH-symmetric and x∈H⊥x\in H^\perp. Then

FHK⊆⋄HK⊆MHKF_HK\subseteq\diamond_HK\subseteq M_HK

for every such HH and every K∈KnK\in\mathcal{K}^n.

The corresponding convergence result is as follows. As in Theorem 1.1, we allow the rotations to depend on the sequence of subspaces alone.

Corollary 4.2. Consider a sequence (Hm)(H_m) of proper linear subspaces of Rn\mathbb{R}^n, possibly of different dimensions. Then there exists a sequence (Rm)⊂SO(n)(R_m) \subset SO(n) satisfying (9) with the following property. For every K∈KnK \in\mathcal{K}^n and every sequence of HmH_m-symmetrizations ⋄m\diamond_m such that ⋄mA⊆MHmA\diamond_m A \subseteq M_{H_m}A for every A∈KnA \in\mathcal{K}^n, the sequence

K0=K,Km=⋄mKm−1,m∈N,K_0 = K,\qquad K_m = \diamond_m K_{m-1},\qquad m \in\mathbb{N},

satisfies

dH(RmKm,L)→0d_{\mathrm{H}}(R_mK_m,L) \to0

for some L∈KnL \in\mathcal{K}^n. In particular, the same rotations apply to fibre and Minkowski symmetrizations and to the symmetrizations in Proposition 4.1. If one of these uncorrected sequences does not converge, the rotations satisfy (1).

Proof. The containment in the Minkowski symmetral implies (4), and HmH_m-symmetry gives Tmfm=fmT_m f_m = f_m. These observations still hold for a tuple of processes that use different symmetrizations along the same subspaces. When each HmH_m is a hyperplane, the proof of Theorem 1.1 therefore applies without further changes.

For arbitrary subspaces we need to reconsider the last step in the proof of Lemma 2.3, since ρHi\rho_{H_i} can have determinant 11. Before this step, the proof gives (5) with the minimum over O(n)O(n), also for tuples. As in the proof of Theorem 2.1, it follows that fmf_m approaches the O(n)O(n)-orbit of the support-function tuple of any subsequential limit LL. This orbit consists of at most two SO(n)SO(n)-orbits. If there are two, they are compact and disjoint, hence have positive distance. On the other hand, (10) holds in the product norm and shows that ∥fm−fm−1∥→0\lVert f_m - f_{m-1}\rVert\to0. Thus, for sufficiently large mm, the tuple cannot pass from a small neighbourhood of one orbit to a small neighbourhood of the other. Since LL is a subsequential limit, fmf_m approaches the SO(n)SO(n)-orbit corresponding to LL.

We can now use the proof of Theorem 3.2 for tuples, followed by the minimal-stabilizer argument in § 3.2. In choosing a minimal stabilizer there, we consider the tuple limits of all the processes in the statement, so that the resulting rotations are common to them all. The same argument at the end of § 3 gives the first half of (1) if any uncorrected sequence fails to converge.

We conclude with a further extension that applies to Minkowski symmetrization. As pointed out in [14], the behaviour of Minkowski symmetrization on Kn\mathcal{K}^n can be used to describe the one on the family of non-empty compact subsets of Rn\mathbb{R}^n, which we denote by Cn\mathcal{C}^n. The corresponding result in this case looks as follows.

Corollary 4.3. For C∈CnC \in\mathcal{C}^n and a sequence (Hm)(H_m) as in Corollary 4.2, consider the sequence

C0=C,Cm=MHmCm−1,m∈N,C_0 = C,\qquad C_m = M_{H_m}C_{m-1},\qquad m \in\mathbb{N},

Then the same rotations (Rm)⊂SO(n)(R_m) \subset SO(n) as in Corollary 4.2 give

dH(RmCm,L)→0d_{\mathrm{H}}(R_mC_m,L) \to0

for some L∈KnL \in\mathcal{K}^n. In particular, one sequence of rotations works for every C∈CnC \in\mathcal{C}^n. If (Cm)(C_m) does not converge for some C∈CnC \in\mathcal{C}^n, these rotations satisfy (1).

Proof. Put Km=conv⁡CmK_m = \operatorname{conv} C_m. First, recall that the convex hull commutes with Minkowski addition and, therefore, with Minkowski symmetrization. In particular, (Km)(K_m) is the Minkowski symmetrization process starting from conv⁡C\operatorname{conv} C. Moreover, the estimate from the Shapley–Folkman–Starr theorem [12 Theorem 3.1.6] gives

dH(Cm,Km)≤n2mdiam⁡C⟶0.d_H(C_m,K_m) \leq\frac{\sqrt{n}}{2^m}\operatorname{diam} C \longrightarrow0.

See [14 Theorem 3.4 and Corollary 3.5] for a similar argument. Applying the Corollary 4.2 to (Km)(K_m) and using rotation invariance of dHd_H, we obtain convergence of (RmCm)(R_m C_m) to the same limit as (RmKm)(R_m K_m). Finally, the estimate shows that (Cm)(C_m) converges if and only if (Km)(K_m) converges, so the last claim also follows Corollary 4.2. □

Acknowledgments. The author was supported, in part, by the Austrian Science Fund (FWF): 10.55776/PAT3787224. The author thanks Georg Hofstätter for insightful discussions.

AI disclosure. This manuscript, both in writing and mathematical proofs, was prepared with the assistance of AI. In particular, GPT-6 Astra was used. The mathematical results were assisted and refined through various rounds of prompting. The final form of proofs and writing in the manuscript were handcrafted by the author, and double-checked with AI for typos and stylistic suggestions.

Jacopo Ulivelli

Institut für Diskrete Mathematik und Geometrie

TU Wien

Wiedner Hauptstraße 8–10/1046

1040 Wien, Austria

e-mail: [email protected]

References

  1. [1]G. Bianchi, A. Burchard, P. Gronchi, and A. Volčič, Convergence in shape of Steiner symmetrizations, Indiana Univ. Math. J. 61 (2012), pp. 1695–1710.DOI
  2. [2]G. Bianchi, R. J. Gardner, and P. Gronchi, Symmetrization in geometry, Adv. Math. 306 (2017), pp. 51–88.DOI
  3. [3]G. Bianchi, R. J. Gardner, and P. Gronchi, Convergence of symmetrization processes, Indiana Univ. Math. J. 71 (2022), pp. 785–817.DOI
  4. [4]G. Bianchi and P. Gronchi, private communication, 2019.
  5. [5]G. Bianchi and P. Gronchi, Symmetrizations, in Convex Geometry, Lecture Notes in Math., vol. 2332, Springer, Cham, 2023, pp. 233–292.DOI
  6. [6]T. Bröcker and T. tom Dieck, Representations of Compact Lie Groups, Graduate Texts in Mathematics, vol. 98, Springer-Verlag, New York, 1995, corrected reprint of the 1985 translation.
  7. [7]A. Burchard and M. Fortier, Random polarizations, Adv. Math. 234 (2013), pp. 550–573.arxiv.org/abs/1104.4103
  8. [8]D. Coupier and Y. Davydov, Random symmetrizations of convex bodies, Adv. in Appl. Probab. 46 (2014), pp. 603–621.arxiv.org/abs/1211.1785
  9. [9]W. Groß, Die Minimaleigenschaft der Kugel, Monatsh. Math. Phys. 28 (1917), pp. 77–97.
  10. [10]J. M. Lee, Introduction to Smooth Manifolds, Graduate Texts in Mathematics, vol. 218, 2nd ed., Springer, New York, 2013.
  11. [11]P. Mani-Levitska, Random Steiner symmetrizations, Studia Sci. Math. Hungar. 21 (1986), pp. 373–378.
  12. [12]R. Schneider, Cover Bodies: The Brunn–Minkowski Theory, Encyclopedia of Mathematics and its Applications, vol. 151, 2nd expanded ed., Cambridge University Press, Cambridge, 2014.
  13. [13]J. Ulivelli, Convergence properties of symmetrization processes, Adv. in Appl. Math. 146 (2023), Paper No. 102484, 18 pp.DOI
  14. [14]J. Ulivelli, Generalization of Klain’s theorem to Minkowski symmetrization of compact sets and related topics, Canad. Math. Bull. 66 (2023), pp. 124–141.

Paper details

vlnc:W_qg3wend6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

GPT-6 Astra

View statement
Looking for related papers…
Contents