betaFeedback

WorksMathematics

The dimension-free symmetric Gaussian isoperimetric profile: finite attainment and critical rank

Abstract

We determine the infimum of Gaussian perimeter over centrally symmetric sets when the ambient dimension is allowed to vary. The answer is the lower envelope of the perimeters of centered balls and their complements. At every mass other than one half the infimum is attained in finite dimension, proving the finite-attainment conjecture stated by Heilman. Every attaining ball rank is at most 1/(2∣a−1/2∣2)1/(2|a-1/2|^2). At mass one half the value is 1/π1/\sqrt\pi and no finite-dimensional centrally symmetric set attains it. Every optimal ball rank satisfies k∣a−1/2∣2→1/(9π)k|a-1/2|^2\to1/(9\pi) as a→1/2a\to1/2. The profile has expansion π−1/2−(3π/2)∣a−1/2∣2+O(∣a−1/2∣4)\pi^{-1/2}-(3\sqrt\pi/2)|a-1/2|^2+O(|a-1/2|^4), and it yields the sharp even Gaussian Cheeger inequality. The geometric input is Viana's classification of isoperimetric regions in real projective spaces. The additional step is a stability estimate that controls the smaller of the two spherical ranks away from half volume and identifies the only regime in which both ranks can escape to infinity.

The profile and the main results

The symmetric Gaussian isoperimetric problem asks how central symmetry changes the least perimeter at a prescribed Gaussian volume. In this paper the volume is fixed but the ambient dimension may vary. We determine the resulting lower envelope, prove that a finite-dimensional ball or its complement attains it away from half volume, and describe the diverging optimal ball rank as the volume approaches one half.

Let γd\gamma_d be standard Gaussian probability measure on Rd\mathbb{R}^d, with density

ϕd(x)=(2π)−d/2e−∣x∣2/2.\phi_d(x) = (2\pi)^{-d/2}e^{-|x|^2/2}.

For a set of locally finite perimeter put

Pγd(E)=∫∂∗Eϕd(x) dHd−1(x).P_{\gamma_d}(E) = \int_{\partial^* E} \phi_d(x)\,d\mathcal{H}^{d-1}(x).

The perimeter is allowed to be infinite and is defined to be infinite for measurable sets which do not have locally finite perimeter. Symmetry always means central symmetry, E=−EE=-E up to a null set; neither convexity nor invariance under the separate coordinate reflections is assumed. Define

Id(a)=inf⁡{Pγd(E):E=−E, γd(E)=a}.(1)I_d(a) = \inf\{P_{\gamma_d}(E): E=-E,\ \gamma_d(E)=a\}. \tag*{(1)}
I∞(a)=inf⁡d≥1Id(a),0<a<1.(2)I_\infty(a) = \inf_{d\geq1} I_d(a), \qquad0<a<1. \tag*{(2)}

Our perimeter uses the ambient probability density on the boundary. The surface-area convention in [6, 7] is 2πPγd\sqrt{2\pi}P_{\gamma_d}.

For k≥1k\geq1 let

ρk(r)=21−k/2Γ(k/2)rk−1e−r2/2,Fk(r)=∫0rρk(t) dt,r>0.(3)\rho_k(r) = \frac{2^{1-k/2}}{\Gamma(k/2)}r^{k-1}e^{-r^2/2}, \qquad F_k(r) = \int_0^r \rho_k(t)\,dt,\qquad r>0. \tag*{(3)}

Write Rk(a)=Fk−1(a)R_k(a)=F_k^{-1}(a) and

jk(a)=ρk(Rk(a)).(4)j_k(a)=\rho_k(R_k(a)). \tag*{(4)}

Thus jk(a)j_k(a) is the Gaussian perimeter of the centered ball of mass aa in Rk\mathbb{R}^k. A product with a Gaussian Euclidean factor does not change either its mass or its perimeter. An attaining ball rank at mass aa is an integer kk for which jk(max⁡{a,1−a})=I∞(a)j_k(\max\{a,1-a\})=I_\infty(a). This is the dimension of the ball factor, not the ambient dimension of an arbitrary minimizing set.

Background and relation to previous work

The symmetric question originates in Barthe’s work [3]. His theorem for unconditional sets concerns an anisotropic perimeter; the present paper uses the usual Euclidean Gaussian perimeter. For that perimeter, Barchiesi and Julin [2] proved that strips minimize at volumes sufficiently close to one, with a threshold independent of dimension. Heilman [6, 7] obtained partial classification results through Gaussian variations. La Manna’s local stability threshold for balls [8] will enter our quantitative rank bound through an explicit second-variation calculation.

The question of determining (1.2) is stated explicitly in [7 Section 1.3], together with the conjecture that its value is realized in a finite dimension at every mass different from 1/21/2. The half-mass lower bound 1/π1/\sqrt{\pi} is proposed in the introductory discussion of [6]. The results below address these dimension-free questions. Recent work of Barchiesi, Cagnetti, and Julin [1] develops a double-arc rearrangement for unconditional sets and discusses a route through high-dimensional spheres. Milman [10] resolves the fixed-dimensional symmetric problem in the plane. Our result concerns the infimum over all finite dimensions and does not require a classification in each prescribed dimension.

Theorem 1.1 (Exact profile and finite attainment). For 0<a<10<a<1, put m=max⁡{a,1−a}m=\max\{a,1-a\}. Then

I∞(a)=inf⁡k≥1min⁡{jk(a),jk(1−a)}=inf⁡k≥1jk(m).(5)I_\infty(a)=\inf_{k\geq1}\min\{j_k(a),j_k(1-a)\}=\inf_{k\geq1}j_k(m). \tag*{(5)}

If a≠1/2a\ne1/2, these infima are minima. A centered ball of mass aa when a>1/2a>1/2, and the complement of a centered ball of mass 1−a1-a when a<1/2a<1/2, realizes I∞(a)I_\infty(a) in a finite dimension. Every integer kk with jk(m)=I∞(a)j_k(m)=I_\infty(a) satisfies

k≤12∣a−1/2∣2.(6)k\leq\frac{1}{2|a-1/2|^2}. \tag*{(6)}

Consequently,

Id(a)=I∞(a)ifd≥⌈12∣a−1/2∣2⌉,(7)I_d(a)=I_\infty(a)\quad\text{if}\quad d\geq\left\lceil\frac{1}{2|a-1/2|^2}\right\rceil, \tag*{(7)}

and a round cylinder or its complement realizes this value. At the exceptional mass,

I∞(1/2)=1π.(8)I_\infty(1/2)=\frac{1}{\sqrt{\pi}}. \tag*{(8)}

This value is approached by centered half-mass balls as their dimension tends to infinity, but it is not realized by any finite-dimensional centrally symmetric set.

Theorem 1.2 (Critical rank). Let δ>0\delta>0 tend to zero and choose any attaining ball rank k=k(δ)k=k(\delta) in I∞(1/2+δ)=jk(1/2+δ)I_\infty(1/2+\delta)=j_k(1/2+\delta). Then

k(δ)δ2⟶19π.(9)k(\delta)\delta^2\longrightarrow\frac{1}{9\pi}. \tag*{(9)}

Moreover,

I∞(1/2+δ)=1π−3π2δ2+O(δ4).(10)I_\infty(1/2+\delta)=\frac{1}{\sqrt{\pi}}-\frac{3\sqrt{\pi}}{2}\delta^2+O(\delta^4). \tag*{(10)}

By symmetry the same expansion holds on the other side of 1/21/2. In particular, I∞I_\infty is differentiable at 1/21/2, with derivative zero.

Corollary 1.3 (Sharp even Gaussian Cheeger inequality). Every centrally symmetric set of finite Gaussian perimeter satisfies

Pγd(E)≥2πmin⁡{γd(E),1−γd(E)}.(11)P_{\gamma_d}(E) \ge\frac{2}{\sqrt{\pi}}\min\{\gamma_d(E),1-\gamma_d(E)\}. \tag*{(11)}

For every even Lipschitz function f:Rd→Rf:\mathbb{R}^d\to\mathbb{R},

inf⁡c∈R∫∣f−c∣ dγd≤π2∫∣∇f∣ dγd.(12)\inf_{c\in\mathbb{R}}\int|f-c|\,\mathrm{d}\gamma_d \le\frac{\sqrt{\pi}}{2}\int|\nabla f|\,\mathrm{d}\gamma_d. \tag*{(12)}

Both constants are optimal when all finite dimensions are considered. The constant cc on the left of (1.12) can be any median of ff.

Remark 1.4 (Scope). Theorem 1.1 determines the profile with the dimension unrestricted and proves finite attainment away from half mass. It does not assert the formula

Id(a)=min⁡1≤k≤dmin⁡{jk(a),jk(1−a)}I_d(a)=\min_{1\le k\le d}\min\{j_k(a),j_k(1-a)\}

for every prescribed smaller dimension dd, and it does not classify every equality case. Those fixed-dimension assertions are stronger. The results here also do not assert that every intermediate integer occurs as an optimal rank. In particular, (1.9) describes the ranks in the ball representation of the dimension-free profile, not a classification of all minimizers or a formula for their least possible ambient dimension.

Geometric input and proof strategy

Viana’s theorem [11 Corollary 1.3] classifies antipodally invariant isoperimetric regions on a sphere as thresholds of a squared coordinate projection. Their masses are beta probabilities. We first use their second variation to locate the threshold within one unit of the first block dimension. A two-block central limit argument then shows that both block dimensions can diverge only at mass one half. Fixed ranks converge to Gaussian balls, while the escaping ranks have perimeter tending to 1/π1/\sqrt{\pi}. We transfer this profile limit to arbitrary Gaussian sets using the exact perimeter formula for a spherical lift.

The related Poincaré-limit strategy and the role of the smaller spherical rank are discussed in [1 Appendix B, Remark 8]. The rank estimate proved below is what makes that limiting argument compact away from half mass. The self-shrinker classification from [7 Theorem 6.1] is used only in the last step excluding finite-dimensional attainment at half mass; it is not needed to compute the profile. More precisely, Sections 2–4 determine the profile and prove finite attainment. Section 5 proves the quantitative bound. Section 6 obtains the critical expansion without assuming nonattainment. Its differentiability consequence then supplies the zero-multiplier hypothesis needed in Section 7.

Spherical profiles and stability

Throughout the spherical argument N≥4N\ge4 is even. This subsequence is sufficient for the Gaussian limit and permits an exactly balanced test at half mass. Let σN\sigma_N be normalized surface measure on SN−1(N)⊂RNS^{N-1}(\sqrt{N})\subset\mathbb{R}^N. Spherical perimeter is the boundary area divided by the total area of this sphere, with the metric of radius N\sqrt{N}. Denote the antipodal isoperimetric profile by

IN(a)=inf⁡{PσN(A):A=−A, σN(A)=a}.I_N(a)=\inf\{P_{\sigma_N}(A):A=-A,\ \sigma_N(A)=a\}.

Proposition 2.1 (Viana’s classification in probability coordinates). For 1≤r≤N−11\le r\le N-1, write s=N−rs=N-r and let t=tN,r(a)t=t_{N,r}(a) be the aa-quantile of Beta⁡(r/2,s/2)\operatorname{Beta}(r/2,s/2). Then

IN(a)=min⁡1≤r≤N−1JN,r(a),JN,r(a)=2Nt(1−t)tr/2−1(1−t)s/2−1B(r/2,s/2).(13)I_N(a)=\min_{1\le r\le N-1}J_{N,r}(a),\qquad J_{N,r}(a)=\frac{2}{\sqrt{N}}\sqrt{t(1-t)}\frac{t^{r/2-1}(1-t)^{s/2-1}}{B(r/2,s/2)}. \tag*{(13)}

The corresponding region is, after rotation,

AN,r,t={x∈SN−1(N):x12+⋯+xr2N≤t}.(14)A_{N,r,t}=\left\{x\in S^{N-1}(\sqrt{N}):\frac{x_1^2+\cdots+x_r^2}{N}\le t\right\}. \tag*{(14)}

Proof. The quotient map to real projective space is a two-sheeted local isometry. An antipodally invariant finite-perimeter set descends to the quotient; its volume, boundary area, and the total ambient volume are all divided by two. Thus the normalized isoperimetric problems agree. Apply [11 Corollary 1.3 and Section 4.1]; the latter section includes finite-perimeter regions and the possible singular set. Rescaling the metric to radius N\sqrt{N} preserves the classification and rescales normalized perimeter by N−1/2N^{-1/2}. Complements can be put in the form (14) by interchanging the two coordinate blocks. For a uniform point on the sphere the variable T=(x12+⋯+xr2)/NT=(x_1^2+\cdots+x_r^2)/N has the indicated beta distribution. Its tangent gradient has length 2T(1−T)/N2\sqrt{T(1-T)}/\sqrt{N}. Coarea gives (13). □

Lemma 2.2 (Uniform continuity of the spherical profiles). The functions INI_N are concave and symmetric about 1/21/2, and

0≤IN(a)≤2/π.(15)0\le I_N(a)\le\sqrt{2/\pi}. \tag*{(15)}

They have a common Lipschitz bound on every compact subinterval of (0,1)(0,1).

Proof. Put θ=arcsin⁡T\theta=\arcsin\sqrt{T}. Its density on (0,π/2)(0,\pi/2) is

g(θ)=2B(r/2,s/2)sin⁡r−1θcos⁡s−1θ.g(\theta)=\frac{2}{B(r/2,s/2)}\sin^{r-1}\theta\cos^{s-1}\theta.

The function log⁡g\log g is concave. If GG is its distribution function, then a↦g(G−1(a))a\mapsto g(G^{-1}(a)) is concave, since its second derivative is (log⁡g)′′(G−1(a))/g(G−1(a))(\log g)''(G^{-1}(a))/g(G^{-1}(a)). This function divided by N\sqrt{N} is JN,rJ_{N,r}. The infimum of concave functions is concave, so (13) proves concavity of INI_N. Complementation gives its symmetry.

For r=1r=1, the maximum of gg is 2/B(1/2,(N−1)/2)2/B(1/2,(N-1)/2). The gamma-ratio estimate

Γ(x+1/2)Γ(x)≤x,x>0,\frac{\Gamma(x+1/2)}{\Gamma(x)}\le\sqrt{x},\qquad x>0,

follows from Cauchy–Schwarz applied to a gamma random variable. It yields

IN(a)≤2Γ(N/2)πNΓ((N−1)/2)<2/π.I_N(a)\le\frac{2\Gamma(N/2)}{\sqrt{\pi N}\Gamma((N-1)/2)}<\sqrt{2/\pi}.

The lower envelope tends to zero at both endpoints: use r=N−1r=N-1 at zero and r=1r=1 at one. A nonnegative concave function bounded by a fixed constant and vanishing at the endpoints has a uniformly bounded slope on [ε,1−ε][\varepsilon,1-\varepsilon]. This proves the last assertion. □

Lemma 2.3 (Stable thresholds). Suppose (14) minimizes the antipodal perimeter problem and r,s≥2r,s\ge2. Then

r−1N≤t≤r+1N.(16)\frac{r-1}{N}\le t\le\frac{r+1}{N}. \tag*{(16)}

Proof. Write Σ=∂AN,r,t\Sigma=\partial A_{N,r,t}. The boundary is Sr−1(Nt)×Ss−1(N(1−t))S^{r-1}(\sqrt{Nt})\times S^{s-1}(\sqrt{N(1-t)}). Up to the common choice of orientation, its principal curvatures are (1−t)/(Nt)\sqrt{(1-t)/(Nt)} with multiplicity r−1r-1, and −t/(N(1−t))-\sqrt{t/(N(1-t))} with multiplicity s−1s-1. Since the ambient Ricci curvature in a unit direction is (N−2)/N(N-2)/N, the Jacobi potential is

∣A∣2+Ric⁡(ν,ν)=1N(r−1t+s−11−t).(17)|A|^2+\operatorname{Ric}(\nu,\nu)=\frac{1}{N}\left(\frac{r-1}{t}+\frac{s-1}{1-t}\right). \tag*{(17)}

For an even smooth normal speed ff with ∫Σf=0\int_{\Sigma} f=0, stability gives

QΣ(f)=∫Σ(∣∇Σf∣2−(∣A∣2+Ric⁡(ν,ν))f2)dHN−2≥0.Q_{\Sigma}(f)=\int_{\Sigma}\left(|\nabla_{\Sigma}f|^2-\left(|A|^2+\operatorname{Ric}(\nu,\nu)\right)f^2\right)d\mathcal{H}^{N-2}\geq0.

Indeed, ν(−x)=−ν(x)\nu(-x)=-\nu(x), so an even normal speed produces an antipodal normal graph. Adding a constant normal displacement of order two enforces the volume constraint exactly: the derivative of volume in the constant direction is the positive boundary area, so the implicit function theorem applies. This correction does not alter the constrained quadratic form above.

Take a nonzero degree-two spherical harmonic on the first factor and make it constant on the second. It is even and has zero mean on the product, and its Laplace eigenvalue is 2r/(Nt)2r/(Nt). Substitution in QΣ(f)≥0Q_{\Sigma}(f)\geq0 gives

r−1t+s−11−t≤2rt.\frac{r-1}{t}+\frac{s-1}{1-t}\leq\frac{2r}{t}.

which is Nt≤r+1Nt\leq r+1. A degree-two harmonic in the second block gives Nt≥r−1Nt\geq r-1. These harmonics exist precisely in the range needed here, r,s≥2r,s\geq2. No pinching claim is made when a block has dimension one. □

Lemma 2.4 (Escape of both ranks forces half mass). Let rN+sN=Nr_N+s_N=N and rN,sN→∞r_N,s_N\to\infty. If ∣NtN−rN∣≤1|Nt_N-r_N|\leq1, then

P{Beta⁡(rN/2,sN/2)≤tN}⟶12.(18)\mathbb{P}\left\{\operatorname{Beta}\left(r_N/2,s_N/2\right)\leq t_N\right\}\longrightarrow\frac{1}{2}. \tag*{(18)}

Consequently, at each fixed mass a≠1/2a\ne1/2, the smaller rank of any spherical isoperimetric minimizer stays bounded as N→∞N\to\infty.

Proof. Suppress the subscript NN. Let UU and VV be independent chi-square variables of degrees rr and ss. Then T=U/(U+V)T=U/(U+V) has the beta law. Set D=2rsND=\sqrt{2rsN} and

W=s(U−r)−r(V−s)D=sNU−r2r−rNV−s2s.W=\frac{s(U-r)-r(V-s)}{D}=\sqrt{\frac{s}{N}}\frac{U-r}{\sqrt{2r}}-\sqrt{\frac{r}{N}}\frac{V-s}{\sqrt{2s}}.

The two standardized variables converge to independent standard normals. Along any subsequence the two coefficients have a further convergent subsequence, and their squared sum is one. Thus WW converges to a standard normal without a restriction on r/Nr/N. Also

EU+VD=12r+12s⟶0.\mathbb{E}\frac{U+V}{D}=\sqrt{\frac{1}{2r}+\frac{1}{2s}}\longrightarrow0.

For c=±1c=\pm1 there is the exact equivalence

T≤r+cN⟺W≤cU+VD.T\leq\frac{r+c}{N}\Longleftrightarrow W\leq c\frac{U+V}{D}.

The probabilities at both endpoints tend to 1/21/2, by convergence in distribution of WW and convergence in probability of the right-hand threshold to zero. Squeezing proves (18).

If the smaller minimizing rank were unbounded at a fixed nonhalf mass, a subsequence would have both ranks tending to infinity. Lemma 2.3 would then contradict (18). A rank equal to one already has bounded smaller rank and needs no stability estimate. □

The two profile limits

Lemma 3.1 (Fixed-rank and escaping-rank limits). For every fixed k≥1k\geq1,

JN,k(a)⟶jk(a)(19)J_{N,k}(a)\longrightarrow j_k(a) \tag*{(19)}

locally uniformly for a∈(0,1)a\in(0,1). If r+s=Nr+s=N, both rr and ss tend to infinity, and ∣Nt−r∣≤1|Nt-r|\leq1, then the perimeter of (14) tends to 1/π1/\sqrt{\pi}.

Proof. For fixed kk, the density of X=NTX = NT, with T∼Beta⁡(k/2,(N−k)/2)T \sim\operatorname{Beta}(k/2,(N-k)/2), is

Γ(N/2)N−k/2Γ(k/2)Γ((N−k)/2)xk/2−1(1−x/N)(N−k)/2−1,0<x<N.\frac{\Gamma(N/2)N^{-k/2}}{\Gamma(k/2)\Gamma((N-k)/2)}x^{k/2-1}(1-x/N)^{(N-k)/2-1}, \qquad0<x<N.

Denote this density by fN,kf_{N,k}. The gamma-ratio asymptotic and (1−x/N)(N−k)/2−1→e−x/2(1-x/N)^{(N-k)/2-1}\to e^{-x/2} show that it converges locally uniformly on (0,∞)(0,\infty) to the chi-square density fkf_k. For convergence in distribution, write X=U/((U+V)/N)X=U/((U+V)/N) with independent U∼χk2U\sim\chi_k^2 and V∼χN−k2V\sim\chi_{N-k}^2. The denominator tends to one in probability, so X⇒χk2X\Rightarrow\chi_k^2. The limiting distribution function is continuous and strictly increasing; hence its interior quantiles converge, uniformly on each compact mass interval. In particular, if xN,k(a)=NtN,k(a)x_{N,k}(a)=Nt_{N,k}(a), then

xN,k(a)⟶Rk(a)2,JN,k(a)=2xN,k(a)1−xN,k(a)/N fN,k(xN,k(a)).x_{N,k}(a)\longrightarrow R_k(a)^2,\qquad J_{N,k}(a)=2\sqrt{x_{N,k}(a)}\sqrt{1-x_{N,k}(a)/N}\,f_{N,k}(x_{N,k}(a)).

The quantiles stay in a compact subset of (0,∞)(0,\infty) on such a mass interval. Passing to the limit in this formula, and using 2rfk(r2)=ρk(r)2rf_k(r^2)=\rho_k(r), proves (3.1) with the asserted uniformity.

For the second assertion put p=r/Np=r/N. At t=pt=p the perimeter equals

2Np(r−1)/2(1−p)(s−1)/2B(r/2,s/2).(20)\frac{2}{\sqrt{N}}\frac{p^{(r-1)/2}(1-p)^{(s-1)/2}}{B(r/2,s/2)}. \tag*{(20)}

Stirling’s formula gives, uniformly whenever both rr and ss tend to infinity,

B(r/2,s/2)=2πr(r−1)/2s(s−1)/2N(N−1)(N−1)/2exp⁡(O(1/r+1/s)).B(r/2,s/2)=2\sqrt{\pi}\frac{r^{(r-1)/2}s^{(s-1)/2}}{N(N-1)^{(N-1)/2}}\exp(O(1/r+1/s)).

All powers therefore cancel in (3.2), leaving π−1/2exp⁡(O(1/r+1/s))\pi^{-1/2}\exp(O(1/r+1/s)). For completeness, the derivative of the logarithm of the numerator, at pp, is

N2(1s−1r).\frac{N}{2}\left(\frac{1}{s}-\frac{1}{r}\right).

and its second derivative between pp and a point at distance at most 1/N1/N is O(N2(1/r+1/s))O(N^2(1/r+1/s)). Moving tt within the pinching interval therefore changes the logarithm by O(1/r+1/s)O(1/r+1/s). These estimates remain valid when p→0p\to0 or p→1p\to1, since the relative shift in each block is at most the reciprocal of its rank. □

Lemma 3.2 (A strict half-mass gap in every finite rank). For every k≥1k\ge1,

jk(1/2)>e1/(18k)π>1π,lim⁡k→∞jk(1/2)=1π.(21)j_k(1/2)>\frac{e^{1/(18k)}}{\sqrt{\pi}}>\frac{1}{\sqrt{\pi}},\qquad\lim_{k\to\infty}j_k(1/2)=\frac{1}{\sqrt{\pi}}. \tag*{(21)}

Moreover, as k→∞k\to\infty,

Rk(1/2)2=k−23+O(k−1),jk(1/2)=1π(1+118k+O(k−2)).(22)R_k(1/2)^2=k-\frac{2}{3}+O(k^{-1}),\qquad j_k(1/2)=\frac{1}{\sqrt{\pi}}\left(1+\frac{1}{18k}+O(k^{-2})\right). \tag*{(22)}

Proof. Let M(x)M(x) be the median of a gamma variable of shape xx and scale one. The Chen–Rubin lower bound [5] and [4 Proposition 3.6] give

x−13<M(x)<xexp⁡(−13x),x>0.x-\frac{1}{3}<M(x)<x\exp\left(-\frac{1}{3x}\right),\qquad x>0.

For the upper estimate, the cited proposition in the published version of [4] states that xlog⁡(x/M(x))x\log(x/M(x)) decreases to 1/31/3, which yields the displayed strict bound. Hence, writing uk=Rk(1/2)2u_k=R_k(1/2)^2,

k−23<uk<ke−2/(3k)=:vk.(23)k-\frac{2}{3}<u_k<ke^{-2/(3k)}=:v_k. \tag*{(23)}

The chi density decreases above its mode, whose square is k−1k-1; this includes k=1k=1, for which the mode is zero. Both quantities in (3.5) lie above that mode. The upper Stirling bound

Γ(x)<2π xx−1/2e−x+1/(12x)\Gamma(x)<\sqrt{2\pi}\,x^{x-1/2}e^{-x+1/(12x)}

therefore gives

log⁡(π jk(1/2))>log⁡(π ρk(uk))>−k−13k+k2(1−e−2/(3k))−16k>118k.\log(\sqrt{\pi}\,j_k(1/2))>\log(\sqrt{\pi}\,\rho_k(\sqrt{u_k})) >-\frac{k-1}{3k}+\frac{k}{2}\left(1-e^{-2/(3k)}\right)-\frac{1}{6k} >\frac{1}{18k}.

The last step uses 1−e−x>x−x2/21-e^{-x}>x-x^2/2 for x>0x>0. The two median bounds also imply the first expansion in (3.4). Put b=uk−k=−2/3+O(k−1)b=u_k-k=-2/3+O(k^{-1}). The density formula and Stirling’s expansion give explicitly

log⁡(π jk(1/2))=−16k+k−12log⁡(1+b/k)−b2+O(k−3)=118k+O(k−2),\log(\sqrt{\pi}\,j_k(1/2))=-\frac{1}{6k}+\frac{k-1}{2}\log(1+b/k)-\frac{b}{2}+O(k^{-3}) =\frac{1}{18k}+O(k^{-2}),

which proves the remaining assertions. □

Define for now

C(a)=inf⁡k≥1min⁡{jk(a),jk(1−a)}.(24)C(a)=\inf_{k\geq1}\min\{j_k(a),j_k(1-a)\}. \tag*{(24)}

We set C(0)=C(1)=0C(0)=C(1)=0. This gives continuity at the endpoints, since 0≤C(a)≤j1(1−a)0\leq C(a)\leq j_1(1-a) near zero and 0≤C(a)≤j1(a)0\leq C(a)\leq j_1(a) near one. Interior continuity will follow from the proposition below.

Proposition 3.3 (Limit of the spherical profiles). As even N→∞N\to\infty,

IN(a)⟶C(a)(25)I_N(a)\longrightarrow C(a) \tag*{(25)}

locally uniformly on (0,1)(0,1). The function CC is concave and symmetric. If a≠1/2a\neq1/2, the infimum in (3.6) is attained by some finite rank. At a=1/2a=1/2, C(a)=1/πC(a)=1/\sqrt{\pi}.

Proof. Fix a≠1/2a\neq1/2 and choose a minimizing rank for each sphere. Lemma 2.4 bounds its smaller rank. On a subsequence this rank is a fixed integer kk, and one of the two block orientations is fixed. Lemma 3.1 then identifies the limiting perimeter as jk(a)j_k(a) or jk(1−a)j_k(1-a), which is at least C(a)C(a). Conversely, testing any fixed rank in either orientation gives

lim sup⁡N→∞IN(a)≤inf⁡k≥1min⁡{jk(a),jk(1−a)}=C(a).\limsup_{N\to\infty} I_N(a)\leq\inf_{k\geq1}\min\{j_k(a),j_k(1-a)\}=C(a).

Every subsequential limit is therefore C(a)C(a), and a fixed rank on a subsequence realizes this value. At a=1/2a=1/2, the test r=s=N/2r=s=N/2 has quantile exactly t=1/2t=1/2 and its perimeter tends to 1/π1/\sqrt{\pi} by Stirling’s formula. A bounded smaller minimizing rank on any subsequence would instead have a limit jk(1/2)>1/πj_k(1/2)>1/\sqrt{\pi}, by Lemma 3.2. Thus both minimizing ranks tend to infinity. Lemmas 2.3 and 3.1 give the matching limit 1/π1/\sqrt{\pi}. This is also C(1/2)C(1/2) by Lemma 3.2.

Finally Lemma 2.2 supplies equicontinuity on each compact mass interval. The pointwise limit is consequently continuous and the convergence is locally uniform. Concavity and symmetry pass to the limit. □

Transfer to arbitrary Gaussian sets

Lemma 4.1 (Exact perimeter of a spherical lift). Let E⊂RdE \subset\mathbb{R}^d be bounded with smooth boundary, and let

E^N={x∈SN−1(N):(x1,…,xd)∈E}.\widehat{E}_N=\{x\in S^{N-1}(\sqrt{N}):(x_1,\ldots,x_d)\in E\}.

For all sufficiently large NN,

σN(E^N)=∫EwN,d(x) dx,(26)\sigma_N(\widehat{E}_N)=\int_E w_{N,d}(x)\,dx, \tag*{(26)}
PσN(E^N)=∫∂EwN,d(x)1−⟨x,νE(x)⟩2N dHd−1(x),(27)P_{\sigma_N}(\widehat{E}_N)=\int_{\partial E}w_{N,d}(x)\sqrt{1-\frac{\langle x,\nu_E(x)\rangle^2}{N}}\,d\mathcal{H}^{d-1}(x), \tag*{(27)}

where

wN,d(x)=Γ(N/2)Γ((N−d)/2)(πN)d/2(1−∣x∣2N)(N−d−2)/2,∣x∣<N.(28)w_{N,d}(x)=\frac{\Gamma(N/2)}{\Gamma((N-d)/2)(\pi N)^{d/2}}\left(1-\frac{|x|^2}{N}\right)^{(N-d-2)/2},\qquad|x|<\sqrt{N}. \tag*{(28)}

In particular, the mass and perimeter in (26)–(27) converge to γd(E)\gamma_d(E) and Pγd(E)P_{\gamma_d}(E).

Proof. The projection of normalized spherical measure onto the first dd coordinates has density (28), which gives (26). Let gg be a signed distance function to ∂E\partial E near the boundary. On the sphere its pullback has tangent gradient of squared length

∣∇g∣2−⟨x,∇g⟩2N=1−⟨x,νE⟩2N.|\nabla g|^2-\frac{\langle x,\nabla g\rangle^2}{N}=1-\frac{\langle x,\nu_E\rangle^2}{N}.

For clarity, let ηε\eta_\varepsilon be an approximate identity on the real line, supported near zero, and let πd\pi_d be projection onto the first dd coordinates. Coarea on the sphere expresses the perimeter of the lift as the limit of

∫SN−1(N)ηε(g(πdy))∣∇S(g∘πd)(y)∣ dσN(y).\int_{S^{N-1}(\sqrt{N})}\eta_\varepsilon(g(\pi_dy))|\nabla_S(g\circ\pi_d)(y)|\,d\sigma_N(y).

Disintegration replaces this integral by the integral over Rd\mathbb{R}^d against wN,d(x) dxw_{N,d}(x)\,dx, with the same square-root gradient factor. Euclidean coarea for the signed distance gg then gives (27) as ε↓0\varepsilon\downarrow0. Since EE is bounded and NN is large, the projection is a submersion near this boundary and has no additional boundary contribution there. On each fixed compact set, wN,d→ϕdw_{N,d}\to\phi_d uniformly and the square-root factor tends uniformly to one. The asserted limits follow. □

Lemma 4.2 (Smooth approximation with central symmetry). If E⊂RdE\subset\mathbb{R}^d is centrally symmetric and Pγd(E)<∞P_{\gamma_d}(E)<\infty, there are bounded open centrally symmetric sets EjE_j with smooth boundary such that

γd(Ej△E)⟶0,Pγd(Ej)⟶Pγd(E).\gamma_d(E_j\mathbin{\triangle}E)\longrightarrow0,\qquad P_{\gamma_d}(E_j)\longrightarrow P_{\gamma_d}(E).

The approximating volumes need not equal γd(E)\gamma_d(E) exactly.

Proof. We give the weighted and symmetric details of the usual strict finite-perimeter approximation; see [9] for the underlying BVBV approximation and coarea theorems. For almost every R>0R>0,

Pγd(E∩BR)≤Pγd(E;BR)+Pγd(BR).P_{\gamma_d}(E\cap B_R)\leq P_{\gamma_d}(E;B_R)+P_{\gamma_d}(B_R).

The last term is a constant times Rd−1e−R2/2R^{d-1}e^{-R^2/2}, hence tends to zero. Together with weighted lower semicontinuity, this proves strict convergence of the symmetric truncations as R→∞R\to\infty. It remains to approximate a bounded symmetric set AA.

Let uε=1A∗ζεu_\varepsilon=1_A*\zeta_\varepsilon, where ζε\zeta_\varepsilon is nonnegative, smooth, compactly supported, even, and has integral one. Then 0≤uε≤10\leq u_\varepsilon\leq1, uε(−x)=uε(x)u_\varepsilon(-x)=u_\varepsilon(x), and its support is bounded. Convolution and Fubini yield

∫ϕd∣∇uε∣ dx≤∫(ϕd∗ζε) d∣D1A∣⟶Pγd(A).\int\phi_d|\nabla u_\varepsilon|\,dx\leq\int(\phi_d*\zeta_\varepsilon)\,d|D1_A|\longrightarrow P_{\gamma_d}(A).

Here AA has finite ordinary perimeter because the weight is bounded away from zero on its compact support. Weighted lower semicontinuity gives the reverse limiting inequality, so the displayed weighted variations converge.

Fix 0<η<1/20 < \eta< 1/2. For every t∈[η,1−η]t \in[\eta, 1 - \eta],

γd({uε>t}△A)≤η−1∥uε−1A∥L1(γd).\gamma_d(\{u_\varepsilon> t\} \mathbin{\triangle} A) \leq\eta^{-1}\lVert u_\varepsilon- 1_A\rVert_{L^1(\gamma_d)}.

Weighted coarea and Sard’s theorem allow a regular value tε∈[η,1−η]t_\varepsilon\in[\eta, 1 - \eta] with

Pγd({uε>tε})≤∫ϕd∣∇uε∣ dx1−2η+ε.P_{\gamma_d}(\{u_\varepsilon> t_\varepsilon\}) \leq\frac{\int\phi_d\lvert\nabla u_\varepsilon\rvert\,dx}{1-2\eta}+\varepsilon.

These superlevel sets are smooth, bounded, and centrally symmetric. Let η\eta decrease to zero, choose ε\varepsilon so that the L1L^1 error is o(η)o(\eta), and use lower semicontinuity again. This produces strict approximation of AA. A final diagonal choice over the truncations proves the lemma. □

Proposition 4.3 (Gaussian lower bound and exact profile). For every centrally symmetric Gaussian finite-perimeter set,

Pγd(E)≥C(γd(E)).(29)P_{\gamma_d}(E) \geq C(\gamma_d(E)). \tag*{(29)}

Consequently I∞=CI_\infty= C.

Proof. The assertion at mass zero or one is immediate. For interior mass, first suppose EE is bounded and smooth. Its lift is antipodally invariant, so

PσN(E^N)≥IN(σN(E^N)).P_{\sigma_N}(\widehat{E}_N) \geq I_N(\sigma_N(\widehat{E}_N)).

Write aN=σN(E^N)a_N = \sigma_N(\widehat{E}_N) and a=γd(E)a = \gamma_d(E). By Lemma 4.1, aN→aa_N \to a, so these masses belong to a fixed compact interval K⊂(0,1)K \subset(0,1) for all large NN. In particular,

∣IN(aN)−C(a)∣≤sup⁡b∈K∣IN(b)−C(b)∣+∣C(aN)−C(a)∣⟶0.\lvert I_N(a_N)-C(a)\rvert\leq\sup_{b\in K}\lvert I_N(b)-C(b)\rvert+\lvert C(a_N)-C(a)\rvert\longrightarrow0.

by Proposition 3.3. Taking the perimeter limit now proves (4.4) for smooth bounded sets. For a general set of finite Gaussian perimeter, apply Lemma 4.2 and continuity of CC. This also explains why neither the spherical lift nor the smooth approximation has to preserve the mass exactly. Infinite perimeter causes no difficulty.

Thus I∞(a)≥C(a)I_\infty(a) \geq C(a). Centered balls of mass aa in every rank, and complements of centered balls of mass 1−a1-a, give the reverse inequality in (3.6). At half mass their limiting value is provided by Lemma 3.2. Hence I∞=CI_\infty= C at all interior masses. □

Orientation and a quantitative rank bound

Lemma 5.1 (The larger-mass ball is preferred). For k≥1k \geq1 and 0<a<1/20 < a < 1/2,

jk(a)>jk(1−a).(30)j_k(a) > j_k(1-a). \tag*{(30)}

Proof. For k=1k=1, the density ρ1\rho_1 is strictly decreasing. Suppose k>1k>1 and put

V(r)=r22−(k−1)log⁡r.V(r)=\frac{r^2}{2}-(k-1)\log r.

It is strictly convex, has its minimum at r0=k−1r_0=\sqrt{k-1}, and V′′V'' is strictly decreasing. For v>V(r0)v>V(r_0), let r−(v)<r0<r+(v)r_-(v)<r_0<r_+(v) be the two inverse branches. Define p−(v)=−V′(r−(v))p_-(v)=-V'(r_-(v)) and p+(v)=V′(r+(v))p_+(v)=V'(r_+(v)). Both vanish at the minimum level, and

ddvp±(v)2=2V′′(r±(v)).\frac{d}{dv}p_\pm(v)^2=2V''(r_\pm(v)).

@page 9/14 starts_mid=1 ends_mid=1 Here AA has finite ordinary perimeter because the weight is bounded away from zero on its compact support. Weighted lower semicontinuity gives the reverse limiting inequality, so the displayed weighted variations converge.

Fix 0<η<1/20 < \eta< 1/2. For every t∈[η,1−η]t \in[\eta, 1 - \eta],

γd({uε>t}△A)≤η−1∥uε−1A∥L1(γd).\gamma_d(\{u_\varepsilon> t\} \mathbin{\triangle} A) \leq\eta^{-1}\lVert u_\varepsilon- 1_A\rVert_{L^1(\gamma_d)}.

Weighted coarea and Sard’s theorem allow a regular value tε∈[η,1−η]t_\varepsilon\in[\eta, 1 - \eta] with

Pγd({uε>tε})≤∫ϕd∣∇uε∣ dx1−2η+ε.P_{\gamma_d}(\{u_\varepsilon> t_\varepsilon\}) \leq\frac{\int\phi_d\lvert\nabla u_\varepsilon\rvert\,dx}{1-2\eta}+\varepsilon.

These superlevel sets are smooth, bounded, and centrally symmetric. Let η\eta decrease to zero, choose ε\varepsilon so that the L1L^1 error is o(η)o(\eta), and use lower semicontinuity again. This produces strict approximation of AA. A final diagonal choice over the truncations proves the lemma. □

Proposition 4.3 (Gaussian lower bound and exact profile). For every centrally symmetric Gaussian finite-perimeter set,

Pγd(E)≥C(γd(E)).(31)P_{\gamma_d}(E) \geq C(\gamma_d(E)). \tag*{(31)}

Consequently I∞=CI_\infty= C.

Proof. The assertion at mass zero or one is immediate. For interior mass, first suppose EE is bounded and smooth. Its lift is antipodally invariant, so

PσN(E^N)≥IN(σN(E^N)).P_{\sigma_N}(\widehat{E}_N) \geq I_N(\sigma_N(\widehat{E}_N)).

Write aN=σN(E^N)a_N = \sigma_N(\widehat{E}_N) and a=γd(E)a = \gamma_d(E). By Lemma 4.1, aN→aa_N \to a, so these masses belong to a fixed compact interval K⊂(0,1)K \subset(0,1) for all large NN. In particular,

∣IN(aN)−C(a)∣≤sup⁡b∈K∣IN(b)−C(b)∣+∣C(aN)−C(a)∣⟶0.\lvert I_N(a_N)-C(a)\rvert\leq\sup_{b\in K}\lvert I_N(b)-C(b)\rvert+\lvert C(a_N)-C(a)\rvert\longrightarrow0.

by Proposition 3.3. Taking the perimeter limit now proves (4.4) for smooth bounded sets. For a general set of finite Gaussian perimeter, apply Lemma 4.2 and continuity of CC. This also explains why neither the spherical lift nor the smooth approximation has to preserve the mass exactly. Infinite perimeter causes no difficulty.

Thus I∞(a)≥C(a)I_\infty(a) \geq C(a). Centered balls of mass aa in every rank, and complements of centered balls of mass 1−a1-a, give the reverse inequality in (3.6). At half mass their limiting value is provided by Lemma 3.2. Hence I∞=CI_\infty= C at all interior masses. □

Orientation and a quantitative rank bound

Lemma 5.1 (The larger-mass ball is preferred). For k≥1k \geq1 and 0<a<1/20 < a < 1/2,

jk(a)>jk(1−a).(32)j_k(a) > j_k(1-a). \tag*{(32)}

Proof. For k=1k=1, the density ρ1\rho_1 is strictly decreasing. Suppose k>1k>1 and put

V(r)=r22−(k−1)log⁡r.V(r)=\frac{r^2}{2}-(k-1)\log r.

It is strictly convex, has its minimum at r0=k−1r_0=\sqrt{k-1}, and V′′V'' is strictly decreasing. For v>V(r0)v>V(r_0), let r−(v)<r0<r+(v)r_-(v)<r_0<r_+(v) be the two inverse branches. Define p−(v)=−V′(r−(v))p_-(v)=-V'(r_-(v)) and p+(v)=V′(r+(v))p_+(v)=V'(r_+(v)). Both vanish at the minimum level, and

ddvp±(v)2=2V′′(r±(v)).\frac{d}{dv}p_\pm(v)^2=2V''(r_\pm(v)).

It follows that p−(v)>p+(v)p_{-}(v) > p_{+}(v) above the minimum. At points of equal density, the left and right tail probabilities are proportional to

∫v∞e−wp−(w) dwand∫v∞e−wp+(w) dw,\int_v^\infty\frac{e^{-w}}{p_{-}(w)}\,dw \quad\text{and} \quad\int_v^\infty\frac{e^{-w}}{p_{+}(w)}\,dw,

respectively. The right tail is strictly larger.

If Rk(a)≥r0R_k(a) \ge r_0, both relevant quantiles lie on the decreasing branch and (5.1) is immediate. Otherwise let y>r0y > r_0 have the same density as Rk(a)R_k(a). The right-tail comparison gives 1−Fk(y)>a1-F_k(y) > a, whence y<Rk(1−a)y < R_k(1-a). The density is decreasing on this branch, so again (5.1) follows.

Lemma 5.2 (A uniform density bound). For every k≥1k \ge1,

sup⁡r>0ρk(r)≤2π.(33)\sup_{r>0} \rho_k(r) \le\sqrt{\frac{2}{\pi}}. \tag*{(33)}

Proof. Equality holds for the maximum of ρ1\rho_1. For k≥2k \ge2, the lower Stirling bound for Γ(k/2)\Gamma(k/2) gives

ρk(k−1)<e1/2π(1−1k)(k−1)/2≤e2π<2π.\rho_k(\sqrt{k-1}) < \frac{e^{1/2}}{\sqrt{\pi}}\left(1-\frac{1}{k}\right)^{(k-1)/2} \le\sqrt{\frac{e}{2\pi}} < \sqrt{\frac{2}{\pi}}.

Here (1+1/(k−1))k−1≥2(1+1/(k-1))^{k-1} \ge2 by the binomial formula, and e<4e < 4.

Proposition 5.3 (Every active rank is bounded). Let m>1/2m > 1/2 and suppose jk(m)=I∞(m)j_k(m)=I_\infty(m). If k≥2k \ge2, then

Rk(m)2≤k+1(34)R_k(m)^2 \le k+1 \tag*{(34)}

and

k≤2518π(m−1/2)2<12(m−1/2)2.(35)k \le\frac{25}{18\pi(m-1/2)^2} < \frac{1}{2(m-1/2)^2}. \tag*{(35)}

The last, weaker bound also holds for k=1k=1.

Proof. We include the second-variation argument for the necessary radius bound; the same threshold appears in La Manna’s local minimality and instability theorem [8]. The centered kk-ball of radius R=Rk(m)R=R_k(m) is a symmetric Gaussian perimeter minimizer in Rk\mathbb{R}^k, by Proposition 4.3. For an even mean-zero normal speed ff on its boundary, Gaussian second variation gives

∫∂BR(∣∇f∣2−(1+k−1R2)f2)ϕk dHk−1≥0.\int_{\partial B_R}\left(|\nabla f|^2-\left(1+\frac{k-1}{R^2}\right)f^2\right)\phi_k\,d\mathcal{H}^{k-1}\ge0.

Take ff to be a nonzero degree-two spherical harmonic, of Laplace eigenvalue 2k/R22k/R^2. Its symmetry and zero mean make it admissible; an even constant correction of second order enforces exact volume. The displayed inequality becomes ((k+1)/R2−1)∫f2ϕk≥0((k+1)/R^2-1)\int f^2\phi_k\ge0, proving (5.3).

Put δ=m−1/2\delta=m-1/2. The median lower bound (3.5) and Lemma 5.2 give

δ=∫Rk(1/2)Rk(m)ρk(r) dr\delta=\int_{R_k(1/2)}^{R_k(m)}\rho_k(r)\,dr
≤2π(k+1−k−2/3)≤562π k−1/2.\le\sqrt{\frac{2}{\pi}}\left(\sqrt{k+1}-\sqrt{k-2/3}\right)\le\frac{5}{6}\sqrt{\frac{2}{\pi}}\,k^{-1/2}.

For the last inequality, k+1+k−2/3≥2k\sqrt{k+1}+\sqrt{k-2/3}\ge2\sqrt{k} when k≥2k\ge2. Squaring reduces this fact to

(k+1)(k−2/3)−(k−1/6)2=2k3−2536>0.(k+1)(k-2/3)-(k-1/6)^2=\frac{2k}{3}-\frac{25}{36}>0.

This proves (5.4); the comparison with 1/21/2 uses π>3\pi>3. For k=1k=1, the weaker bound is automatic since 0<δ<1/20<\delta<1/2.

Propositions 3.3 and 4.3, together with Lemma 5.1, prove (1.5), finite attainment away from 1/21/2, and the asserted half-mass value. Proposition 5.3 proves (1.6). Taking the product of an attaining ball or its complement with Rd−k\mathbb{R}^{d-k} proves (1.7). The nonattainment assertion at half mass is proved in § 8.

The critical-rank asymptotics

Lemma 6.1 (A uniform expansion near half mass). Put u=k−1/2u=k^{-1/2} and c0=π−1/2c_0=\pi^{-1/2}. There are constants k0≥1k_0\ge1, t0>0t_0>0, and C<∞C<\infty, independent of kk and tt, such that, for k≥k0k\ge k_0 and ∣t∣≤t0\lvert t\rvert\le t_0,

j˙k(1/2+t)=c0+u218π−tu3−πt2+O((u+∣t∣)4).(36)\dot{j}_k(1/2+t)=c_0+\frac{u^2}{18\sqrt{\pi}}-\frac{tu}{3}-\sqrt{\pi}t^2+O((u+\lvert t\rvert)^4). \tag*{(36)}

Here the absolute value of the error is at most C(u+∣t∣)4C(u+\lvert t\rvert)^4.

Proof. Differentiation of the quantile and density gives the exact identities

Rk′=1jk,j˙k=k−1Rk−Rk,j¨k=−1+(k−1)/Rk2jk.(37)R'_k=\frac{1}{j_k},\qquad\dot{j}_k=\frac{k-1}{R_k}-R_k,\qquad\ddot{j}_k=-\frac{1+(k-1)/R_k^2}{j_k}. \tag*{(37)}

Lemma 3.2 consequently yields

j˙k(1/2)=c0+u218π+O(u4),(38)\dot{j}_k(1/2)=c_0+\frac{u^2}{18\sqrt{\pi}}+O(u^4), \tag*{(38)}
Equation not extracted

\dot{j}_k(1/2)=-\frac{u}{3}+O(u^3),

{#eq:6.4}
j¨k(1/2)=−2π+O(u2).(39)\ddot{j}_k(1/2)=-2\sqrt{\pi}+O(u^2). \tag*{(39)}

We justify a remainder uniform in both parameters. Let

Z(t)=Rk(1/2+t)−k,J(t)=jk(1/2+t).Z(t)=R_k(1/2+t)-\sqrt{k},\qquad J(t)=j_k(1/2+t).

Then

Z′=1J,J′=A(u,Z):=−2Z+uZ2−11+uZ.(40)Z'=\frac{1}{J},\qquad J'=A(u,Z):=-2Z+u\frac{Z^2-1}{1+uZ}. \tag*{(40)}

The initial conditions are Z(0)=−u/3+O(u3)Z(0)=-u/3+O(u^3) and (6.3). On one fixed small interval ∣t∣≤t0\lvert t\rvert\le t_0, and for all sufficiently small uu, these ordinary differential equations give J≥c0/2J\ge c_0/2 and Z=O(u+∣t∣)Z=O(u+\lvert t\rvert), with uniform constants. For example, these bounds follow by continuation in a compact rectangle about (Z,J)=(0,c0)(Z,J)=(0,c_0), since the right-hand sides are uniformly smooth there. Explicitly,

AZ=−1−1−u2(1+uZ)2,AZZ=2u(1−u2)(1+uZ)3.A_Z=-1-\frac{1-u^2}{(1+uZ)^2},\qquad A_{ZZ}=\frac{2u(1-u^2)}{(1+uZ)^3}.

It follows from (6.6) that

J′′′=AZZ−AZAJ2=O(u+∣t∣).J'''=\frac{A_{ZZ}-A_ZA}{J^2}=O(u+\lvert t\rvert).

Taylor’s formula with integral remainder, combined with (6.3)–(6.5), now has error bounded by

C(u4+∣t∣u3+t2u2+∣t∣3u+t4),C(u^4+\lvert t\rvert u^3+t^2u^2+\lvert t\rvert^3u+t^4),

which proves (6.1). □\square

Proof of Theorem 1.2. Every attaining rank tends to infinity as δ↓0\delta\downarrow0. Otherwise a bounded-rank subsequence, continuity of I∞=CI_\infty=C, and (3.3) would contradict I∞(1/2)=c0I_\infty(1/2)=c_0.

Choose an integer kk nearest to (9πδ2)−1(9\pi\delta^2)^{-1}. For this test, u=k−1/2=3π δ+O(δ3)u=k^{-1/2}=3\sqrt{\pi}\,\delta+O(\delta^3). The O(δ3)O(\delta^3) error follows from the bounded integer-rounding error and the derivative of x↦x−1/2x\mapsto x^{-1/2} at x≍δ−2x\asymp\delta^{-2}. Lemma 6.1 gives

I∞(1/2+δ)≤c0−3π2δ2+O(δ4).(41)I_\infty(1/2+\delta)\le c_0-\frac{3\sqrt{\pi}}{2}\delta^2+O(\delta^4). \tag*{(41)}

For an attaining rank put u=k(δ)−1/2u = k(\delta)^{-1/2}. One first has u=O(δ)u = O(\delta), uniformly over all choices of attaining ranks. If this uniform bound failed, we could choose a sequence of attaining ranks with u/δ→∞u/\delta\to\infty. Along that sequence, u→0u \to0 and (36) gives jk(1/2+δ)=c0+u2/(18π)+o(u2)>c0j_k(1/2+\delta)=c_0+u^2/(18\sqrt{\pi})+o(u^2)>c_0, contrary to (41).

The expansion is therefore uniform with remainder O(δ4)O(\delta^4) at every attaining rank. Completing the square gives

I∞(1/2+δ)=c0−3π2δ2+(u−3πδ)218π+O(δ4).(42)I_\infty(1/2+\delta)=c_0-\frac{3\sqrt{\pi}}{2}\delta^2+\frac{(u-3\sqrt{\pi}\delta)^2}{18\sqrt{\pi}}+O(\delta^4). \tag*{(42)}

Comparison with (41) yields u=3π δ+O(δ2)u=3\sqrt{\pi}\,\delta+O(\delta^2), hence (9). The nonnegative square in (42) and (41) prove (10). Symmetry gives the expansion on the left of half mass and differentiability at that point. No differentiability of the lower envelope in a whole neighborhood of that point is used. □

Half mass is not attained in finite dimension

Proposition 7.1. If E⊂RdE\subset\mathbb{R}^d is centrally symmetric and γd(E)=1/2\gamma_d(E)=1/2, then

Pγd(E)>1π.(43)P_{\gamma_d}(E)>\frac{1}{\sqrt{\pi}}. \tag*{(43)}

Proof. Only the equality case needs to be excluded. Suppose equality holds. Then EE is a symmetric perimeter minimizer in its ambient dimension; indeed it realizes the infimum over all dimensions. A product with R\mathbb{R} retains this property, so we may assume the ambient dimension is at least two.

Choose the usual finite-perimeter representative of EE. The regularity and first-variation statements for symmetric Gaussian minimizers, in [7 Lemmas 4.1 and 4.3], give a smooth reduced boundary Σ\Sigma and a constant λ\lambda such that

H−⟨x,ν⟩=λ(44)H-\langle x,\nu\rangle=\lambda \tag*{(44)}

there, where ν\nu is the outward unit normal and H=div⁡ΣνH=\operatorname{div}_{\Sigma}\nu. We use only compactly supported variations on the regular part; the classification invoked below includes the possible singular set. Choose a regular point x≠0x\ne0 and a small regular patch UU such that UU and −U-U are disjoint. Such a point exists because the positive perimeter is carried by the reduced boundary. Choose a nonzero, nonnegative, even smooth function ff supported on U∪(−U)U\cup(-U). Since ν(−x)=−ν(x)\nu(-x)=-\nu(x), the normal vector field fνf\nu is odd. Extend it to a compactly supported smooth ambient field and antisymmetrize the extension to obtain X(−x)=−X(x)X(-x)=-X(x) with X⋅ν=fX\cdot\nu=f on Σ\Sigma. The flow of XX commutes with x↦−xx\mapsto-x and thus produces centrally symmetric sets EsE_s for both signs of small ss. Writing a(s)=γd(Es)a(s)=\gamma_d(E_s) and p(s)=Pγd(Es)p(s)=P_{\gamma_d}(E_s), we have

a′(0)=∫Σfφd dHd−1>0,p′(0)=λa′(0).a'(0)=\int_{\Sigma}f\varphi_d\,d\mathcal{H}^{d-1}>0,\qquad p'(0)=\lambda a'(0).

By the already proved profile inequality, p(s)≥I∞(a(s))p(s)\ge I_\infty(a(s)), with equality at s=0s=0. Theorem 1.2 shows that I∞I_\infty is differentiable at 1/21/2 with derivative zero. Consequently s↦p(s)−I∞(a(s))s\mapsto p(s)-I_\infty(a(s)) is differentiable at zero and has a local minimum there, so p′(0)=0p'(0)=0. The first-variation identity and a′(0)>0a'(0)>0 imply λ=0\lambda=0.

We now apply the existing self-shrinker classification [7 Theorem 6.1]. Its hypotheses are precisely minimization in the centrally symmetric class at a fixed Gaussian volume and H=⟨x,ν⟩H=\langle x,\nu\rangle on the reduced boundary; both have just been verified. Compactness and convexity are not extra assumptions in that theorem. It says that the boundary is a hyperplane through zero or a self-shrinking round cylinder. In the latter case its active dimension k≥2k\ge2 has radius k−1\sqrt{k-1}. The indicator of a finite-perimeter set is constant almost everywhere on each connected component where its distributional derivative vanishes. The two components of the complement of this cylinder therefore force EE to be its interior or exterior, up to null sets; the two constant choices would have mass zero or one. But (23) gives Rk(1/2)2>k−2/3>k−1R_k(1/2)^2 > k - 2/3 > k - 1, so neither side has mass 1/21/2. A hyperplane through zero exchanges its two sides under x↦−xx \mapsto-x and cannot bound a nontrivial centrally symmetric set. Both possibilities contradict the assumed mass. This proves (43) and completes Theorem 1.1. □

The sharp even Gaussian Cheeger inequality

Proof of Corollary 1.3. The concave symmetric profile C=I∞C = I_\infty, extended by zero at the endpoints, satisfies C(1/2)=1/πC(1/2) = 1/\sqrt{\pi}. Hence

C(a)≥2πmin⁡{a,1−a}.C(a) \ge\frac{2}{\sqrt{\pi}}\min\{a,1-a\}.

Together with Proposition 4.3, this proves (11).

Let cc be a median of an even Lipschitz function ff. The sets {f>c+t}\{f > c + t\} and {f<c−t}\{f < c - t\} are centrally symmetric and have mass at most 1/21/2 for t>0t > 0. Layer cake and (11), followed by Gaussian coarea, give

∫∣f−c∣ dγd=∫0∞[γd(f>c+t)+γd(f<c−t)] dt≤π2∫0∞[Pγd({f>c+t})+Pγd({f<c−t})] dt=π2∫∣∇f∣ dγd.\begin{aligned} \int|f-c|\,d\gamma_d &= \int_0^\infty\left[\gamma_d(f>c+t)+\gamma_d(f<c-t)\right]\,dt \\ &\le\frac{\sqrt{\pi}}{2}\int_0^\infty\left[P_{\gamma_d}(\{f>c+t\})+P_{\gamma_d}(\{f<c-t\})\right]\,dt \\ &= \frac{\sqrt{\pi}}{2}\int|\nabla f|\,d\gamma_d. \end{aligned}

A median minimizes the integral over constants, proving (12). Finally, centered half-mass balls have perimeter tending to 1/π1/\sqrt{\pi} by Lemma 3.2. Their indicators, approximated by even radial Lipschitz functions, prove sharpness in both formulations. More explicitly, for a fixed half-mass ball BB, radial transition layers give functions 0≤fε≤10 \le f_\varepsilon\le1 with

fε⟶1B in L1(γk),∫∣∇fε∣ dγk⟶Pγk(B).f_\varepsilon\longrightarrow\mathbf{1}_B \text{ in } L^1(\gamma_k), \qquad\int|\nabla f_\varepsilon|\,d\gamma_k \longrightarrow P_{\gamma_k}(B).

The map f↦inf⁡c∥f−c∥L1(γk)f \mapsto\inf_c \|f-c\|_{L^1(\gamma_k)} is 1-Lipschitz in L1L^1, and its value at 1B\mathbf{1}_B is 1/21/2. First letting ε↓0\varepsilon\downarrow0 and then k→∞k \to\infty shows that the functional constant cannot be decreased. □

Remark 8.1 (Strictness and the centering constant). The set inequality is strict in each finite dimension whenever the mass is nontrivial. At half mass this is Proposition 7.1. If 0<a<1/20 < a < 1/2, choose an attaining rank kk for C(a)=jk(1−a)C(a) = j_k(1-a). The concavity of q↦jk(1−q)q \mapsto j_k(1-q), its zero value at q=0q=0, and Lemma 3.2 give

C(a)≥2a jk(1/2)>2aπ.C(a) \ge2a\,j_k(1/2) > \frac{2a}{\sqrt{\pi}}.

Complementation handles a>1/2a > 1/2. The functional statement is centered at a median, equivalently at the best constant in L1L^1; no replacement of this constant by the mean is asserted.

Declaration of generative AI use

OpenAI’s ChatGPT and Codex were used in developing and examining mathematical arguments, checking calculations, locating literature, and drafting and revising the exposition and \LaTeX{} source. AI-generated output is not treated as an independent source of mathematical verification. Responsibility for the final content, including all proofs and references, rests with the author.

References

References

  1. [1]M. Barchiesi, F. Cagnetti, and V. Julin, Double-arc rearrangement and the symmetric isoperimetric problem in Gauss space, preprint (2026), arXiv:2610.15133v1. https://arxiv.org/abs/2610.15133v1.
  2. [2]M. Barchiesi and V. Julin, Symmetry of minimizers of a Gaussian isoperimetric problem, Probab. Theory Related Fields 177 (2020), 217–256. doi:10.1007/s00440-019-00947-9.DOI
  3. [3]F. Barthe, An isoperimetric result for the Gaussian measure and unconditional sets, Bull. London Math. Soc. 33 (2001), no. 4, 408–416. doi:10.1017/S0024609301008411.DOI
  4. [4]C. Berg and H. L. Pedersen, The Chen–Rubin conjecture in a continuous setting, Methods Appl. Anal. 13 (2006), no. 1, 63–88. doi:10.4310/MAA.2006.v13.n1.a4.DOI
  5. [5]J. Chen and H. Rubin, Bounds for the difference between median and mean of gamma and Poisson distributions, Statist. Probab. Lett. 4 (1986), no. 6, 281–283. doi:10.1016/0167-7152(86)90044-1.DOI
  6. [6]S. Heilman, Symmetric convex sets with minimal Gaussian surface area, Amer. J. Math. 143 (2021), no. 1, 53–94. doi:10.1353/ajm.2021.0000.DOI
  7. [7]S. Heilman, Convex cylinders and the symmetric Gaussian isoperimetric problem, preprint (2022), arXiv:2204.12003. https://arxiv.org/abs/2204.12003.
  8. [8]D. A. La Manna, Local minimality of the ball for the Gaussian perimeter, Adv. Calc. Var. 12 (2019), no. 2, 193–210. doi:10.1515/acv-2017-0007.
  9. [9]F. Maggi, Sets of Finite Perimeter and Geometric Variational Problems: An Introduction to Geometric Measure Theory, Cambridge Studies in Advanced Mathematics, vol. 135, Cambridge University Press, 2012.DOI
  10. [10]E. Milman, The symmetric Gaussian isoperimetric problem in the plane, preprint (2026), arXiv:2611.12456v1. https://arxiv.org/abs/2611.12456v1.
  11. [11]C. Viana, Isoperimetry and volume preserving stability in real projective spaces, J. Differential Geom. 125 (2023), no. 1, 187–205. doi:10.4310/jdg/1695236595.DOI

Paper details

vlnc:W_quhnj63k
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

OpenAI ChatGPT GPT 6 Astra

View statement
Looking for related papers…
Contents