betaFeedback

WorksMathematics

Unique determination of ellipsoids by their intrinsic volumes

Abstract

We prove that ellipsoids in Rn\mathbb{R}^n, n≥2n\ge 2, are uniquely determined by their intrinsic volumes up to congruence, resolving the conjecture posed by Gusakova and Zaporozhets in 2017.

Introduction

Let KK be a convex body in Rn\mathbb{R}^n and denote by B2nB_2^n the unit Euclidean ball in Rn\mathbb{R}^n. The Steiner formula asserts that for every t≥0t \ge0 we have

Vol⁡n(K+tB2n)=∑k=0nκn−kVk(K)tn−k,\operatorname{Vol}_n(K+tB_2^n)=\sum_{k=0}^{n}\kappa_{n-k}V_k(K)t^{n-k},

where

κk=πk/2Γ(k/2+1)\kappa_k=\frac{\pi^{k/2}}{\Gamma(k/2+1)}

is the volume of B2kB_2^k.

The coefficients Vk(K)V_k(K), k=0,1,…,nk=0,1,\ldots,n, are called the intrinsic volumes. Note that Vn(K)V_n(K) is equal to the volume of KK, Vn−1(K)V_{n-1}(K) is a multiple of its surface area, V1(K)V_1(K) is a multiple of its mean width, and V0(K)=1V_0(K)=1 for any convex body.

Gusakova and Zaporozhets asked whether ellipsoids in Rn\mathbb{R}^n are uniquely determined (up to an isometry) by their intrinsic volumes. In the three-dimensional case, this question was answered affirmatively by Petrov and Tarasov [6]. We solve the problem in all dimensions.

Theorem 1.1. Two ellipsoids in Rn\mathbb{R}^n, n≥2n \ge2, with the same intrinsic volumes V1,…,VnV_1,\ldots,V_n are congruent.

The analogous result for dual volumes was obtained in [5].

Throughout the paper, we assume that all ellipsoids are full dimensional. The degenerate case can be reduced to subspaces of lower dimensions.

Preliminaries

To every ellipsoid EE whose semiaxes squared are λ1,…,λn>0\lambda_1,\ldots,\lambda_n>0, we associate the polynomial

Q(u)=∏i=1n(1+λiu)=∑j=0nsj(Q)uj.(1)Q(u)=\prod_{i=1}^{n}(1+\lambda_i u)=\sum_{j=0}^{n}s_j(Q)u^j. \tag*{(1)}

Here sj=sj(Q)s_j=s_j(Q) is the jjth elementary symmetric polynomial in λi\lambda_i defined by

sj=∑1≤i1<⋯<ij≤n∏p=1jλip,if j≥1,s_j=\sum_{1\le i_1<\cdots<i_j\le n}\prod_{p=1}^{j}\lambda_{i_p},\qquad\text{if }j\ge1,

and s0=1s_0=1. The following is a consequence of Newton’s inequalities (see [3 Theorem 53]):

sj2≥sj−1sj+1,1≤j<n.(2)s_j^2 \ge s_{j-1}s_{j+1}, \qquad1 \le j < n. \tag*{(2)}

The coefficients s1,…,sns_1,\ldots,s_n determine the semiaxes of EE up to permutation. Our goal is to show that the intrinsic volumes determine these coefficients.

Throughout the paper, we will use the following rescaled version of intrinsic volumes:

vk(K)=Vk(K)κk,0≤k≤n.v_k(K)=\frac{V_k(K)}{\kappa_k}, \qquad0 \le k \le n.

Since an ellipsoid is uniquely determined (up to isometry) by the associated polynomial QQ, it is enough to work with such polynomials and we will use the following notation:

vk(Q)=vk(E).v_k(Q)=v_k(E).

If the ellipsoid or its polynomial QQ are clear from the context, we will just write vkv_k.

It will often be convenient to work with a polynomial

Q(u)=1+s1u+⋯+sn−1un−1+snun,Q(u)=1+s_1u+\cdots+s_{n-1}u^{n-1}+s_nu^n,

whose coefficients are positive, but QQ is not necessarily associated with an ellipsoid.

For such a polynomial and 0≤k≤n0\le k\le n, we define

Tk(u)=∑p<kspup,T_k(u)=\sum_{p<k}s_pu^p,
Ak(u)=∑p<k(k−p)spup,A_k(u)=\sum_{p<k}(k-p)s_pu^p,
Bk(u)=∑p>k(p−k)spup.B_k(u)=\sum_{p>k}(p-k)s_pu^p.

In particular, A0=Bn=0A_0=B_n=0. Note that

Ak(u)=kTk(u)−uTk′(u)A_k(u)=kT_k(u)-uT_k'(u)

and

Ak+1=Ak+Tk+1.A_{k+1}=A_k+T_{k+1}.

Similarly,

Bk(u)=u(Q−Tk)′−k(Q−Tk).B_k(u)=u(Q-T_k)'-k(Q-T_k).

Whenever we need to emphasize that AkA_k and BkB_k are associated with a given polynomial QQ, we will write AkQA_k^Q and BkQB_k^Q.

Auxiliary lemmas

First we derive a formula for vkv_k using an existing formula for intrinsic volumes of ellipsoids proved in [2].

Lemma 3.1. Let EE be an ellipsoid with the corresponding polynomials QQ and TkT_k. Then

vk=12∫0∞u−k/2TkQ′−Tk′QQ3/2 du,1≤k≤n.(3)v_k=\frac{1}{2}\int_0^\infty u^{-k/2}\frac{T_kQ'-T_k'Q}{Q^{3/2}}\,du,\qquad1\le k\le n. \tag*{(3)}

Proof. By [2 Theorem 1], we have

vk=∑i=1nλisk−1(i)∫0∞tk−1(1+λit2)∏ℓ=1n1+λℓt2 dt,v_k=\sum_{i=1}^n \lambda_i s_{k-1}^{(i)}\int_0^\infty\frac{t^{k-1}}{(1+\lambda_i t^2)\prod_{\ell=1}^n\sqrt{1+\lambda_\ell t^2}}\,dt,

where sk−1(i)s_{k-1}^{(i)} is the (k−1)(k-1)st elementary symmetric function of λ1,…,λn\lambda_1,\ldots,\lambda_n with λi\lambda_i omitted. Set u=t2u=t^2. Then tk−1 dt=12uk/2−1 dut^{k-1}\,\mathrm{d}t=\frac{1}{2}u^{k/2-1}\,\mathrm{d}u and ∏ℓ=1n1+λℓt2=Q(u)\prod_{\ell=1}^{n}\sqrt{1+\lambda_{\ell}t^2}=\sqrt{Q(u)}. Bringing the sum inside the integral yields

vk=12∫0∞uk/2−1Q(u)∑i=1nλisk−1(i)1+λiu du.(4)v_k=\frac{1}{2}\int_0^\infty\frac{u^{k/2-1}}{\sqrt{Q(u)}}\sum_{i=1}^{n}\frac{\lambda_i s_{k-1}^{(i)}}{1+\lambda_i u}\,\mathrm{d}u. \tag*{(4)}

Next we express the sum using only the coefficients of QQ and the partial sum TkT_k. Observe that

sp=sp(i)+λisp−1(i),s_p=s_p^{(i)}+\lambda_i s_{p-1}^{(i)},

where s−1(i)=0s_{-1}^{(i)}=0 by convention.

Thus

Tk(u)1+λiu=∑p=0k−1spup1+λiu=∑p=0k−1sp(i)up+λi∑p=0k−1sp−1(i)up1+λiu=sk−1(i)uk−1+∑p=0k−2sp(i)up+λi∑p=0k−2sp(i)up+11+λiu=∑p=0k−2sp(i)up+sk−1(i)uk−11+λiu.\begin{aligned} \frac{T_k(u)}{1+\lambda_i u}=\frac{\sum_{p=0}^{k-1}s_pu^p}{1+\lambda_i u}=\frac{\sum_{p=0}^{k-1}s_p^{(i)}u^p+\lambda_i\sum_{p=0}^{k-1}s_{p-1}^{(i)}u^p}{1+\lambda_i u} \\ &=\frac{s_{k-1}^{(i)}u^{k-1}+\sum_{p=0}^{k-2}s_p^{(i)}u^p+\lambda_i\sum_{p=0}^{k-2}s_p^{(i)}u^{p+1}}{1+\lambda_i u}=\sum_{p=0}^{k-2}s_p^{(i)}u^p+\frac{s_{k-1}^{(i)}u^{k-1}}{1+\lambda_i u}. \end{aligned}

The formula is valid for all 1≤k≤n1\leq k\leq n. In the case k=1k=1, the sum from 00 to k−2k-2 is zero. Now multiply by λi\lambda_i and sum over ii. Since

∑i=1nλi1+λiu=Q′Q,(5)\sum_{i=1}^{n}\frac{\lambda_i}{1+\lambda_i u}=\frac{Q'}{Q}, \tag*{(5)}

and

∑i=1nλisp(i)=(p+1)sp+1,\sum_{i=1}^{n}\lambda_i s_p^{(i)}=(p+1)s_{p+1},

we obtain

TkQ′Q=∑p=0k−2(p+1)sp+1up+uk−1∑i=1nλisk−1(i)1+λiu=Tk′+uk−1∑i=1nλisk−1(i)1+λiu.T_k\frac{Q'}{Q}=\sum_{p=0}^{k-2}(p+1)s_{p+1}u^p+u^{k-1}\sum_{i=1}^{n}\frac{\lambda_i s_{k-1}^{(i)}}{1+\lambda_i u}=T_k'+u^{k-1}\sum_{i=1}^{n}\frac{\lambda_i s_{k-1}^{(i)}}{1+\lambda_i u}.

Rearranging,

∑i=1nλisk−1(i)1+λiu=TkQ′−Tk′Quk−1Q.\sum_{i=1}^{n}\frac{\lambda_i s_{k-1}^{(i)}}{1+\lambda_i u}=\frac{T_kQ'-T_k'Q}{u^{k-1}Q}.

Substituting this into integral (4) gives (3).

Remark 3.2. If

Q(u)=∑j=0nsjujQ(u)=\sum_{j=0}^{n}s_ju^j

is a polynomial with positive coefficients, not necessarily associated with an ellipsoid, we use the right-hand side of (3) as the definition of vk(Q)v_k(Q). The identity

TkQ′−Tk′Q=∑p<k≤q(q−p)spsqup+q−1T_kQ'-T_k'Q=\sum_{p<k\leq q}(q-p)s_ps_qu^{p+q-1}

shows positivity and convergence. The integrand is O(uk/2−1)O(u^{k/2-1}) at zero and O(u(k−n)/2−2)O(u^{(k-n)/2-2}) at infinity.

The following lemma is a version of the Alexandrov-Fenchel inequalities for vkv_k.

Lemma 3.3. Let K⊂RnK\subset\mathbb{R}^n be a convex body. Then

vk(K)2≥vk−1(K)vk+1(K),1≤k<n.v_k(K)^2\geq v_{k-1}(K)v_{k+1}(K),\qquad1\leq k<n.

Proof. For any convex body K⊂RnK \subset\mathbb{R}^n, the intrinsic volumes can be written as

Vk(K)=1κn−k(nk)V(K[k],B2n[n−k]).V_k(K) = \frac{1}{\kappa_{n-k}} \binom{n}{k} V\left(K[k], B_2^n[n-k]\right).

where V(K[k],B2n[n−k])V\left(K[k], B_2^n[n-k]\right) is the corresponding mixed volume (see, for example, [1 (A.26)]). Thus

vk(K)=(nk)κkκn−kV(K[k],B2n[n−k]).v_k(K) = \frac{\binom{n}{k}}{\kappa_k\kappa_{n-k}} V\left(K[k], B_2^n[n-k]\right).

Dividing by vk(B2n)v_k(B_2^n) and using V(B2n[k],B2n[n−k])=κnV\left(B_2^n[k], B_2^n[n-k]\right) = \kappa_n gives

vk(K)vk(B2n)=V(K[k],B2n[n−k])κn.\frac{v_k(K)}{v_k(B_2^n)} = \frac{V\left(K[k], B_2^n[n-k]\right)}{\kappa_n}.

Formula (7.66) in [7] shows that this ratio is log-concave as a sequence in kk. Finally observe that the sequence

vk(B2n)=(nk)κnκkκn−k=π Γ((n+1)/2)Γ((k+1)/2)Γ((n−k+1)/2)v_k(B_2^n) = \frac{\binom{n}{k}\kappa_n}{\kappa_k\kappa_{n-k}} = \frac{\sqrt{\pi}\,\Gamma((n+1)/2)}{\Gamma((k+1)/2)\Gamma((n-k+1)/2)}

is also log-concave. Indeed, this follows from the convexity of log⁡Γ\log\Gamma, since both arguments of Γ\Gamma in the denominator are affine functions of kk. Thus, the sequence vk(K)v_k(K) is log-concave as the product of two positive log-concave sequences.

In the next lemma, we will derive an alternative representation for vkv_k and use it in conjunction with the Cauchy–Schwarz inequality to obtain a bound on the ratio of two consecutive intrinsic volumes of ellipsoids.

Lemma 3.4. For any ellipsoid in Rn\mathbb{R}^n we have

vk+1vk<sksk−1,1≤k<n.(6)\frac{v_{k+1}}{v_k} < \sqrt{\frac{s_k}{s_{k-1}}}, \qquad1 \leq k < n. \tag*{(6)}

Proof. Let λ1,…,λn>0\lambda_1,\ldots,\lambda_n > 0 be the squared semiaxes, and write

Q(u)=∏i=1n(1+λiu)=∑j=0nsjuj,Hj(u)=∑i=1nλisj−1(i)1+λiu.Q(u) = \prod_{i=1}^{n}(1+\lambda_i u) = \sum_{j=0}^{n}s_j u^j,\qquad H_j(u) = \sum_{i=1}^{n}\frac{\lambda_i s_{j-1}^{(i)}}{1+\lambda_i u}.

Formula (4) gives

vj=12∫0∞uj/2−1Hj(u)Q(u) du,1≤j≤n.(7)v_j = \frac{1}{2}\int_0^\infty\frac{u^{j/2-1}H_j(u)}{\sqrt{Q(u)}}\,du,\qquad1 \leq j \leq n. \tag*{(7)}

Fix 1≤k<n1 \leq k < n and consider the measure on (0,∞)(0,\infty) given by

dμ(u)=uk/2−12Q(u) du.d\mu(u) = \frac{u^{k/2-1}}{2\sqrt{Q(u)}}\,du.

Then (7) gives

vk=∫0∞Hk(u) dμ(u),vk+1=∫0∞u Hk+1(u) dμ(u).v_k = \int_0^\infty H_k(u)\,d\mu(u),\qquad v_{k+1} = \int_0^\infty\sqrt{u}\,H_{k+1}(u)\,d\mu(u).

We will apply the Cauchy–Schwarz inequality to the second integral and use the following two facts:

Hk+1(u)≤sksk−1Hk(u),(8)H_{k+1}(u) \leq\frac{s_k}{s_{k-1}}H_k(u), \tag*{(8)}

and

∫0∞uHk+1(u) dμ(u)=∫0∞Hk(u) dμ(u)=vk.(9)\int_0^\infty uH_{k+1}(u)\,d\mu(u) = \int_0^\infty H_k(u)\,d\mu(u) = v_k. \tag*{(9)}

To prove (8), we use the identities sj=sj(i)+λisj−1(i)s_j=s_j^{(i)}+\lambda_i s_{j-1}^{(i)}, which give

sksk−1(i)−sk−1sk(i)=λi((sk−1(i))2−sk−2(i)sk(i))≥0.s_k s_{k-1}^{(i)}-s_{k-1}s_k^{(i)}=\lambda_i\left((s_{k-1}^{(i)})^2-s_{k-2}^{(i)}s_k^{(i)}\right)\ge0.

The last inequality follows from (2) for the tuple λ1,…,λn\lambda_1,\ldots,\lambda_n with λi\lambda_i removed. Since s−1(i)=0s_{-1}^{(i)}=0 by convention, the inequality also holds for k=1k=1. Multiplying by λi/(1+λiu)\lambda_i/(1+\lambda_i u) and summing over ii, we obtain (8).

We now prove (9). A direct computation gives

uHk+1−Hk=sku∑i=1nλi1+λiu−u∑i=1nλi2sk−1(i)1+λiu−Hk=skuQ′Q−∑i=1n(λi−λi1+λiu)sk−1(i)−∑i=1nλisk−1(i)1+λiu=sk(uQ′Q−k),\begin{aligned} uH_{k+1}-H_k &=s_k u\sum_{i=1}^{n}\frac{\lambda_i}{1+\lambda_i u}-u\sum_{i=1}^{n}\frac{\lambda_i^2s_{k-1}^{(i)}}{1+\lambda_i u}-H_k\\ &=s_k\frac{uQ'}{Q}-\sum_{i=1}^{n}\left(\lambda_i-\frac{\lambda_i}{1+\lambda_i u}\right)s_{k-1}^{(i)}-\sum_{i=1}^{n}\frac{\lambda_i s_{k-1}^{(i)}}{1+\lambda_i u}\\ &=s_k\left(\frac{uQ'}{Q}-k\right), \end{aligned}

where we used (5), uλi2/(1+λiu)=λi−λi/(1+λiu)u\lambda_i^2/(1+\lambda_i u)=\lambda_i-\lambda_i/(1+\lambda_i u) and ∑i=1nλisk−1(i)=ksk\sum_{i=1}^{n}\lambda_i s_{k-1}^{(i)}=ks_k. After multiplication by dμd\mu, the last expression can be written as a total derivative:

sk(uQ′Q−k)uk/2−12Q du=−sk(uk/2Q)′du.s_k\left(\frac{uQ'}{Q}-k\right)\frac{u^{k/2-1}}{2\sqrt{Q}}\,du=-s_k\left(\frac{u^{k/2}}{\sqrt{Q}}\right)'du.

Since 0<k<n0<k<n, uk/2/Qu^{k/2}/\sqrt{Q} vanishes at the endpoints of (0,∞)(0,\infty). Thus, integrating the preceding equality yields (9).

Combining (8) and (9), we obtain

vk+12=(∫0∞u Hk+1(u) dμ(u))2<(∫0∞Hk+1(u) dμ(u))(∫0∞uHk+1(u) dμ(u))≤sksk−1vk2.\begin{aligned} v_{k+1}^2=\left(\int_0^\infty\sqrt{u}\,H_{k+1}(u)\,d\mu(u)\right)^2 \\ &<\left(\int_0^\infty H_{k+1}(u)\,d\mu(u)\right)\left(\int_0^\infty uH_{k+1}(u)\,d\mu(u)\right)\le\frac{s_k}{s_{k-1}}v_k^2. \end{aligned}

The Cauchy-Schwarz inequality is strict because Hk+1 dμH_{k+1}\,d\mu has positive density on (0,∞)(0,\infty), where u\sqrt{u} is not constant. This proves (6).

All integrals above are finite. Indeed, Hj(u)=O(1)H_j(u)=O(1) at zero and Hj(u)=O(u−1)H_j(u)=O(u^{-1}) at infinity. The density of Hk+1 dμH_{k+1}\,d\mu is therefore O(uk/2−1)O(u^{k/2-1}) at zero and O(u(k−n)/2−2)O(u^{(k-n)/2-2}) at infinity. Both its total integral and its first moment converge for 1≤k<n1\le k<n. □

We will also need the following elementary fact about matrix invertibility.

Lemma 3.5. Let M=(Mk,j)M=(M_{k,j}) be an m×mm\times m matrix with positive entries. Suppose

Mk−1,j−2Mk,j+Mk+1,j≥0,for 1≤k≠j≤m,M_{k-1,j}-2M_{k,j}+M_{k+1,j}\ge0,\qquad\text{for }1\le k\ne j\le m,

where M0,j=0=Mm+1,jM_{0,j}=0=M_{m+1,j} by convention. Then MM is invertible.

Proof. Consider an m×mm\times m matrix G=(Gk,j)G=(G_{k,j}) such that

Gk,j=2Mk,j−Mk−1,j−Mk+1,j.G_{k,j}=2M_{k,j}-M_{k-1,j}-M_{k+1,j}.

By assumption, Gk,j≤0G_{k,j}\le0 when k≠jk\ne j. Also, note that

∑k=1mGk,j=M1,j+Mm,j>0\sum_{k=1}^{m}G_{k,j}=M_{1,j}+M_{m,j}>0

where we used that MM has positive entries. This implies that

Gj,j>−∑k≠jGk,j=∑k≠j∣Gk,j∣≥0.(10)G_{j,j}>-\sum_{k\ne j}G_{k,j}=\sum_{k\ne j}|G_{k,j}|\ge0. \tag*{(10)}

We claim that GG is invertible. Indeed, assume that there is x∈Rmx \in\mathbb{R}^{m}, x≠0x \ne0 such that GTx=0G^{T}x=0, and choose jj such that ∣xj∣=max⁡1≤k≤m∣xk∣≠0|x_j|=\max_{1\le k\le m}|x_k|\ne0. Then, using (10), we have

Gj,j∣xj∣=∣∑k≠jGk,jxk∣≤∑k≠j∣Gk,j∣∣xk∣≤∣xj∣∑k≠j∣Gk,j∣<Gj,j∣xj∣,G_{j,j}|x_j|=\left|\sum_{k\ne j}G_{k,j}x_k\right|\le\sum_{k\ne j}|G_{k,j}||x_k|\le|x_j|\sum_{k\ne j}|G_{k,j}|<G_{j,j}|x_j|,

which is impossible. Thus, GG is invertible. Since the rows of GG are linear combinations of the rows of MM, Mx=0Mx=0 implies Gx=0Gx=0. Thus, MM is also invertible. □

Main Lemmas

Consider the space of polynomials Q(u)=1+s1u+⋯+sn−1un−1+snunQ(u)=1+s_1u+\cdots+s_{n-1}u^{n-1}+s_nu^n with positive coefficients, where the coefficient sns_n is fixed. This assumption comes from the fact that we are working with ellipsoids of fixed volume. Thus, this space of polynomials can be identified with (0,∞)n−1(0,\infty)^{n-1}, and we can think of vk(Q)v_k(Q), 1≤k≤n−11\le k\le n-1, as a function of (s1,…,sn−1)∈(0,∞)n−1(s_1,\ldots,s_{n-1})\in(0,\infty)^{n-1}. The lemma below gives the derivatives of this function.

Lemma 4.1. For 1≤k,j<n1\le k,j<n,

∂vk∂sj={14∫0∞uj−k2−1Q−32Ak du,j≥k,14∫0∞uj−k2−1Q−32Bk du,j<k.(11)\frac{\partial v_k}{\partial s_j}= \begin{cases} \frac{1}{4}\int_{0}^{\infty}u^{j-\frac{k}{2}-1}Q^{-\frac{3}{2}}A_k\,du, & j\ge k,\\[6pt] \frac{1}{4}\int_{0}^{\infty}u^{j-\frac{k}{2}-1}Q^{-\frac{3}{2}}B_k\,du, & j<k. \end{cases} \tag*{(11)}

When j=kj=k, either AkA_k or BkB_k gives the same integral. In particular, all the partial derivatives ∂vk/∂sj\partial v_k/\partial s_j are positive for 1≤k,j<n1\le k,j<n.

Proof. To obtain (11), we treat the cases j≥kj\ge k and j<kj<k separately. First, let j≥kj\ge k. In this case, TkT_k does not depend on sjs_j. Differentiating (3) gives

∂vk∂sj=12∫0∞u−k2((jTkuj−1−Tk′uj)Q−32−32(TkQ′−Tk′Q)Q−52uj) du=12∫0∞uj−k2−1Q−32(jTk+12uTk′−32uTkQ′Q−1) du.(12)\begin{aligned} \frac{\partial v_k}{\partial s_j} &=\frac{1}{2}\int_{0}^{\infty}u^{-\frac{k}{2}}\left((jT_ku^{j-1}-T'_ku^j)Q^{-\frac{3}{2}}-\frac{3}{2}(T_kQ'-T'_kQ)Q^{-\frac{5}{2}}u^j\right)\,du\\ &=\frac{1}{2}\int_{0}^{\infty}u^{j-\frac{k}{2}-1}Q^{-\frac{3}{2}}\left(jT_k+\frac{1}{2}uT'_k-\frac{3}{2}uT_kQ'Q^{-1}\right)\,du. \tag*{(12)} \end{aligned}

The differentiation is justified since the integrand in (12) is of order O(uk/2−1)O(u^{k/2-1}) at zero and O(u(k−n)/2−2)O(u^{(k-n)/2-2}) at infinity, where the bounds in OO are locally uniform in s1,…,sn−1s_1,\ldots,s_{n-1}.

Now we observe that

12(uj−k2TkQ−32)′=12(j−k2)uj−k2−1TkQ−32+12uj−k2Tk′Q−32−34uj−k2TkQ−52Q′.\frac{1}{2}\left(u^{j-\frac{k}{2}}T_kQ^{-\frac{3}{2}}\right)'=\frac{1}{2}\left(j-\frac{k}{2}\right)u^{j-\frac{k}{2}-1}T_kQ^{-\frac{3}{2}}+\frac{1}{2}u^{j-\frac{k}{2}}T'_kQ^{-\frac{3}{2}}-\frac{3}{4}u^{j-\frac{k}{2}}T_kQ^{-\frac{5}{2}}Q'.

The boundary values of uj−k/2TkQ−3/2u^{j-k/2}T_kQ^{-3/2} vanish at 00 and at ∞\infty since j−k2>0j-\frac{k}{2}>0 and j+k2−1<3n2j+\frac{k}{2}-1<\frac{3n}{2}. Subtracting this total derivative from the integrand in (12) yields

∂vk∂sj=14∫0∞uj−k2−1Q−32(kTk−uTk′) du=14∫0∞uj−k2−1Q−32Ak du\frac{\partial v_k}{\partial s_j}=\frac{1}{4}\int_{0}^{\infty}u^{j-\frac{k}{2}-1}Q^{-\frac{3}{2}}(kT_k-uT'_k)\,du=\frac{1}{4}\int_{0}^{\infty}u^{j-\frac{k}{2}-1}Q^{-\frac{3}{2}}A_k\,du

which gives (11) for j≥kj\ge k. Now, let j<kj<k. We can rewrite (3) as

vk=12∫0∞u−k2(Q−Tk)′Q−(Q−Tk)Q′Q32 duv_k=\frac{1}{2}\int_{0}^{\infty}u^{-\frac{k}{2}}\frac{(Q-T_k)'Q-(Q-T_k)Q'}{Q^{\frac{3}{2}}}\,du

where Q−TkQ-T_k does not depend on sjs_j. Performing a similar computation as in (12) and adding the total derivative 12(uj−k/2(Q−Tk)Q−3/2)′\frac{1}{2}\left(u^{j-k/2}(Q-T_k)Q^{-3/2}\right)' to the integrand gives

∂vk∂sj=14∫0∞uj−k/2−1Q−3/2(u(Q−Tk)′−k(Q−Tk)) du=14∫0∞uj−k/2−1Q−3/2Bk du.\frac{\partial v_k}{\partial s_j}=\frac{1}{4}\int_0^\infty u^{j-k/2-1}Q^{-3/2}\left(u(Q-T_k)'-k(Q-T_k)\right)\,\mathrm{d}u=\frac{1}{4}\int_0^\infty u^{j-k/2-1}Q^{-3/2}B_k\,\mathrm{d}u.

The boundary values also vanish at 00 and at ∞\infty. This confirms (11) for j<kj<k.

Finally, since Ak−Bk=kQ−uQ′A_k-B_k=kQ-uQ', we get

∫0∞uk/2−1(Ak−Bk)Q−3/2 du=∫0∞uk/2−1(kQ−uQ′)Q−3/2 du=2∫0∞(uk/2Q)′ du=0\begin{aligned} \int_0^\infty u^{k/2-1}(A_k-B_k)Q^{-3/2}\,\mathrm{d}u=\int_0^\infty u^{k/2-1}(kQ-uQ')Q^{-3/2}\,\mathrm{d}u \\ &=2\int_0^\infty\left(\frac{u^{k/2}}{\sqrt{Q}}\right)'\,\mathrm{d}u=0 \end{aligned}

where in the last equality the boundary terms vanish at 00 and at ∞\infty since 0<k<n0<k<n. Thus, when j=kj=k, either expression in (11) can be used. The coefficients of AkA_k and BkB_k are positive, so ∂vk∂sj\frac{\partial v_k}{\partial s_j} are also positive for any 1≤k,j<n1\leq k,j<n.

To the integrands in (11), we associate the normalized functions

ak(u)=u−k/2Ak(u)vk,bk(u)=u−k/2Bk(u)vk,0≤k≤n.a_k(u)=\frac{u^{-k/2}A_k(u)}{v_k},\qquad b_k(u)=\frac{u^{-k/2}B_k(u)}{v_k},\qquad0\leq k\leq n.

Here v0=1v_0=1, and a0=bn=0a_0=b_n=0. The normalization by vkv_k will be useful when comparing two ellipsoids with the same intrinsic volumes.

Lemma 4.2. Let EE be an ellipsoid with the corresponding polynomial QQ. For every u>0u>0, the sequences ak(u)a_k(u) and bk(u)b_k(u) are strictly convex, that is,

ak−1−2ak+ak+1>0,bk−1−2bk+bk+1>0a_{k-1}-2a_k+a_{k+1}>0,\qquad b_{k-1}-2b_k+b_{k+1}>0

for every 1≤k<n1\leq k<n.

Proof. First, we treat the sequence ak(u)a_k(u). Fix 1≤k<n1\leq k<n. By Newton’s inequalities (2), we have

spsp−1≥sksk−1>(vk+1vk)2,1≤p≤k,\frac{s_p}{s_{p-1}}\geq\frac{s_k}{s_{k-1}}>\left(\frac{v_{k+1}}{v_k}\right)^2,\qquad1\leq p\leq k,

where we use Lemma 3.4. Therefore

Tk+1−(vk+1vk)2uTk=s0+∑p=1k(sp−(vk+1vk)2sp−1)up>0(13)T_{k+1}-\left(\frac{v_{k+1}}{v_k}\right)^2uT_k=s_0+\sum_{p=1}^{k}\left(s_p-\left(\frac{v_{k+1}}{v_k}\right)^2s_{p-1}\right)u^p>0 \tag*{(13)}

since every coefficient in this expression is positive.

We now consider

uk+12vk+1(ak−1−2ak+ak+1)=Ak+1−2vk+1vkuAk+uvk+1vk−1Ak−1≥Ak+1−2vk+1vkuAk+(vk+1vk)2uAk−1=(1−vk+1vku)2Ak+Tk+1−(vk+1vk)2uTk>0\begin{aligned} u^{\frac{k+1}{2}}v_{k+1}(a_{k-1}-2a_k+a_{k+1})=A_{k+1}-2\frac{v_{k+1}}{v_k}\sqrt{u}A_k+u\frac{v_{k+1}}{v_{k-1}}A_{k-1} \\ &\geq A_{k+1}-2\frac{v_{k+1}}{v_k}\sqrt{u}A_k+\left(\frac{v_{k+1}}{v_k}\right)^2uA_{k-1} \\ &=\left(1-\frac{v_{k+1}}{v_k}\sqrt{u}\right)^2A_k+T_{k+1}-\left(\frac{v_{k+1}}{v_k}\right)^2uT_k>0 \end{aligned}

where we used Lemma 3.3 and the identities Ak+1=Ak+Tk+1A_{k+1}=A_k+T_{k+1} and Ak−1=Ak−TkA_{k-1}=A_k-T_k. Since the first term is nonnegative and the second is strictly positive by (13),

ak−1−2ak+ak+1>0.a_{k-1}-2a_k+a_{k+1}>0.

We now prove the assertion for bkb_k. Without loss of generality, assume that EE is centered at the origin. Consider

Q∗(u)=unsnQ(1u)=∏i=1n(1+λi−1u)=∑p=0nsn−psnup.Q^*(u)=\frac{u^n}{s_n}Q\left(\frac{1}{u}\right)=\prod_{i=1}^{n}(1+\lambda_i^{-1}u)=\sum_{p=0}^{n}\frac{s_{n-p}}{s_n}u^p.

We note that Q∗Q^* is the polynomial of the polar ellipsoid E∗E^* of EE. By [4 Prop. 4.8],

vj(E∗)=vn−j(E)sn.v_j(E^*)=\frac{v_{n-j}(E)}{\sqrt{s_n}}.

Since the coefficients of Q∗Q^* are sp∗=sn−psns_p^*=\frac{s_{n-p}}{s_n}, we get

An−kQ∗(1u)=∑p<n−k(n−k−p)sp∗u−p=1sn∑p<n−k(n−k−p)sn−pu−p=1sn∑k<p(p−k)spup−n=u−nsnBkQ(u).\begin{aligned} A_{n-k}^{Q^*}\left(\frac{1}{u}\right)=\sum_{p<n-k}(n-k-p)s_p^*u^{-p}=\frac{1}{s_n}\sum_{p<n-k}(n-k-p)s_{n-p}u^{-p} \\ &=\frac{1}{s_n}\sum_{k<p}(p-k)s_pu^{p-n}=\frac{u^{-n}}{s_n}B_k^Q(u). \end{aligned}

Therefore,

an−kQ∗(1u)=un−k2An−kQ∗(1u)vn−k(E∗)=u−n+k2BkQ(u)snvk(E)=u−n2snbkQ(u).a_{n-k}^{Q^*}\left(\frac{1}{u}\right)=\frac{u^{\frac{n-k}{2}}A_{n-k}^{Q^*}\left(\frac{1}{u}\right)}{v_{n-k}(E^*)}=\frac{u^{-\frac{n+k}{2}}B_k^Q(u)}{\sqrt{s_n}v_k(E)}=\frac{u^{-\frac{n}{2}}}{\sqrt{s_n}}b_k^Q(u).

Thus,

bk−1Q(u)−2bkQ(u)+bk+1Q(u)=snun2(an−k+1Q∗(1u)−2an−kQ∗(1u)+an−k−1Q∗(1u))>0b_{k-1}^Q(u)-2b_k^Q(u)+b_{k+1}^Q(u)=\sqrt{s_n}u^{\frac{n}{2}}\left(a_{n-k+1}^{Q^*}\left(\frac{1}{u}\right)-2a_{n-k}^{Q^*}\left(\frac{1}{u}\right)+a_{n-k-1}^{Q^*}\left(\frac{1}{u}\right)\right)>0

where the result follows from the assertion for aka_k applied to Q∗Q^*. □\square

Proof of Theorem 1.1

Proof. Let E1E_1 and E2E_2 be ellipsoids with the same intrinsic volumes, and P,QP,Q be the associated polynomials:

P(u)=∑i=0nsi(P)ui;Q(u)=∑i=0nsi(Q)ui.P(u)=\sum_{i=0}^{n}s_i(P)u^i;\qquad Q(u)=\sum_{i=0}^{n}s_i(Q)u^i.

Denote

vk(P)=vk(Q)=vk,1≤k≤n,v_k(P)=v_k(Q)=v_k,\qquad1\leq k\leq n,

and observe that

sn(P)=sn(Q)=vn2.s_n(P)=s_n(Q)=v_n^2.

Let

Qt=(1−t)P+tQ,0≤t≤1.Q_t=(1-t)P+tQ,\qquad0\leq t\leq1.

The coefficients of QtQ_t are given by si(Qt)=(1−t)si(P)+tsi(Q)s_i(Q_t)=(1-t)s_i(P)+ts_i(Q), which are positive numbers. Thus, formula (3) defines vk(Qt)v_k(Q_t) for all 0≤t≤10\leq t\leq1.

Consider the (n−1)×(n−1)(n-1)\times(n-1) matrix M=(Mk,j)M=(M_{k,j}), where

Mk,j=1vk∫01∂vk(Qt)∂sj dt,1≤k,j<n.(14)M_{k,j}=\frac{1}{v_k}\int_0^1\frac{\partial v_k(Q_t)}{\partial s_j}\,dt,\qquad1\leq k,j<n. \tag*{(14)}

By Lemma 4.1, all entries of MM are positive.

We now verify the condition of Lemma 3.5 for the matrix MM. By formula (11),

∂vk(Qt)∂sj=14∫0∞uj−k2−1Qt−32C(Qt) du.(15)\frac{\partial v_k(Q_t)}{\partial s_j}=\frac{1}{4}\int_0^\infty u^{j-\frac{k}{2}-1}Q_t^{-\frac{3}{2}}C(Q_t)\,du. \tag*{(15)}

where C(Qt)C(Q_t) is either AkQtA_k^{Q_t} or BkQtB_k^{Q_t}, depending on whether j≥kj \ge k or j<kj < k. In both cases, C(Qt)C(Q_t) depends linearly on the coefficients of QtQ_t, and so C(Qt)=(1−t)C(P)+tC(Q)C(Q_t)=(1-t)C(P)+tC(Q). Thus, we can evaluate the following integral:

∫01C(Qt)Qt3/2 dt=C(P)∫011−t((1−t)P+tQ)3/2 dt+C(Q)∫01t((1−t)P+tQ)3/2 dt=2(P+Q)2(C(P)P+C(Q)Q),(16)\begin{aligned} \int_0^1 \frac{C(Q_t)}{Q_t^{3/2}}\,dt=C(P)\int_0^1\frac{1-t}{((1-t)P+tQ)^{3/2}}\,dt+C(Q)\int_0^1\frac{t}{((1-t)P+tQ)^{3/2}}\,dt \\ &=\frac{2}{(\sqrt{P}+\sqrt{Q})^2}\left(\frac{C(P)}{\sqrt{P}}+\frac{C(Q)}{\sqrt{Q}}\right), \tag*{(16)} \end{aligned}

where the last equality follows by direct integration.

For j≥kj\ge k, combining (14), (15), (16), and the definition of aka_k for the polynomials PP and QQ:

akP=u−k/2AkPvk,akQ=u−k/2AkQvk,a_k^P=\frac{u^{-k/2}A_k^P}{v_k},\qquad a_k^Q=\frac{u^{-k/2}A_k^Q}{v_k},

we get

Mk,j=12∫0∞uj−1(P+Q)2(akPP+akQQ) du.(17)M_{k,j}=\frac{1}{2}\int_0^\infty\frac{u^{j-1}}{(\sqrt{P}+\sqrt{Q})^2}\left(\frac{a_k^P}{\sqrt{P}}+\frac{a_k^Q}{\sqrt{Q}}\right)\,du. \tag*{(17)}

For j<kj<k, the same formula holds with bkb_k in place of aka_k. When j=kj=k, either formula is valid by Lemma 4.1.

Fix a column jj, and set M0,j=Mn,j=0M_{0,j}=M_{n,j}=0. If k<jk<j, we can express each of Mk−1,jM_{k-1,j}, Mk,jM_{k,j}, and Mk+1,jM_{k+1,j} using formula (17) with the corresponding functions aa. Thus

Mk−1,j−2Mk,j+Mk+1,j=12∫0∞uj−1(P+Q)2(ak−1P−2akP+ak+1PP+ak−1Q−2akQ+ak+1QQ) du>0,M_{k-1,j}-2M_{k,j}+M_{k+1,j}=\frac{1}{2}\int_0^\infty\frac{u^{j-1}}{(\sqrt{P}+\sqrt{Q})^2}\left(\frac{a_{k-1}^P-2a_k^P+a_{k+1}^P}{\sqrt{P}}+\frac{a_{k-1}^Q-2a_k^Q+a_{k+1}^Q}{\sqrt{Q}}\right)\,du>0,

by Lemma 4.2. If k>jk>j, we use the formula with bk−1b_{k-1}, bkb_k, bk+1b_{k+1} in the same way. The conventions M0,j=Mn,j=0M_{0,j}=M_{n,j}=0 agree with a0=0a_0=0 and bn=0b_n=0. Therefore, by Lemma 3.5 the matrix MM is invertible.

Now let

Δsj=sj(Q)−sj(P).\Delta s_j=s_j(Q)-s_j(P).

Observe that Δsn=0\Delta s_n=0 due to sn(P)=sn(Q)s_n(P)=s_n(Q). Since

ddtvk(Qt)=∑j=1n−1∂vk(Qt)∂sjΔsj,\frac{d}{dt}v_k(Q_t)=\sum_{j=1}^{n-1}\frac{\partial v_k(Q_t)}{\partial s_j}\Delta s_j,

integrating this identity and using (14), we obtain

0=vk(Q)−vk(P)vk=∑j=1n−1Mk,jΔsj.0=\frac{v_k(Q)-v_k(P)}{v_k}=\sum_{j=1}^{n-1}M_{k,j}\Delta s_j.

Since MM is invertible, the last equality implies Δsj=0\Delta s_j=0 for all 1≤j≤n−11\le j\le n-1. Recall also that Δsn=0\Delta s_n=0. Thus P=QP=Q, and their roots determine the same squared semiaxes, with multiplicities. □

AI Disclaimer. ChatGPT was used in exploratory discussions and in the preparation of this manuscript.

References

S. Myroshnychenko, Department of Mathematics and Statistics, University of the Fraser Valley, Abbotsford, BC V2S 7M8, Canada Email address: [email protected]

K. Tatarko, Department of Pure Mathematics, University of Waterloo, Waterloo, ON N2L 3G1, Canada Email address: [email protected]

V. Yaskin, Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton, AB T6G 2N8, Canada Email address: [email protected]

References

  1. [1]R. J. Gardner, Geometric tomography, 2nd edition, Encyclopedia Math. Appl., 58, Cambridge University Press, 2006.DOI
  2. [2]A. G usakova, E. Spodarev and D. Zaporozhets, Intrinsic volumes of ellipsoids, Zap. Nauchn. Sem. POMI 515 (2022), 121–140.
  3. [3]G. H. Hardy, J. E. Littlewood, and G. Pólya, Inequalities, 2nd ed., Cambridge University Press, Cambridge, 1952.
  4. [4]Z. Kabluchko, D. Zaporozhets, Intrinsic volumes of Sobolev balls with applications to Brownian convex hulls, Trans. Amer. Math. Soc. 368 (2016), 8873–8899.arxiv.org/abs/1404.6113
  5. [5]S. Myroshnychenko, K. Tatarko, and V. Yaskin, Unique determination of ellipsoids by their dual volumes, International Mathematics Research Notices 2022, no. 17 (2022), 13569–13589.DOI
  6. [6]F. Petrov and A. Tarasov, Uniqueness of a three-dimensional ellipsoid with given intrinsic volumes, Arnold Mathematical Journal 6 (2020), 163–171.DOI
  7. [7]R. Schneider, Convex Bodies: The Brunn–Minkowski Theory, second expanded edition, Cambridge University Press, 2014.

Paper details

vlnc:W_ypcsnwyk
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

Not recorded for this paper

View statement
Looking for related papers…
Contents