betaFeedback

WorksMathematics

Large deviations and the asymptotic geometry of Lorentz balls

Abstract

We study the uniform distribution on finite-dimensional Lorentz balls Bq,pnB_{q,p}^n in the range 1<p<q<∞1<p<q<\infty. We prove a Sanov-type large deviation principle for the empirical measures of the absolute coordinates of a random vector uniformly distributed on n1/qBq,pnn^{1/q}B_{q,p}^n. The rate function is given by an entropy deficit under a weighted quantile constraint. We show that the entropy maximizer is unique and characterize its distribution function as the unique increasing solution of a nonlinear boundary-value problem. The limiting distribution has compact support, for which we obtain an explicit bound. These results establish the conjectured empirical limit of Kabluchko, Prochno and Sonnleitner [Journal of Functional Analysis, 288(1):110682, 2025] in the range 1<p<q<∞1<p<q<\infty. As consequences, we determine the exponential volume asymptotics of Lorentz balls and obtain asymptotic independence of any fixed number of coordinates. We also prove a law of large numbers for Lorentz norms with finite parameters and use it, together with the volume asymptotics, to obtain a Schechtman–Schmuckenschläger type theorem for intersections of volume-normalized Lorentz balls. We further derive a variational formula for the logarithmic intersection asymptotics. The proof combines an exponential change of measure with quantile interpolation and a variational argument that does not assume regularity of the maximizing quantile in advance.

Introduction and main results

The probabilistic study of high-dimensional convex bodies is a central part of asymptotic geometric analysis. Endowing a convex body with normalized Lebesgue measure connects its geometric properties to the distribution of a random vector and makes probabilistic limit theorems available for their study. The unit balls of the classical sequence spaces ℓpn\ell_p^n provide a particularly instructive setting. A fundamental tool is the probabilistic representation obtained by Schechtman and Zinn [21] and independently by Rachev and Rüschendorf [19]: a uniformly distributed point can be represented using independent pp-generalized Gaussian random variables, normalization by their ℓpn\ell_p^n-norm, and an independent radial factor. This representation underlies the analysis of coordinate distributions, norms, and asymptotic intersection volumes for these balls.

An important impetus for the systematic study of large deviations in this setting came from Gantert, Kim and Ramanan [8], who established annealed and quenched large deviation principles for one-dimensional projections of uniformly distributed points in ℓpn\ell_p^n-balls. The dependence of the speeds and rate functions on pp reveals geometric information beyond the Gaussian fluctuation regime. Alonso-Gutiérrez, Prochno and Thäle [1] subsequently studied the Euclidean norms of projections onto random subspaces whose dimensions may grow with the ambient dimension. Kabluchko, Prochno and Thäle [13] obtained central limit theorems and large deviation principles for ℓq\ell_q-norms of random points in ℓpn\ell_p^n-balls, and later complemented these results by moderate deviation principles [16], describing the intermediate scale between central limit and large deviation behaviour.

A closely related development concerns the empirical distribution of coordinates. Kim and Ramanan [17] proved Sanov-type large deviation principles for empirical measures under the cone and surface probability measures on high-dimensional ℓpn\ell_p^n-spheres and used them to derive conditional limit theorems. In the matrix setting, Kabluchko, Prochno and Thäle [15] established corresponding large deviation principles for empirical eigenvalue measures of self-adjoint matrix balls and singular-value measures of Schatten balls. Such results identify limiting empirical distributions through variational problems for the rate function. For an account of these developments and their connections to geometric functional analysis, we refer to the survey [18].

The extension of this programme to Orlicz balls brought maximum entropy methods and Gibbs measures to the fore. Kabluchko and Prochno [11] related the uniform distribution on an Orlicz ball to Gibbs distributions with potential given by the defining Orlicz function. This provided an alternative to the explicit probabilistic representation for ℓpn\ell_p^n-balls and led to asymptotic volume formulas (see also [2]) and results on intersection volumes. Building on this approach, Frühwirth and Prochno [6] established Sanov-type large deviation principles for the empirical measures of the coordinates, together with conditional limit theorems. These results extend the connection between empirical distributions, entropy maximization, and asymptotic geometry beyond the classical power constraints.

Lorentz balls form another natural extension of the classical ℓpn\ell_p^n-balls. For p<qp<q, their defining constraint depends on the order of the absolute coordinates and therefore cannot be written as a sum of one-coordinate potentials. To make this precise, let x1∗≥⋯≥xn∗x_1^* \ge\cdots\ge x_n^* denote the decreasing rearrangement of (∣xi∣)i=1n(|x_i|)_{i=1}^n. For 1≤p≤q<∞1 \le p \le q < \infty, the finite-dimensional Lorentz space ℓq,pn\ell_{q,p}^n is Rn\mathbb{R}^n equipped with the norm

∥x∥q,p:=(∑i=1n(in)p/q−1(xi∗)p)1/p.\|x\|_{q,p}:=\left(\sum_{i=1}^n \left(\frac{i}{n}\right)^{p/q-1}(x_i^*)^p\right)^{1/p}.

and we denote its unit ball by

Bq,pn:={x∈Rn:∥x∥q,p≤1}.B_{q,p}^n:=\{x\in\mathbb{R}^n:\|x\|_{q,p}\le1\}.

When p=qp=q, this is the usual ℓpn\ell_p^n-ball. The dependence on the decreasing rearrangement is the main additional obstacle when p<qp<q.

Kabluchko, Prochno and Sonnleitner [12] developed a probabilistic approach to Lorentz balls in the case p=1p=1. They obtained an explicit probabilistic representation of the uniform distribution, identified the limiting empirical law, and derived coordinate and volumetric limit theorems, including a Schechtman–Schmuckenschläger type result for intersections with classical ℓrn\ell_r^n-balls. For general 1≤p≤q<∞1\le p\le q<\infty, they formulated a conjecture describing the limiting distribution through a nonlinear differential equation with boundary conditions. Their heuristic derivation uses a weighted quantile constraint and a maximum entropy principle.

In this paper, we establish this conjectured description in the range 1<p<q<∞1<p<q<\infty. We first prove a large deviation principle for the empirical measures on the positive orthant. We then show that the associated entropy maximization problem has a unique solution and derive the differential equation together with its boundary conditions. A converse argument proves uniqueness of the admissible solution of the boundary-value problem. In particular, the differential equation characterizes the limiting law. The same analysis determines the exponential volume asymptotics and yields a Poincaré–Maxwell–Borel principle for Lorentz balls. We also prove a law of large numbers for Lorentz norms with parameters 1≤s≤r<∞1\le s\le r<\infty. Together with the volume asymptotics, this yields a Schechtman–Schmuckenschläger type theorem for intersections of two volume-normalized Lorentz balls. We further determine the exponential decay rate of the intersection volumes below the threshold.

Main results

Throughout the paper, unless stated otherwise, we fix 1<p<q<∞1<p<q<\infty and use the normalization

Dn:=n1/qBq,pn={x∈Rn:1n∑i=1n(in)p/q−1(xi∗)p≤1}.(1)D_n:=n^{1/q}B_{q,p}^n=\left\{x\in\mathbb{R}^n:\frac{1}{n}\sum_{i=1}^n\left(\frac{i}{n}\right)^{p/q-1}(x_i^*)^p\le1\right\}. \tag*{(1)}

Let Dn+:=Dn∩[0,∞)nD_n^+:=D_n\cap[0,\infty)^n, let U(n)U^{(n)} be uniformly distributed on Dn+D_n^+, and put

Ln:=1n∑i=1nδUi(n).L_n:=\frac{1}{n}\sum_{i=1}^n\delta_{U_i^{(n)}}.

We equip P([0,∞))\mathcal{P}([0,\infty)) with the topology of weak convergence. For μ∈P([0,∞))\mu\in\mathcal{P}([0,\infty)), its increasing quantile function is

Qμ(t):=inf⁡{x≥0:μ([0,x])≥t},0<t<1.Q_\mu(t):=\inf\{x\ge0:\mu([0,x])\ge t\},\qquad0<t<1.

For t∈(0,1)t\in(0,1), define

w(t):=(1−t)p/q−1,E(μ):=∫01w(t)Qμ(t)p dt,K:={μ∈P([0,∞)):E(μ)≤1}.(2)w(t):=(1-t)^{p/q-1},\qquad\mathcal{E}(\mu):=\int_0^1w(t)Q_\mu(t)^p\,dt,\qquad K:=\{\mu\in\mathcal{P}([0,\infty)):\mathcal{E}(\mu)\le1\}. \tag*{(2)}

Every measure in KK has ppth moment at most one. We write Leb⁡\operatorname{Leb} for Lebesgue measure on [0,∞)[0,\infty) and Leb⁡A\operatorname{Leb}_A for its restriction to a Borel set A⊆[0,∞)A\subseteq[0,\infty). For a measure μ\mu and a measurable map TT defined on its underlying space, we write T#μT_{\#}\mu for the pushforward of μ\mu under TT, defined by (T#μ)(A):=μ(T−1(A))(T_{\#}\mu)(A):=\mu(T^{-1}(A)).

for every measurable set AA in the target space. For a probability measure μ\mu with finite first moment, we write

h(μ):={−∫0∞f(x)log⁡f(x) dx:μ=f dx,−∞:μ≪̸Leb⁡,h(\mu):= \begin{cases} -\int_{0}^{\infty} f(x)\log f(x)\,\mathrm{d}x & : \mu=f\,\mathrm{d}x,\\ -\infty& : \mu\not\ll\operatorname{Leb}, \end{cases}

and put 0log⁡0:=00\log0:=0. This entropy is well defined in [−∞,∞)[-\infty,\infty). Let us define

h∗:=sup⁡μ∈Kh(μ).(3)h^*:=\sup_{\mu\in K}h(\mu). \tag*{(3)}

Theorem A (Large deviations and volume asymptotics). Let 1<p<q<∞1<p<q<\infty. Then h∗∈Rh^*\in\mathbb{R}, and the empirical measures (Ln)n∈N(L_n)_{n\in\mathbb{N}} satisfy a large deviation principle on P([0,∞))\mathcal{P}([0,\infty)) with speed nn and good rate function

I(μ):={h∗−h(μ):E(μ)≤1,+∞:E(μ)>1.(4)I(\mu):= \begin{cases} h^*-h(\mu) & : \mathcal{E}(\mu)\le1,\\ +\infty& : \mathcal{E}(\mu)>1. \end{cases} \tag*{(4)}

Moreover,

lim⁡n→∞n1/qvol⁡n(Bq,pn)1/n=2eh∗.(5)\lim_{n\to\infty} n^{1/q}\operatorname{vol}_{n}(B_{q,p}^{n})^{1/n}=2e^{h^*}. \tag*{(5)}

Thus the exponential volume asymptotics are determined by a one-dimensional variational problem. Our second theorem identifies its maximizer and the corresponding empirical limit.

Theorem B (The limiting distribution). Let 1<p<q<∞1<p<q<\infty.

(i) There is a unique measure μ∗∈K\mu_*\in K with h(μ∗)=h∗h(\mu_*)=h^*, and E(μ∗)=1\mathcal{E}(\mu_*)=1. Its quantile Q=Qμ∗Q=Q_{\mu_*} extends to Q∈C1([0,1))Q\in C^1([0,1)), with Q(0)=0Q(0)=0, and satisfies

1Q′(s)=∫s1Q(t)p−1(1−t)p/q−1 dt,0≤s<1.(6)\frac{1}{Q'(s)}=\int_s^1 Q(t)^{p-1}(1-t)^{p/q-1}\,\mathrm{d}t,\qquad0\le s<1. \tag*{(6)}

(ii) The support of μ∗\mu_* is [0,rp,q][0,r_{p,q}], where

0<rp,q≤(p2q−p)1/p.(7)0<r_{p,q}\le\left(\frac{p^2}{q-p}\right)^{1/p}. \tag*{(7)}

Its distribution function GG belongs to C1([0,rp,q])∩C2((0,rp,q))C^1([0,r_{p,q}])\cap C^2((0,r_{p,q})) and is the unique increasing solution, with positive derivative in the interior, of

G′′(x)=−xp−1(1−G(x))p/q−1G′(x),0<x<r,G''(x)=-x^{p-1}(1-G(x))^{p/q-1}G'(x),\qquad0<x<r,
G(0)=0,lim⁡x↑rG(x)=1,lim⁡x↑rG′(x)=0.(8)G(0)=0,\qquad\lim_{x\uparrow r}G(x)=1,\qquad\lim_{x\uparrow r}G'(x)=0. \tag*{(8)}

Here uniqueness is among pairs (r,G)(r,G) with r∈(0,∞]r\in(0,\infty], G∈C([0,r))∩C2((0,r))G\in C([0,r))\cap C^2((0,r)) and 0<G<10<G<1, G′>0G'>0 on (0,r)(0,r); the solution necessarily has r=rp,q<∞r=r_{p,q}<\infty. Derivatives at finite endpoints are understood one-sidedly.

(iii) The empirical measures LnL_n converge to μ∗\mu_* in probability with respect to the topology of weak convergence. If X(n)X^{(n)} is uniformly distributed on the full scaled ball DnD_n, then

1n∑i=1nδXi(n)→Pνq,p,νq,p:=12μ∗+12(−id⁡)#μ∗.(9)\frac{1}{n}\sum_{i=1}^{n}\delta_{X_i^{(n)}}\xrightarrow{\mathbb{P}}\nu_{q,p},\qquad \nu_{q,p}:=\frac{1}{2}\mu_*+\frac{1}{2}(-\operatorname{id})_\#\mu_*. \tag*{(9)}

The measure νq,p\nu_{q,p} has the continuous density

fq,p(x)={12G′(∣x∣),∣x∣≤rp,q,0,∣x∣>rp,q.(10)f_{q,p}(x)= \begin{cases} \frac{1}{2}G'(|x|), & |x|\le r_{p,q},\\ 0, & |x|>r_{p,q}. \end{cases} \tag*{(10)}
Numerical limiting densities $f_{q,p}$ of the coordinate distributions on high-dimensional Lorentz balls

Figure 1. Numerical limiting densities fq,pf_{q,p} of the coordinate distributions on high-dimensional Lorentz balls. The left panel fixes p=2p=2 and varies qq; the right panel fixes q=6q=6 and varies pp. Each density is symmetric, has total mass one, and vanishes outside a compact interval.

Theorem B proves the empirical-limit conjecture in [12] for 1<p<q<∞1<p<q<\infty. The cases p=1p=1 and p=qp=q are discussed in that paper and correspond, respectively, to its explicit Lorentz limit and to the classical pp-Gaussian law. The compact support in (7) is a feature of the strict inequality p<qp<q. Our assertion of uniqueness in (8) includes the terminal derivative condition; the differential equation with G(0)=0G(0)=0 alone does not determine a unique solution.

The permutation symmetry of DnD_n turns convergence of the empirical measures into convergence of any fixed number of coordinates.

Corollary 1.1 (A Poincaré–Maxwell–Borel principle). Let k∈Nk\in\mathbb{N} be fixed and let X(n)X^{(n)} be uniformly distributed on DnD_n. Then

Law⁡(X1(n),…,Xk(n))⇒νq,p⊗kas n→∞.\operatorname{Law}(X_1^{(n)},\ldots,X_k^{(n)}) \Rightarrow\nu_{q,p}^{\otimes k}\qquad\text{as }n\to\infty.

We also obtain an intersection-volume result in the sense of Schechtman and Schmuckenschläger [20]. It complements [7, 9, 10, 11, 14], extends the intersection result of [12] to p>1p>1 and allows a second Lorentz ball in place of a classical ℓrn\ell_r^n-ball. For 1≤b≤a<∞1\le b\le a<\infty, write

B^a,bn:=vol⁡n(Ba,bn)−1/nBa,bn,ca,b:=lim⁡n→∞n1/avol⁡n(Ba,bn)1/n.(11)\widehat{B}_{a,b}^n:=\operatorname{vol}_n(B_{a,b}^n)^{-1/n}B_{a,b}^n,\qquad c_{a,b}:=\lim_{n\to\infty}n^{1/a}\operatorname{vol}_n(B_{a,b}^n)^{1/n}. \tag*{(11)}

If 1<b<a1<b<a, let ha,b∗h_{a,b}^* denote the entropy maximum for the pair (a,b)(a,b), so that hq,p∗=h∗h_{q,p}^*=h^*. Theorem A, the volume asymptotics in [12], and the classical ℓan\ell_a^n-ball volume formula give, respectively,

ca,b={2eha,b∗,1<b<a,2e1/a/a,b=1<a,2Γ(1+1/a)(ae)1/a,b=a.(12)c_{a,b}= \begin{cases} 2e^{h_{a,b}^*}, & 1<b<a,\\ 2e^{1/a}/a, & b=1<a,\\ 2\Gamma(1+1/a)(ae)^{1/a}, & b=a. \end{cases} \tag*{(12)}

Corollary 1.2 (Intersection volumes). Let 1<p<q<∞1<p<q<\infty and 1≤s≤r<∞1\le s\le r<\infty, and let Q=Qμ∗Q=Q_{\mu_*} be the quantile from Theorem B. Define

mq,p;r,s:=(∫01(1−u)s/r−1Q(u)s du)1/s,(13)m_{q,p;r,s}:=\left(\int_0^1(1-u)^{s/r-1}Q(u)^s\,\mathrm{d}u\right)^{1/s}, \tag*{(13)}
τq,p;r,s:=cr,scq,pmq,p;r,s.(14)\tau_{q,p;r,s}:=\frac{c_{r,s}}{c_{q,p}}m_{q,p;r,s}. \tag*{(14)}

Then, for every t>0t>0 with t≠τq,p;r,st\ne\tau_{q,p;r,s},

lim⁡n→∞vol⁡n(B^q,pn∩tB^r,sn)={0t<τq,p;r,s,1t>τq,p;r,s.(15)\lim_{n\to\infty}\operatorname{vol}_n\left(\widehat{B}_{q,p}^n\cap t\widehat{B}_{r,s}^n\right)= \begin{cases} 0 & t<\tau_{q,p;r,s},\\ 1 & t>\tau_{q,p;r,s}. \end{cases} \tag*{(15)}

For every fixed t∈(0,τq,p;r,s)t\in(0,\tau_{q,p;r,s}), there exists J(t)∈(0,∞)J(t)\in(0,\infty) such that

vol⁡n(B^q,pn∩tB^r,sn)=exp⁡{−nJ(t)+o(n)}as n→∞.\operatorname{vol}_n\left(\widehat{B}_{q,p}^n\cap t\widehat{B}_{r,s}^n\right)=\exp\{-nJ(t)+o(n)\}\qquad\text{as }n\to\infty.

Moreover, τq,p;r,s≥1\tau_{q,p;r,s}\ge1. If 1<s<r<∞1<s<r<\infty, then τq,p;r,s=1\tau_{q,p;r,s}=1 if and only if (r,s)=(q,p)(r,s)=(q,p). In particular, the intersection volume at t=1t=1 tends to zero for two distinct parameter pairs in the strict parameter range.

For 1<s<r1<s<r, the threshold can equivalently be written as τq,p;r,s=ehr,s∗−hq,p∗mq,p;r,s\tau_{q,p;r,s}=e^{h^*_{r,s}-h^*_{q,p}}m_{q,p;r,s}. For a classical ℓrn\ell_r^n-ball, corresponding to s=rs=r, it is

τq,p;r,r=Γ(1+1/r)(re)1/re−h∗(∫0∞xr μ∗(dx))1/r.(16)\tau_{q,p;r,r}=\Gamma(1+1/r)(re)^{1/r}e^{-h^*}\left(\int_0^\infty x^r\,\mu_*(\mathrm{d}x)\right)^{1/r}. \tag*{(16)}

The variational characterization of J(t)J(t) is given in (55), and the exponential asymptotics are proved in Proposition 6.4.

Outline of the proofs

The first step is to express the Lorentz constraint as an integral against the empirical quantile. The resulting weight depends on nn, so the admissible sets vary with the dimension. We control these sets inside a common weakly compact moment ball and identify all their possible weak limits. An exponential change of measure then relates restricted volumes to the empirical measures of independent exponential random variables. Sanov’s theorem gives the upper bound after truncation of the first-moment tilt. For the lower bound, we use boundedly supported comparison measures and approximate arbitrary finite-entropy admissible measures by truncation and dilation.

The variational analysis is carried out in quantile coordinates. Concavity of the entropy along quantile interpolation and strict convexity of the energy give uniqueness. Before taking variations, we remove any initial offset and singular part of the maximizing quantile. Multiplicative perturbations of its derivative then yield (6) with an explicit integrable majorant. The integral equation supplies the required regularity, the boundary conditions, and the compact-support estimate. Finally, an entropy comparison shows that every admissible solution of the boundary-value problem is the same maximizer.

For the intersection results, convexity yields uniform exponential tails for the one-coordinate marginals. This strengthens empirical convergence to Wasserstein convergence and gives a law of large numbers for every Lorentz norm with finite parameters. Combining this with the volume asymptotics yields the threshold. A second quantile constraint in the restricted-volume argument gives the logarithmic intersection asymptotics.

Section 2 contains the quantile representation and the compactness and entropy estimates. Section 3 proves Theorem A. Section 4 establishes uniqueness of the entropy maximizer and the empirical and coordinate limits. Section 5 completes the characterization of the limiting distribution. Section 6 proves the intersection-volume results.

Quantiles, compactness and entropy

We write μn⇒μ\mu_n\Rightarrow\mu for weak convergence of probability measures. The space P([0,∞))\mathcal{P}([0,\infty)) is Polish for this topology; throughout, dd denotes any compatible metric. We use the facts that weak convergence implies Qμn(t)→Qμ(t)Q_{\mu_n}(t) \to Q_\mu(t) at every continuity point of QμQ_\mu, hence for almost every t∈(0,1)t \in(0,1), and that

(Qμ)#Leb⁡(0,1)=μ(17)(Q_\mu)_\#\operatorname{Leb}_{(0,1)} = \mu \tag*{(17)}

In particular, integration against μ\mu can be written as integration of a function of QμQ_\mu on (0,1)(0,1). We refer to [3 Appendix D] for the weak-topology facts used below.

The discrete constraint

For n∈Nn \in\mathbb{N}, we define

wn(t)=(1−j−1n)p/q−1ifj−1n<t≤jn,En(μ)=∫01wn(t)Qμ(t)p dt.(18)w_n(t) = \left(1-\frac{j-1}{n}\right)^{p/q-1}\quad\text{if}\quad\frac{j-1}{n}<t\leq\frac{j}{n},\qquad\mathcal{E}_n(\mu)=\int_0^1w_n(t)Q_\mu(t)^p\,\mathrm{d}t. \tag*{(18)}

We recall the finite-dimensional quantile identity from [12] and include its short proof to fix the mesh convention.

Lemma 2.1. For x∈[0,∞)nx \in[0,\infty)^n, let Lx:=n−1∑i=1nδxiL_x := n^{-1}\sum_{i=1}^n\delta_{x_i}. Then

En(Lx)=1n∑i=1n(in)p/q−1(xi∗)phencex∈Dn+⟺En(Lx)≤1.(19)\mathcal{E}_n(L_x)=\frac{1}{n}\sum_{i=1}^n\left(\frac{i}{n}\right)^{p/q-1}(x_i^*)^p\qquad\text{hence}\qquad x\in D_n^+\Longleftrightarrow\mathcal{E}_n(L_x)\leq1. \tag*{(19)}

Moreover, for any 0<t<10<t<1,

1≤wn(t)≤w(t),wn(t)→w(t),∫01w(t) dt=qp.(20)1\leq w_n(t)\leq w(t),\qquad w_n(t)\to w(t),\qquad\int_0^1w(t)\,\mathrm{d}t=\frac{q}{p}. \tag*{(20)}

Proof. On ((j−1)/n,j/n]((j-1)/n,j/n], the increasing quantile of LxL_x equals xn−j+1∗x_{n-j+1}^*. This remains true when coordinates coincide. Integrating over these intervals and replacing n−j+1n-j+1 by ii gives (2.3). Since p/q−1∈(−1,0)p/q-1\in(-1,0), ww is increasing and integrable, and wnw_n takes its value at the left endpoint of each interval. This proves (2.4).

A common compact set

Let

Cp:={μ∈P([0,∞)):∫0∞xp μ(dx)≤1},m1(μ)=∫0∞x μ(dx).C_p:=\left\{\mu\in\mathcal{P}([0,\infty)):\int_0^\infty x^p\,\mu(\mathrm{d}x)\leq1\right\},\qquad m_1(\mu)=\int_0^\infty x\,\mu(\mathrm{d}x).

The inequality wn,w≥1w_n,w\geq1 implies that all the sets {En≤1}\{\mathcal{E}_n\leq1\} and KK are contained in CpC_p.

Lemma 2.2. The set CpC_p is weakly compact and m1m_1 is continuous on CpC_p. The functionals En\mathcal{E}_n and E\mathcal{E} are weakly lower semicontinuous. If μn⇒μ\mu_n\Rightarrow\mu, then

E(μ)≤lim inf⁡n→∞En(μn).(21)\mathcal{E}(\mu)\leq\liminf_{n\to\infty}\mathcal{E}_n(\mu_n). \tag*{(21)}

Consequently, KK is compact and, for every δ>0\delta>0,

{En≤1}⊆Cp∩Kδfor all sufficiently large n,Kδ={μ:d(μ,K)≤δ}.(22)\{\mathcal{E}_n\leq1\}\subseteq C_p\cap K_\delta\quad\text{for all sufficiently large }n,\qquad K_\delta=\{\mu:d(\mu,K)\leq\delta\}. \tag*{(22)}

Proof. The ppth-moment functional is lower semicontinuous, as it is the supremum of the continuous functionals μ↦∫(xp∧R) dμ\mu\mapsto\int(x^p\wedge R)\,\mathrm{d}\mu. Thus CpC_p is closed. The bound μ((R,∞))≤R−p\mu((R,\infty))\leq R^{-p} holds uniformly on CpC_p, so Prohorov’s theorem gives compactness. Furthermore,

0≤m1(μ)−m1,R(μ)≤∫(R,∞)x μ(dx)≤R1−p,wherem1,R(μ)=∫(x∧R) dμ.(23)0\leq m_1(\mu)-m_{1,R}(\mu)\leq\int_{(R,\infty)}x\,\mu(\mathrm{d}x)\leq R^{1-p},\qquad\text{where}\qquad m_{1,R}(\mu)=\int(x\wedge R)\,\mathrm{d}\mu. \tag*{(23)}

Hence the bounded continuous functionals m1,Rm_{1,R} converge uniformly to m1m_1 on CpC_p. If μn⇒μ\mu_n \Rightarrow\mu, quantile convergence and (20) allow us to apply Fatou’s lemma to wnQμnpw_n Q_{\mu_n}^p, giving (21). The same argument with a fixed weight ww or wnw_n proves lower semicontinuity. In particular, all these functionals are Borel measurable and KK is a closed subset of CpC_p and hence compact.

If (22) were false, there would be nk→∞n_k \to\infty and μk∈Cp\mu_k \in C_p with Enk(μk)≤1E_{n_k}(\mu_k) \le1 and d(μk,K)>δd(\mu_k,K) > \delta. A subsequence converges to some μ∈Cp\mu\in C_p. The proof of (21), applied along this subsequence, gives μ∈K\mu\in K. Hence d(μk,K)≤d(μk,μ)→0d(\mu_k,K) \le d(\mu_k,\mu) \to0, contradicting d(μk,K)>δd(\mu_k,K) > \delta.

Entropy

Let ρ(dx)=e−x1[0,∞)(x)dx\rho(\mathrm{d}x)=e^{-x}\mathbf{1}_{[0,\infty)}(x)\mathrm{d}x be the Exp⁡(1)\operatorname{Exp}(1) distribution. We use the relative entropy

H(μ∣ρ)={∫log⁡(dμdρ)dμ:μ≪ρ,+∞:otherwise.H(\mu\mid\rho)= \begin{cases} \int\log\left(\frac{\mathrm{d}\mu}{\mathrm{d}\rho}\right)\mathrm{d}\mu& : \mu\ll\rho,\\ +\infty& : \text{otherwise}. \end{cases}

It is nonnegative and weakly lower semicontinuous; see [3 Section 6.2].

For every probability measure μ\mu with finite first moment, the standard entropy identity gives

h(μ)=m1(μ)−H(μ∣ρ),(24)h(\mu)=m_1(\mu)-H(\mu\mid\rho), \tag*{(24)}

with both sides understood in the extended sense. Since m1m_1 is continuous on CpC_p by Lemma 2.2, the weak lower semicontinuity of relative entropy implies that hh is upper semicontinuous on CpC_p. Moreover, h(μ)≤m1(μ)≤1h(\mu)\le m_1(\mu)\le1 for μ∈Cp\mu\in C_p. On the other hand, the uniform distribution on [0,b][0,b], with 0<b<(p/q)1/p0<b<(p/q)^{1/p}, belongs to KK and has entropy log⁡b\log b. Consequently, h∗∈Rh^*\in\mathbb{R}.

Restricted volumes and large deviations

For a Borel set A⊆P([0,∞))A\subseteq\mathcal{P}([0,\infty)), define

Vn(A):=vol⁡n{x∈Dn+:Lx∈A}.V_n(A):=\operatorname{vol}_n\{x\in D_n^+:L_x\in A\}.

This is well defined because x↦Lxx\mapsto L_x is weakly continuous. Indeed, ∫f dLx=n−1∑if(xi)\int f\,\mathrm{d}L_x=n^{-1}\sum_i f(x_i) is continuous for every f∈Cb([0,∞))f\in C_b([0,\infty)).

Proposition 3.1 (Restricted-volume bounds). For every closed F⊆P([0,∞))F\subseteq\mathcal{P}([0,\infty)) and every open O⊆P([0,∞)),O\subseteq\mathcal{P}([0,\infty)),

lim sup⁡n→∞1nlog⁡Vn(F)≤sup⁡μ∈F∩Kh(μ),(25)\limsup_{n\to\infty}\frac{1}{n}\log V_n(F)\le\sup_{\mu\in F\cap K}h(\mu), \tag*{(25)}
lim inf⁡n→∞1nlog⁡Vn(O)≥sup⁡μ∈O∩Kh(μ);(26)\liminf_{n\to\infty}\frac{1}{n}\log V_n(O)\ge\sup_{\mu\in O\cap K}h(\mu); \tag*{(26)}

we use the conventions log⁡0=sup⁡∅=−∞\log0=\sup\varnothing=-\infty.

Proof. The upper bound. Let Y1,Y2,…Y_1,Y_2,\ldots be independent with law ρ\rho and put LnY=n−1∑i=1nδYiL_n^Y=n^{-1}\sum_{i=1}^n\delta_{Y_i}. Since ρ⊗n(dx)=e−∑ixidx\rho^{\otimes n}(\mathrm{d}x)=e^{-\sum_i x_i}\mathrm{d}x on the positive orthant, Lemma 2.1 gives the exact identity

Vn(A)=E[enm1(LnY)1{LnY∈A,En(LnY)≤1}].(27)V_n(A)=\mathbb{E}\left[e^{n m_1(L_n^Y)}\mathbf{1}_{\{L_n^Y\in A,E_n(L_n^Y)\le1\}}\right]. \tag*{(27)}

By Sanov’s theorem [3 Theorem 6.2.10], the sequence (LnY)(L_n^Y) satisfies an LDP for weak convergence, with speed nn and good rate function H(⋅∣ρ)H(\cdot\mid\rho). Fix R,δ>0R,\delta>0 and set Aδ=F∩Cp∩KδA_\delta=F\cap C_p\cap K_\delta. This set is compact. By (2.6) and (2.7), for all sufficiently large nn,

Vn(F)≤enR1−pE[enm1,R(LnY)1{LnY∈Aδ}].V_n(F)\le e^{nR^{1-p}}\mathbb{E}\left[e^{nm_{1,R}(L_n^Y)}\mathbf{1}_{\{L_n^Y\in A_\delta\}}\right].

The restricted Laplace upper bound for a bounded continuous functional on a closed set [4 Appendix B.2, pp. 321–322] therefore yields

lim sup⁡n→∞1nlog⁡Vn(F)≤R1−p+sup⁡μ∈Aδ(m1,R(μ)−H(μ∣ρ))≤R1−p+sup⁡μ∈Aδh(μ).\begin{aligned} \limsup_{n\to\infty}\frac{1}{n}\log V_n(F)\le R^{1-p}+\sup_{\mu\in A_\delta}\left(m_{1,R}(\mu)-H(\mu\mid\rho)\right) \\ &\le R^{1-p}+\sup_{\mu\in A_\delta}h(\mu). \end{aligned}

After letting R→∞R\to\infty, we consider the compact sets A1/kA_{1/k}, which decrease to F∩KF\cap K. If this intersection is nonempty, upper semicontinuity of hh and compactness imply

lim⁡k→∞sup⁡μ∈A1/kh(μ)=sup⁡μ∈F∩Kh(μ).\lim_{k\to\infty}\sup_{\mu\in A_{1/k}}h(\mu)=\sup_{\mu\in F\cap K}h(\mu).

For completeness, choose a maximizer in each A1/kA_{1/k} and pass to a convergent subsequence in CpC_p; its limit lies in F∩KF\cap K, and upper semicontinuity gives the nontrivial inequality. This argument also applies when the supremum is −∞-\infty. If F∩KF\cap K is empty, compactness implies that A1/kA_{1/k} is empty for some kk. This proves (3.1).

The lower bound for bounded support and strict feasibility. Let μ=f dx∈O\mu=f\,\mathrm{d}x\in O have support in [0,M][0,M], finite entropy, and E(μ)<1\mathcal{E}(\mu)<1. For independent random variables ZiZ_i with law μ\mu, their empirical measures LnZL_n^Z converge weakly to μ\mu almost surely. This follows, for example, by applying the strong law to the indicators of intervals with rational endpoints and using monotonicity of distribution functions. Quantile convergence and the bound

0≤wn(t)QLnZ(t)p≤Mpw(t)0\le w_n(t)Q_{L_n^Z}(t)^p\le M^p w(t)

give En(LnZ)→E(μ)<1\mathcal{E}_n(L_n^Z)\to\mathcal{E}(\mu)<1 almost surely by dominated convergence. Moreover, finite entropy implies flog⁡f∈L1f\log f\in L^1, so

−1n∑i=1nlog⁡f(Zi)⟶h(μ)almost surely.-\frac{1}{n}\sum_{i=1}^n\log f(Z_i)\longrightarrow h(\mu)\qquad\text{almost surely}.

Fix ε>0\varepsilon>0. Let BnB_n consist of the vectors zz for which f(zi)>0f(z_i)>0 for all ii and

Lz∈O,En(Lz)≤1,−1n∑i=1nlog⁡f(zi)≥h(μ)−ε.L_z\in O,\qquad\mathcal{E}_n(L_z)\le1,\qquad-\frac{1}{n}\sum_{i=1}^n\log f(z_i)\ge h(\mu)-\varepsilon.

Then μ⊗n(Bn)→1\mu^{\otimes n}(B_n)\to1, and

μ⊗n(Bn)=∫Bn∏i=1nf(zi) dz≤e−n(h(μ)−ε)Vn(O).\mu^{\otimes n}(B_n)=\int_{B_n}\prod_{i=1}^n f(z_i)\,\mathrm{d}z\le e^{-n(h(\mu)-\varepsilon)}V_n(O).

Taking logarithms and then letting ε↓0\varepsilon\downarrow0 shows that

lim inf⁡n→∞1nlog⁡Vn(O)≥h(μ).(28)\liminf_{n\to\infty}\frac{1}{n}\log V_n(O)\ge h(\mu). \tag*{(28)}

Approximation of an arbitrary admissible measure. Let μ=f dx∈O∩K\mu=f\,\mathrm{d}x\in O\cap K have finite entropy. For zM=μ([0,M])>0z_M=\mu([0,M])>0, let μM\mu_M be its conditional law on [0,M][0,M], with density fM=zM−1f1[0,M]f_M=z_M^{-1}f\mathbf{1}_{[0,M]}. Then μM⇒μ\mu_M\Rightarrow\mu, and

FμM(x)=min⁡{Fμ(x)/zM,1}≥Fμ(x).F_{\mu_M}(x)=\min\{F_\mu(x)/z_M,1\}\ge F_\mu(x).

Consequently, QμM≤QμQ_{\mu_M} \le Q_\mu and E(μM)≤E(μ)≤1\mathcal{E}(\mu_M) \le\mathcal{E}(\mu) \le1. Also,

h(μM)=−1zM∫0Mflog⁡f dx+log⁡zM⟶h(μ).(29)h(\mu_M)=-\frac{1}{z_M}\int_0^M f\log f\,\mathrm{d}x+\log z_M\longrightarrow h(\mu). \tag*{(29)}

Here zM↑1z_M \uparrow1, and flog⁡f∈L1f\log f \in L^1 by finite entropy.

For c>0c>0, write cμc\mu for the pushforward of μ\mu by x↦cxx\mapsto cx. Directly from the density and quantile formulas,

E(cμ)=cpE(μ),h(cμ)=h(μ)+log⁡c.(30)\mathcal{E}(c\mu)=c^p\mathcal{E}(\mu),\qquad h(c\mu)=h(\mu)+\log c. \tag*{(30)}

For large MM and then c<1c<1 sufficiently close to one, the measure cμMc\mu_M belongs to OO, has bounded support and energy strictly below one, and has entropy arbitrarily close to h(μ)h(\mu). Apply (3.4) and take the supremum over the finite-entropy measures in O∩KO\cap K. If there are no such measures, the lower bound is vacuous. This proves (3.2).

Proof of Theorem A. Apply Proposition 3.1 to the whole space to obtain

lim⁡n→∞1nlog⁡vol⁡n(Dn+)=h∗.(31)\lim_{n\to\infty}\frac{1}{n}\log\operatorname{vol}_n(D_n^+)=h^*. \tag*{(31)}

Since U(n)U^{(n)} is uniform on Dn+D_n^+,

P(Ln∈A)=Vn(A)vol⁡n(Dn+).\mathbb{P}(L_n\in A)=\frac{V_n(A)}{\operatorname{vol}_n(D_n^+)}.

Subtracting (3.7) from the restricted-volume bounds gives the LDP upper bound for closed sets and lower bound for open sets, with rate function (1.4). The rate function is nonnegative. For every finite aa,

{I≤a}={μ∈K:h(μ)≥h∗−a}\{I\le a\}=\{\mu\in K:h(\mu)\ge h^*-a\}

is a closed subset of the compact set KK, by upper semicontinuity of hh. Thus II is a good rate function.

Finally, orthant symmetry and (1.1) give

2nvol⁡n(Dn+)=vol⁡n(Dn)=nn/qvol⁡n(Bq,pn).2^n\operatorname{vol}_n(D_n^+)=\operatorname{vol}_n(D_n)=n^{n/q}\operatorname{vol}_n(B_{q,p}^n).

Taking nnth roots and using (3.7) proves (1.5).

The entropy maximizer and empirical limits

Existence of a maximizer follows from compactness of KK and upper semicontinuity of the entropy. To prove uniqueness, we interpolate between quantiles. This requires an entropy formula that remains valid when the quantile has a singular part.

Entropy in quantile coordinates

Lemma 4.1 (Quantile entropy formula). Let Q:(0,1)→[0,∞)Q:(0,1)\to[0,\infty) be nondecreasing and let μ=Q#Leb⁡(0,1)\mu=Q_\#\operatorname{Leb}_{(0,1)}. Then

μ≪Leb⁡⟺Q′>0almost everywhere.\mu\ll\operatorname{Leb}\quad\Longleftrightarrow\quad Q'>0\quad\text{almost everywhere}.

If these conditions hold and μ\mu has finite first moment, then

h(μ)=∫01log⁡Q′(t) dt(32)h(\mu)=\int_0^1\log Q'(t)\,\mathrm{d}t \tag*{(32)}

where both sides are allowed to take the value −∞-\infty. In particular, if h(μ)∈Rh(\mu)\in\mathbb{R}, then log⁡Q′∈L1(0,1)\log Q'\in L^1(0,1).

Proof. Suppose μ=f dx\mu=f\,\mathrm{d}x, and let FF be its distribution function. Since μ\mu has no atoms, QQ is strictly increasing: an interval on which QQ is constant would create an atom. The pushforward identity therefore gives F(Q(t))=tF(Q(t))=t for 0<t<10<t<1. The equality F′=fF'=f holds outside a Lebesgue-null set. Its preimage under QQ has measure zero because Q#Leb⁡(0,1)=μ≪Leb⁡Q_{\#}\operatorname{Leb}_{(0,1)}=\mu\ll\operatorname{Leb}. Since QQ is differentiable with finite derivative almost everywhere, the chain rule gives

f(Q(t))Q′(t)=1for almost every t∈(0,1).(33)f(Q(t))Q'(t)=1 \qquad\text{for almost every }t\in(0,1). \tag*{(33)}

Here f(Q(t))f(Q(t)) is positive and finite almost everywhere: the sets where f=0f=0 or f=∞f=\infty have μ\mu-measure zero after choosing a finite almost everywhere version of the density. This proves Q′>0Q'>0 almost everywhere.

Conversely, assume Q′>0Q'>0 almost everywhere. Let NN be a Lebesgue-null Borel set, and let DD be the set of points in Q−1(N)Q^{-1}(N) at which Q′Q' exists and lies in (0,∞)(0,\infty). Then Q−1(N)∖DQ^{-1}(N)\setminus D is null. The restriction Q∣DQ|_D is injective and differentiable relative to DD. The one-dimensional area formula on a measurable domain [5 Theorem 263D(iv)] gives

∫DQ′(t) dt=Leb⁡(Q(D))≤Leb⁡(N)=0.\int_D Q'(t)\,\mathrm{d}t=\operatorname{Leb}(Q(D))\leq\operatorname{Leb}(N)=0.

Thus DD is null and μ(N)=Leb⁡(Q−1(N))=0\mu(N)=\operatorname{Leb}(Q^{-1}(N))=0, proving absolute continuity of μ\mu. This argument does not assume absolute continuity of QQ.

Under these equivalent conditions, (33) holds. Hence log⁡Q′=−log⁡f∘Q\log Q'=-\log f\circ Q almost everywhere. Apply the pushforward identity to the positive and negative parts separately. If μ\mu has finite first moment, (2.8) ensures that the entropy is well defined, and the resulting equality is (4.1).

Proposition 4.2 (Existence and uniqueness). There is a unique entropy maximizer μ∗\mu_* in KK, and E(μ∗)=1\mathcal{E}(\mu_*)=1.

Proof. Lemma 2.2, upper semicontinuity of hh, and h∗∈Rh^*\in\mathbb{R} give existence and finite maximizing entropy. Every maximizer has energy one: otherwise, a dilation by some c>1c>1 would still belong to KK and would increase the entropy by log⁡c\log c, by (3.6).

Suppose that μ0≠μ1\mu_0\ne\mu_1 are two maximizers, with quantiles Q0,Q1Q_0,Q_1. For 0<θ<10<\theta<1, define

Qθ=(1−θ)Q0+θQ1,μθ=(Qθ)#Leb⁡(0,1).Q_\theta=(1-\theta)Q_0+\theta Q_1,\qquad\mu_\theta=(Q_\theta)_{\#}\operatorname{Leb}_{(0,1)}.

The function QθQ_\theta is nondecreasing and left-continuous, so it is the quantile of μθ\mu_\theta. Since Q0Q_0 and Q1Q_1 differ on a set of positive measure, strict convexity of u↦upu\mapsto u^p and positivity of ww give

E(μθ)<(1−θ)E(μ0)+θE(μ1)=1.(34)\mathcal{E}(\mu_\theta)<(1-\theta)\mathcal{E}(\mu_0)+\theta\mathcal{E}(\mu_1)=1. \tag*{(34)}

By Lemma 4.1, Q0′,Q1′>0Q_0',Q_1'>0 almost everywhere, and hence Qθ′>0Q_\theta'>0 almost everywhere. Since μθ∈K⊆Cp\mu_\theta\in K\subseteq C_p, it has finite first moment. Thus Lemma 4.1 and concavity of the logarithm yield

h(μθ)=∫01log⁡Qθ′(t) dt≥(1−θ)h(μ0)+θh(μ1)=h∗.h(\mu_\theta)=\int_0^1\log Q_\theta'(t)\,\mathrm{d}t\geq(1-\theta)h(\mu_0)+\theta h(\mu_1)=h^*.

In view of (4.3), dilating μθ\mu_\theta slightly gives an admissible measure with entropy greater than h∗h^*, a contradiction.

Empirical measures and fixed coordinates

Proof of Theorem B(iii). Fix ε>0\varepsilon> 0 and put Fε:={μ:d(μ,μ∗)≥ε}F_\varepsilon:= \{\mu: d(\mu,\mu_*) \ge\varepsilon\}. By compactness, upper semicontinuity, and uniqueness of the maximizer,

sup⁡μ∈Fε∩Kh(μ)<h∗.\sup_{\mu\in F_\varepsilon\cap K} h(\mu) < h^*.

with the same conclusion if the intersection is empty. Theorem A gives

lim sup⁡n→∞1nlog⁡P(d(Ln,μ∗)≥ε)≤sup⁡μ∈Fε∩Kh(μ)−h∗<0.(35)\limsup_{n \to\infty} \frac{1}{n} \log\mathbb{P}\left(d(L_n,\mu_*) \ge\varepsilon\right) \le\sup_{\mu\in F_\varepsilon\cap K} h(\mu) - h^* < 0. \tag*{(35)}

This proves Ln→μ∗L_n \to\mu_* in probability.

To pass to the full scaled ball, use its invariance under coordinatewise sign changes. We may represent

X(n)=d(ε1U1(n),…,εnUn(n)),X^{(n)} \mathrel{\overset{d}{=}} \left(\varepsilon_1 U_1^{(n)},\ldots,\varepsilon_n U_n^{(n)}\right),

where the εi\varepsilon_i are independent symmetric signs, independent of U(n)U^{(n)}. Orthant boundaries have Lebesgue measure zero, so this representation is exact in law. For f∈Cb(R)f \in C_b(\mathbb{R}), put fˉ(u)=(f(u)+f(−u))/2\bar{f}(u) = (f(u)+f(-u))/2 for u≥0u \ge0. Conditional on U(n)U^{(n)}, the centered sign fluctuation

Zn=1n∑i=1n(f(εiUi(n))−fˉ(Ui(n)))Z_n = \frac{1}{n}\sum_{i=1}^{n}\left(f(\varepsilon_i U_i^{(n)})-\bar{f}(U_i^{(n)})\right)

has mean zero and variance at most ∥f∥∞2/n\lVert f\rVert_\infty^2/n. Hence Zn→0Z_n \to0 in probability, whereas

1n∑i=1nfˉ(Ui(n))→P∫fˉ dμ∗=∫f dνq,p.\frac{1}{n}\sum_{i=1}^{n}\bar{f}(U_i^{(n)}) \xrightarrow{\mathbb{P}} \int\bar{f}\,d\mu_* = \int f\,d\nu_{q,p}.

Finite intersections of neighborhoods specified by bounded continuous test functions form a neighborhood basis for weak convergence. Thus convergence of these integrals in probability gives (1.9). The density formula (1.10) follows from the regularity established in Section 5.

Proof of Corollary 1.1. Let Mn=n−1∑i=1nδXi(n)M_n=n^{-1}\sum_{i=1}^{n}\delta_{X_i^{(n)}} and fix f1,…,fk∈Cb(R)f_1,\ldots,f_k \in C_b(\mathbb{R}). Theorem B and boundedness give

E∏j=1k∫fj dMn⟶∏j=1k∫fj dνq,p.\mathbb{E}\prod_{j=1}^{k}\int f_j\,dM_n \longrightarrow\prod_{j=1}^{k}\int f_j\,d\nu_{q,p}.

For n≥kn \ge k, expanding the product gives

E∏j=1k∫fj dMn=1nk∑i1,…,ik=1nE∏j=1kfj(Xij(n)).\mathbb{E}\prod_{j=1}^{k}\int f_j\,dM_n = \frac{1}{n^k}\sum_{i_1,\ldots,i_k=1}^{n} \mathbb{E}\prod_{j=1}^{k} f_j\left(X_{i_j}^{(n)}\right).

Write (n)k=n(n−1)⋯(n−k+1)(n)_k=n(n-1)\cdots(n-k+1). By exchangeability, every term with distinct indices has expectation E∏j=1kfj(Xj(n))\mathbb{E}\prod_{j=1}^{k}f_j(X_j^{(n)}). The total contribution of repeated indices therefore gives the estimate

∣E∏j=1kfj(Xj(n))−E∏j=1k∫fj dMn∣≤2(1−(n)knk)∏j=1k∥fj∥∞⟶0.\left|\mathbb{E}\prod_{j=1}^{k}f_j(X_j^{(n)})-\mathbb{E}\prod_{j=1}^{k}\int f_j\,dM_n\right| \le2\left(1-\frac{(n)_k}{n^k}\right)\prod_{j=1}^{k}\lVert f_j\rVert_\infty\longrightarrow0.

The one-coordinate marginals converge to νq,p\nu_{q,p} and are tight, hence the kk-coordinate laws are tight as well. Every subsequential limit has the displayed product integrals of νq,p⊗k\nu_{q,p}^{\otimes k}, which determine this product measure. This proves the claim.

The variational equation and compact support

We now identify the entropy maximizer. The first step justifies working with the derivative of its quantile.

Lemma 5.1. Let Q=Qμ∗Q=Q_{\mu_*}. Then Q′>0Q'>0 almost everywhere on (0,1)(0,1), and for every t∈(0,1)t\in(0,1),

Q(t)=∫0tQ′(s) ds.(36)Q(t)=\int_0^t Q'(s)\,\mathrm{d}s. \tag*{(36)}

In particular, after setting Q(0)=0Q(0)=0, the function QQ is absolutely continuous on [0,b][0,b] for every b∈(0,1)b\in(0,1).

Proof. By Lemma 4.1, v=Q′>0v=Q'>0 almost everywhere. Monotonicity gives

0≤∫atv(s) ds≤Q(t)−Q(a)≤Q(t),0<a<t<1.0\le\int_a^t v(s)\,\mathrm{d}s\le Q(t)-Q(a)\le Q(t),\qquad0<a<t<1.

Consequently, Q~(t)=∫0tv(s) ds\widetilde{Q}(t)=\int_0^t v(s)\,\mathrm{d}s is finite, satisfies 0≤Q~≤Q0\le\widetilde{Q}\le Q, and has derivative vv almost everywhere. Since Q~\widetilde{Q} is continuous and strictly increasing, it is the quantile function of its induced law μ~\widetilde{\mu}. The inequality Q~≤Q\widetilde{Q}\le Q implies μ~∈K\widetilde{\mu}\in K, and the quantile entropy formula gives

h(μ~)=∫01log⁡v(t) dt=h(μ∗)=h∗.h(\widetilde{\mu})=\int_0^1\log v(t)\,\mathrm{d}t=h(\mu_*)=h^*.

Uniqueness of the maximizer gives μ~=μ∗\widetilde{\mu}=\mu_*. Since Q~\widetilde{Q} and QQ are the corresponding quantiles, they agree on (0,1)(0,1). This removes any initial offset, jumps, and singular continuous part of QQ.

Admissible variations

For a quantile PP, write E(P)\mathcal{E}(P) and h(P)h(P) for the energy and entropy of its induced law. If 0<E(P)<∞0<\mathcal{E}(P)<\infty and h(P)∈Rh(P)\in\mathbb{R}, scaling gives

h(P)−1plog⁡E(P)≤h∗.h(P)-\frac{1}{p}\log\mathcal{E}(P)\le h^*.

Since log⁡u≤u−1\log u\le u-1 for u>0u>0, the functional

Φ(P)=h(P)−1pE(P)(37)\Phi(P)=h(P)-\frac{1}{p}\mathcal{E}(P) \tag*{(37)}

therefore satisfies

Φ(P)≤h(P)−1plog⁡E(P)−1p≤h∗−1p=Φ(Q).(38)\Phi(P)\le h(P)-\frac{1}{p}\log\mathcal{E}(P)-\frac{1}{p}\le h^*-\frac{1}{p}=\Phi(Q). \tag*{(38)}

This observation replaces the constrained problem by an unconstrained one without introducing an unknown Lagrange multiplier.

Proposition 5.2. The maximizing quantile belongs to C1([0,1))C^1([0,1)) and satisfies (1.6). If r=lim⁡s↑1Q(s)∈(0,∞]r=\lim_{s\uparrow1}Q(s)\in(0,\infty], then G=Q−1G=Q^{-1} satisfies

G′′(x)=−xp−1(1−G(x))p/q−1G′(x),0<x<r.(39)G''(x)=-x^{p-1}(1-G(x))^{p/q-1}G'(x),\qquad0<x<r. \tag*{(39)}

Proof. Write v=Q′v=Q' and use Lemma 5.1. Let ψ:(0,1)→R\psi:(0,1)\to\mathbb{R} be bounded and measurable. For ε∈R\varepsilon\in\mathbb{R} sufficiently close to zero, define

vε=v(1+εψ),Qε(t)=∫0tvε(s) ds.v_\varepsilon=v(1+\varepsilon\psi),\qquad Q_\varepsilon(t)=\int_0^t v_\varepsilon(s)\,\mathrm{d}s.

Put M=∥ψ∥∞M=\|\psi\|_{\infty} and assume ∣ε∣M<1/2|\varepsilon|M<1/2. Then vε>0v_{\varepsilon}>0 almost everywhere and

(1−∣ε∣M)Q≤Qε≤(1+∣ε∣M)Q.(1-|\varepsilon|M)Q\le Q_{\varepsilon}\le(1+|\varepsilon|M)Q.

Thus QεQ_{\varepsilon} induces a measure with finite positive energy, hence with finite first moment. Moreover,

log⁡vε=log⁡v+log⁡(1+εψ)∈L1(0,1),\log v_{\varepsilon}=\log v+\log(1+\varepsilon\psi)\in L^{1}(0,1),

so it has finite entropy by Lemma 4.1. Both positive and negative small ε\varepsilon are admissible in (5.3).

Set R(t)=∫0tv(s)ψ(s) dsR(t)=\int_{0}^{t}v(s)\psi(s)\,ds, so Qε=Q+εRQ_{\varepsilon}=Q+\varepsilon R and ∣R∣≤MQ|R|\le MQ. Differentiation under the integral sign is justified by the bounds

∣ψ1+εψ∣≤2M,∣wQεp−1R∣≤M(3/2)p−1wQp.\left|\frac{\psi}{1+\varepsilon\psi}\right|\le2M,\qquad|wQ_{\varepsilon}^{p-1}R|\le M(3/2)^{p-1}wQ^{p}.

The second majorant is integrable because E(Q)=1\mathcal{E}(Q)=1. Since Φ(Qε)\Phi(Q_{\varepsilon}) has a maximum at zero, its derivative there vanishes:

0=∫01ψ(s) ds−∫01w(t)Q(t)p−1(∫0tv(s)ψ(s) ds) dt.(40)0=\int_{0}^{1}\psi(s)\,ds-\int_{0}^{1}w(t)Q(t)^{p-1}\left(\int_{0}^{t}v(s)\psi(s)\,ds\right)\,dt. \tag*{(40)}

The double integral is absolutely integrable, since

∫01∫0tw(t)Q(t)p−1v(s)∣ψ(s)∣ ds dt≤M∫01w(t)Q(t)p dt=M.\int_{0}^{1}\int_{0}^{t}w(t)Q(t)^{p-1}v(s)|\psi(s)|\,ds\,dt\le M\int_{0}^{1}w(t)Q(t)^{p}\,dt=M.

Fubini’s theorem turns (5.5) into

∫01ψ(s)[1−v(s)∫s1w(t)Q(t)p−1 dt] ds=0.\int_{0}^{1}\psi(s)\left[1-v(s)\int_{s}^{1}w(t)Q(t)^{p-1}\,dt\right]\,ds=0.

Because ψ\psi is arbitrary,

v(s)A(s)=1almost everywhere,A(s)=∫s1w(t)Q(t)p−1 dt.(41)v(s)A(s)=1\quad\text{almost everywhere},\qquad A(s)=\int_{s}^{1}w(t)Q(t)^{p-1}\,dt. \tag*{(41)}

We now show that Q∈C1([0,1))Q\in C^{1}([0,1)) and that (1.6) holds for every s∈[0,1)s\in[0,1). Hölder’s inequality gives

A(0)≤(∫01wQp)(p−1)/p(∫01w)1/p=(qp)1/p<∞.(42)A(0)\le\left(\int_{0}^{1}wQ^{p}\right)^{(p-1)/p}\left(\int_{0}^{1}w\right)^{1/p}=\left(\frac{q}{p}\right)^{1/p}<\infty. \tag*{(42)}

Hence AA is continuous on [0,1][0,1]. Since Q(t)>0Q(t)>0 for t>0t>0, we have A(s)>0A(s)>0 for s<1s<1, whereas A(1)=0A(1)=0. By (5.1) and (5.6),

Q(t)=∫0t1A(s) ds.Q(t)=\int_{0}^{t}\frac{1}{A(s)}\,ds.

This proves Q∈C1([0,1))Q\in C^{1}([0,1)) and (1.6). On (0,1)(0,1),

A′(s)=−w(s)Q(s)p−1.A'(s)=-w(s)Q(s)^{p-1}.

The inverse function theorem now yields G′(x)=A(G(x))G'(x)=A(G(x)), and therefore

G′′(x)=A′(G(x))G′(x)=−xp−1(1−G(x))p/q−1G′(x).G''(x)=A'(G(x))G'(x)=-x^{p-1}(1-G(x))^{p/q-1}G'(x).

This proves (5.4).

The support and endpoint conditions

Proposition 5.3. The support of μ∗\mu_* is [0,rp,q][0,r_{p,q}], with rp,qr_{p,q} satisfying (7). Its distribution function satisfies

G(0)=0,G(rp,q)=1,G′(0)>0,G′(rp,q)=0.G(0)=0,\qquad G(r_{p,q})=1,\qquad G'(0)>0,\qquad G'(r_{p,q})=0.

Moreover, G∈C1([0,rp,q])∩C2((0,rp,q))G\in C^1([0,r_{p,q}])\cap C^2((0,r_{p,q})), and its density is strictly decreasing on (0,rp,q)(0,r_{p,q}).

Proof. By monotonicity of QQ, for 0<s<10<s<1,

1Q′(s)=∫s1w(t)Q(t)p−1 dt≥qpQ(s)p−1(1−s)p/q.\frac{1}{Q'(s)}=\int_s^1 w(t)Q(t)^{p-1}\,\mathrm{d}t\ge\frac{q}{p}Q(s)^{p-1}(1-s)^{p/q}.

Since Q(s)>0Q(s)>0, it follows that

(Q(s)p)′=pQ(s)p−1Q′(s)≤p2q(1−s)−p/q.(Q(s)^p)'=pQ(s)^{p-1}Q'(s)\le\frac{p^2}{q}(1-s)^{-p/q}.

Integrating from zero and using Q(0)=0Q(0)=0 gives

Q(s)p≤p2q−p(1−(1−s)1−p/q).(43)Q(s)^p\le\frac{p^2}{q-p}\left(1-(1-s)^{1-p/q}\right). \tag*{(43)}

Thus rp,q=Q(1−)<∞r_{p,q}=Q(1-)<\infty and (7) holds.

The continuous strictly increasing quantile maps (0,1)(0,1) onto (0,rp,q)(0,r_{p,q}). Its induced law therefore has support [0,rp,q][0,r_{p,q}], and the inverse GG extends continuously with values zero and one at the endpoints. Since G′=A∘GG'=A\circ G in the interior and AA is continuous on [0,1][0,1],

lim⁡x↓0G′(x)=A(0)>0,lim⁡x↑rp,qG′(x)=A(1)=0.\lim_{x\downarrow0}G'(x)=A(0)>0,\qquad\lim_{x\uparrow r_{p,q}}G'(x)=A(1)=0.

The one-sided endpoint derivatives have these values by the mean value theorem. This proves the stated regularity. Finally, (39) has a strictly negative right-hand side on (0,rp,q)(0,r_{p,q}), so G′G' is strictly decreasing there.

Uniqueness of the boundary-value solution

The preceding results give an admissible solution of (8). To obtain a characterization, we must also show that every such solution yields the entropy maximizer. The following argument includes competitors whose quantiles have a singular part.

Proposition 5.4 (Converse characterization). Let (r,G)(r,G) satisfy the regularity, positivity and boundary assumptions in Theorem B(ii), as well as (8). Then r=rp,qr=r_{p,q} and GG is the distribution function of μ∗\mu_*.

Proof. Let Q=G−1Q=G^{-1} on (0,1)(0,1) and set A(s)=G′(Q(s))=1/Q′(s)A(s)=G'(Q(s))=1/Q'(s). The differential equation gives

A′(s)=−w(s)Q(s)p−1.A'(s)=-w(s)Q(s)^{p-1}.

Integrating from ss to b<1b<1 and letting b↑1b\uparrow1, the terminal condition on G′G' implies

A(s)=∫s1w(t)Q(t)p−1 dt,Q′(s)A(s)=1.(44)A(s)=\int_s^1 w(t)Q(t)^{p-1}\,\mathrm{d}t,\qquad Q'(s)A(s)=1. \tag*{(44)}

No finite-energy assumption is needed for this identity: the improper integral equals the finite quantity A(s)A(s). We have Q(0+)=0Q(0+)=0. The same monotonicity estimate as in Proposition 5.3, integrated first from ε\varepsilon to ss and then with ε↓0\varepsilon\downarrow0, yields (5.8). Consequently,

r=Q(1−)≤(p2q−p)1/p<∞.r=Q(1-)\leq\left(\frac{p^2}{q-p}\right)^{1/p}<\infty.

In particular, A(0)=∫01wQp−1A(0)=\int_0^1 wQ^{p-1} is finite and positive, and (5.9) implies Q∈C1([0,1))Q\in C^1([0,1)) with Q(0)=0Q(0)=0. Therefore Q(t)=∫0tQ′(s) dsQ(t)=\int_0^t Q'(s)\,\mathrm{d}s. Tonelli’s theorem now gives the normalization

E(Q)=∫01w(t)Q(t)p−1(∫0tQ′(s) ds)dt=∫01Q′(s)(∫s1w(t)Q(t)p−1 dt)ds=1.(45)\begin{aligned} \mathcal{E}(Q)=\int_0^1w(t)Q(t)^{p-1}\left(\int_0^tQ'(s)\,\mathrm{d}s\right)\mathrm{d}t \\ &=\int_0^1Q'(s)\left(\int_s^1w(t)Q(t)^{p-1}\,\mathrm{d}t\right)\mathrm{d}s=1. \tag*{(45)} \end{aligned}

The law induced by QQ also has finite entropy. Indeed, Q′≥1/A(0)>0Q'\geq1/A(0)>0, Q′Q' is bounded on [0,1/2][0,1/2], and for s≥1/2s\geq1/2,

A(s)≥qpQ(1/2)p−1(1−s)p/q.A(s)\geq\frac{q}{p}Q(1/2)^{p-1}(1-s)^{p/q}.

Thus Q′(s)≤C(1−s)−p/qQ'(s)\leq C(1-s)^{-p/q} near one, and log⁡Q′∈L1(0,1)\log Q'\in L^1(0,1). The quantile entropy formula from Lemma 4.1 applies.

Let PP be the quantile of any μ∈K\mu\in K with finite entropy. By monotonicity,

∫0tP′(s) ds≤P(t),0<t<1,\int_0^tP'(s)\,\mathrm{d}s\leq P(t),\qquad0<t<1,

whether or not PP is absolutely continuous. Using Tonelli, Hölder, and (5.10), we obtain

∫01P′(s)A(s) ds=∫01w(t)Q(t)p−1(∫0tP′(s) ds)dt≤∫01w(t)Q(t)p−1P(t) dt≤E(Q)(p−1)/pE(P)1/p≤1.\begin{aligned} \int_0^1P'(s)A(s)\,\mathrm{d}s &=\int_0^1w(t)Q(t)^{p-1}\left(\int_0^tP'(s)\,\mathrm{d}s\right)\mathrm{d}t\\ &\leq\int_0^1w(t)Q(t)^{p-1}P(t)\,\mathrm{d}t\\ &\leq\mathcal{E}(Q)^{(p-1)/p}\mathcal{E}(P)^{1/p}\leq1. \end{aligned}

Finally, log⁡u≤u−1\log u\leq u-1 gives

h(μ)−h(Q)=∫01log⁡P′(s)Q′(s) ds≤∫01P′(s)A(s) ds−1≤0.(46)h(\mu)-h(Q)=\int_0^1\log\frac{P'(s)}{Q'(s)}\,\mathrm{d}s\leq\int_0^1P'(s)A(s)\,\mathrm{d}s-1\leq0. \tag*{(46)}

Both logarithms are integrable, and the right-hand side is finite by the preceding estimate, so every integral in this comparison is justified. Measures of entropy −∞-\infty do not affect maximality. Hence the law induced by QQ is an entropy maximizer in KK, and Proposition 4.2 identifies it with μ∗\mu_*. Its quantile determines rr and GG uniquely. □

Completion of the proof of Theorem B. Proposition 4.2 gives the unique maximizer and energy normalization. Proposition 5.2 gives (1.6) and the differential equation. Propositions 5.3 and 5.4 give the support estimate, boundary regularity, and uniqueness of the boundary-value solution. The empirical convergence was proved in § 4. The continuous density in (1.10) follows from G′∈C([0,rp,q])G'\in C([0,r_{p,q}]) and G′(rp,q)=0G'(r_{p,q})=0. □

Intersection volumes

Throughout this section, 1<p<q<∞1 < p < q < \infty are the fixed source parameters and 1≤s≤r<∞1 \le s \le r < \infty are the target parameters. We use the volume normalization and constants from (11)–(12). For 0<u<10 < u < 1, set

wr,s(u)=(1−u)s/r−1,Er,s(μ)=∫01wr,s(u)Qμ(u)s du.(47)w_{r,s}(u)=(1-u)^{s/r-1},\qquad\mathcal{E}_{r,s}(\mu)=\int_0^1 w_{r,s}(u)Q_\mu(u)^s\,\mathrm{d}u. \tag*{(47)}

The corresponding discrete weight and energy are

wn;r,s(u)=(1−j−1n)s/r−1ifj−1n<u≤jn,w_{n;r,s}(u)=\left(1-\frac{j-1}{n}\right)^{s/r-1}\quad\text{if}\quad\frac{j-1}{n}<u\le\frac{j}{n},
En;r,s(μ)=∫01wn;r,s(u)Qμ(u)s du.(48)\mathcal{E}_{n;r,s}(\mu)=\int_0^1 w_{n;r,s}(u)Q_\mu(u)^s\,\mathrm{d}u. \tag*{(48)}

As in Lemma 2.1,

En;r,s(Lx)=n−s/r∥x∥r,ss,x∈[0,∞)n.(49)\mathcal{E}_{n;r,s}(L_x)=n^{-s/r}\lVert x\rVert_{r,s}^s,\qquad x\in[0,\infty)^n. \tag*{(49)}

Also, 1≤wn;r,s≤wr,s1\le w_{n;r,s}\le w_{r,s} and wn;r,s→wr,sw_{n;r,s}\to w_{r,s} pointwise on (0,1)(0,1).

Moment bounds and Lorentz norms

We first strengthen the empirical convergence in Theorem B. For k≥1k\ge1, we use the one-dimensional quantile representation of the kk-Wasserstein distance,

Wk(μ,ν)=(∫01∣Qμ(t)−Qν(t)∣k dt)1/k,W_k(\mu,\nu)=\left(\int_0^1\lvert Q_\mu(t)-Q_\nu(t)\rvert^k\,\mathrm{d}t\right)^{1/k},

for probability measures with finite kkth moments.

Lemma 6.1. There is a constant c0>0c_0>0 such that

P(U1(n)≥t)≤2e−c0t,n∈N,t≥0.(50)\mathbb{P}(U_1^{(n)}\ge t)\le2e^{-c_0t},\qquad n\in\mathbb{N},\quad t\ge0. \tag*{(50)}

Consequently, for every 1≤k<∞1\le k<\infty,

Wk(Ln,μ∗)→P0.(51)W_k(L_n,\mu_*)\xrightarrow{\mathbb{P}}0. \tag*{(51)}

Proof. Permutation symmetry and the bound n−1∑i=1nxip≤1n^{-1}\sum_{i=1}^n x_i^p\le1 on Dn+D_n^+ give

E[(U1(n))p]≤1,EU1(n)≤1.\mathbb{E}\left[(U_1^{(n)})^p\right]\le1,\qquad\mathbb{E}U_1^{(n)}\le1.

Put Sn(t)=P(U1(n)≥t)S_n(t)=\mathbb{P}(U_1^{(n)}\ge t) and Ct=Dn+∩{x1≥t}C_t=D_n^+\cap\{x_1\ge t\}. Since Dn+D_n^+ is convex,

(1−θ)Ca+θCb⊆C(1−θ)a+θb,0<θ<1.(1-\theta)C_a+\theta C_b\subseteq C_{(1-\theta)a+\theta b},\qquad0<\theta<1.

For Sn(a),Sn(b)>0S_n(a),S_n(b)>0, the Brunn–Minkowski inequality, followed by the weighted arithmetic–geometric mean inequality, therefore yields

Sn((1−θ)a+θb)≥((1−θ)Sn(a)1/n+θSn(b)1/n)n≥Sn(a)1−θSn(b)θ.S_n((1-\theta)a+\theta b)\ge\left((1-\theta)S_n(a)^{1/n}+\theta S_n(b)^{1/n}\right)^n\ge S_n(a)^{1-\theta}S_n(b)^\theta.

The resulting log-concavity inequality is automatic if either value is zero. Thus SnS_n is log-concave. Since Sn(0)=1S_n(0)=1 and Sn(2)≤1/2S_n(2)\le1/2 by Markov’s inequality, log-concavity gives Sn(t)≤e−c0tS_n(t)\le e^{-c_0t} for t≥2t\ge2, with c0=(log⁡2)/2c_0=(\log2)/2. This proves (50) for all t≥0t\ge0. In particular,

sup⁡nE[(U1(n))a]<∞for every a>0.(52)\sup_n\mathbb{E}\left[(U_1^{(n)})^a\right]<\infty\qquad\text{for every }a>0. \tag*{(52)}

Fix k≥1k \ge1 and let TR(x)=min⁡{x,R}T_R(x)=\min\{x,R\}. Write LnR=(TR)#LnL_n^R=(T_R)_\#L_n and μ∗R=(TR)#μ∗\mu_*^R=(T_R)_\#\mu_*. The natural coupling and (6.6) imply

EWk(Ln,LnR)k≤E∫(x−R)+kLn(dx)≤1RE[(U1(n))k+1]≤CkR.\mathbb{E}W_k(L_n,L_n^R)^k \le\mathbb{E}\int(x-R)_+^k L_n(\mathrm{d}x) \le\frac{1}{R}\mathbb{E}\left[(U_1^{(n)})^{k+1}\right] \le\frac{C_k}{R}.

For fixed RR, Theorem B gives LnR→μ∗RL_n^R \to\mu_*^R weakly in probability. On [0,R][0,R], weak convergence implies WkW_k-convergence by almost everywhere quantile convergence and dominated convergence. Hence Wk(LnR,μ∗R)→0W_k(L_n^R,\mu_*^R)\to0 in probability. Moreover, Wk(μ∗R,μ∗)→0W_k(\mu_*^R,\mu_*)\to0 as R→∞R\to\infty, since μ∗\mu_* has compact support. Apply the triangle inequality, use Markov’s inequality for the first term, and then let R→∞R\to\infty to obtain (51). □

Proposition 6.2 (A law of large numbers for Lorentz norms). Let 1≤s≤r<∞1\le s\le r<\infty and let X(n)X^{(n)} be uniformly distributed on DnD_n. With mq,p;r,sm_{q,p;r,s} as in (1.13), we have 0<mq,p;r,s<∞0<m_{q,p;r,s}<\infty and

n−1/r∥X(n)∥r,s→Pmq,p;r,s.(53)n^{-1/r}\|X^{(n)}\|_{r,s}\xrightarrow{\mathbb{P}}m_{q,p;r,s}. \tag*{(53)}

Proof. Define

Fn;r,s(μ)=(∫01wn;r,s(t)Qμ(t)s dt)1/s.F_{n;r,s}(\mu)=\left(\int_0^1 w_{n;r,s}(t)Q_\mu(t)^s\,\mathrm{d}t\right)^{1/s}.

We have wn;r,s≤wr,sw_{n;r,s}\le w_{r,s} and pointwise convergence on (0,1)(0,1). Choose k>rk>r. Since wr,s∈Lk/(k−s)(0,1)w_{r,s}\in L^{k/(k-s)}(0,1), Minkowski’s and Hölder’s inequalities give, for measures with finite kkth moments,

∣Fn;r,s(μ)−Fn;r,s(ν)∣≤(∫01wn;r,s∣Qμ−Qν∣s)1/s≤∥wr,s∥Lk/(k−s)1/sWk(μ,ν).|F_{n;r,s}(\mu)-F_{n;r,s}(\nu)| \le\left(\int_0^1 w_{n;r,s}|Q_\mu-Q_\nu|^s\right)^{1/s} \le\|w_{r,s}\|_{L^{k/(k-s)}}^{1/s}W_k(\mu,\nu).

The constant is independent of nn. Dominated convergence gives Fn;r,s(μ∗)→mq,p;r,sF_{n;r,s}(\mu_*)\to m_{q,p;r,s}, because QQ is bounded and wr,sw_{r,s} is integrable. Thus Lemma 6.1 yields Fn;r,s(Ln)→mq,p;r,sF_{n;r,s}(L_n)\to m_{q,p;r,s} in probability. By (6.3), Fn;r,s(Ln)=n−1/r∥U(n)∥r,sF_{n;r,s}(L_n)=n^{-1/r}\|U^{(n)}\|_{r,s}. The absolute coordinates of X(n)X^{(n)} have the law of U(n)U^{(n)}, proving (6.7). Positivity and finiteness of mq,p;r,sm_{q,p;r,s} follow from Q(t)>0Q(t)>0 for t>0t>0, boundedness of QQ, and integrability of wr,sw_{r,s}. □

The intersection threshold

Proof of Corollary 1.2. Put

an=n1/qvol⁡n(Bq,pn)1/n,dn=n1/rvol⁡n(Br,sn)1/n.a_n=n^{1/q}\operatorname{vol}_n(B_{q,p}^n)^{1/n},\qquad d_n=n^{1/r}\operatorname{vol}_n(B_{r,s}^n)^{1/n}.

Then an→cq,pa_n\to c_{q,p} and dn→cr,sd_n\to c_{r,s}. If X(n)X^{(n)} is uniformly distributed on DnD_n, then X(n)/anX^{(n)}/a_n is uniform on B^q,pn\widehat{B}_{q,p}^n. Since the latter body has volume one,

vol⁡n(B^q,pn∩tB^r,sn)=P(n−1/r∥X(n)∥r,s≤tandn).(54)\operatorname{vol}_n(\widehat{B}_{q,p}^n\cap t\widehat{B}_{r,s}^n) =\mathbb{P}\left(n^{-1/r}\|X^{(n)}\|_{r,s}\le t\frac{a_n}{d_n}\right). \tag*{(54)}

Proposition 6.2 now proves (1.15) with the threshold (1.14). The bound τq,p;r,s≥1\tau_{q,p;r,s}\ge1 follows from

vol⁡n(B^q,pn∩tB^r,sn)≤vol⁡n(tB^r,sn)=tn.\operatorname{vol}_n(\widehat{B}_{q,p}^n\cap t\widehat{B}_{r,s}^n) \le\operatorname{vol}_n(t\widehat{B}_{r,s}^n)=t^n.

if the threshold were smaller than one, choosing tt strictly between it and one would contradict (1.15).

Suppose now that 1<s<r<∞1 < s < r < \infty, let μr,s\mu_{r,s} be the entropy maximizer for the target pair, and write m=mq,p;r,sm=m_{q,p;r,s}. The law of Y/mY/m, where Y∼μ∗Y\sim\mu_*, has target energy one and entropy h∗−log⁡mh^*-\log m. Hence

hr,s∗≥h∗−log⁡m.h^*_{r,s}\ge h^*-\log m.

If τq,p;r,s=1\tau_{q,p;r,s}=1, equality holds, and uniqueness of the target maximizer implies μr,s=(m−1id⁡)#μ∗\mu_{r,s}=(m^{-1}\operatorname{id})_\#\mu_*. Thus its distribution function satisfies Gr,s(x)=G(mx)G_{r,s}(x)=G(mx). Comparing the differential equations in Theorem B for the two parameter pairs and cancelling the positive derivative Gr,s′(x)G'_{r,s}(x) gives

mpxp−s(1−Gr,s(x))p/q−s/r=1m^p x^{p-s}(1-G_{r,s}(x))^{p/q-s/r}=1

on the interior of the target support. Letting x↓0x\downarrow0 first forces p=sp=s and then m=1m=1; the same identity then gives q=rq=r. Conversely, if (r,s)=(q,p)(r,s)=(q,p), the equality E(μ∗)=1\mathcal{E}(\mu_*)=1 yields mq,p;q,p=1m_{q,p;q,p}=1 and τq,p;q,p=1\tau_{q,p;q,p}=1. The assertion for t=1t=1 follows. The exponential asymptotics follow from Proposition 6.4 below.

Remark 6.3. Corollary 1.2 does not specify the limit at t=τq,p;r,st=\tau_{q,p;r,s}. A nontrivial critical limit requires control of the norm fluctuations and of the finer volume asymptotics. For identical parameter pairs, the threshold equals one and

vol⁡n(B^q,pn∩tB^q,pn)=min⁡{1,tn},\operatorname{vol}_n(\widehat B_{q,p}^n\cap t\widehat B_{q,p}^n)=\min\{1,t^n\},

so the critical intersection volume equals one in every dimension.

Logarithmic intersection volumes

For t>0t>0, we define the following three quantities

bt=tcq,pcr,s,b_t=t\frac{c_{q,p}}{c_{r,s}},
K(t)={μ∈K:Er,s(μ)≤bts},(55)K(t)=\{\mu\in K:\mathcal{E}_{r,s}(\mu)\le b_t^s\}, \tag*{(55)}
J(t)=h∗−sup⁡μ∈K(t)h(μ).J(t)=h^*-\sup_{\mu\in K(t)}h(\mu).

The following proposition identifies the variational quantity J(t)J(t) with the exponential decay rate appearing in Corollary 1.2.

Proposition 6.4 (Logarithmic intersection volumes). For every t>0t>0, the number J(t)J(t) is finite and nonnegative, and

lim⁡n→∞1nlog⁡vol⁡n(B^q,pn∩tB^r,sn)=−J(t).(56)\lim_{n\to\infty}\frac{1}{n}\log\operatorname{vol}_n(\widehat B_{q,p}^n\cap t\widehat B_{r,s}^n)=-J(t). \tag*{(56)}

Moreover, J(t)>0J(t)>0 for 0<t<τq,p;r,s0<t<\tau_{q,p;r,s}, whereas J(t)=0J(t)=0 for t≥τq,p;r,st\ge\tau_{q,p;r,s}.

Proof. Fix t>0t>0 and put

an=n1/qvol⁡n(B^q,pn)1/n,dn=n1/rvol⁡n(B^r,sn)1/n,bn=tandn⟶bt=tcq,pcr,s.a_n=n^{1/q}\operatorname{vol}_n(\widehat B_{q,p}^n)^{1/n},\qquad d_n=n^{1/r}\operatorname{vol}_n(\widehat B_{r,s}^n)^{1/n},\qquad b_n=t\frac{a_n}{d_n}\longrightarrow b_t=t\frac{c_{q,p}}{c_{r,s}}.

Define

An(t)={μ:En(μ)≤1,En;r,s(μ)≤bns},A_n(t)=\{\mu:\mathcal{E}_n(\mu)\le1,\mathcal{E}_{n;r,s}(\mu)\le b_n^s\},
Wn(t)=vol⁡n{x∈[0,∞)n:Lx∈An(t)}.W_n(t)=\operatorname{vol}_n\{x\in[0,\infty)^n:L_x\in A_n(t)\}.

Orthant symmetry and the definitions of ana_n and dnd_n give

vol⁡n(B^q,pn∩tB^r,sn)=Wn(t)vol⁡n(Dn+).(57)\operatorname{vol}_n(\widehat B_{q,p}^n\cap t\widehat B_{r,s}^n)=\frac{W_n(t)}{\operatorname{vol}_n(D_n^+)}. \tag*{(57)}

We prove that

lim⁡n→∞1nlog⁡Wn(t)=sup⁡μ∈K(t)h(μ).(58)\lim_{n\to\infty} \frac{1}{n}\log W_n(t)=\sup_{\mu\in K(t)} h(\mu). \tag*{(58)}

The upper bound. Since 1≤wn;r,s≤wr,s1\le w_{n;r,s}\le w_{r,s} and wn;r,s→wr,sw_{n;r,s}\to w_{r,s}, the quantile convergence and Fatou arguments in Lemma 2.2 show that Er,s\mathcal{E}_{r,s} and En;r,s\mathcal{E}_{n;r,s} are weakly lower semicontinuous and that

Er,s(μ)≤lim inf⁡n→∞En;r,s(μn)whenever μn⇒μ.\mathcal{E}_{r,s}(\mu)\le\liminf_{n\to\infty}\mathcal{E}_{n;r,s}(\mu_n)\qquad\text{whenever }\mu_n\Rightarrow\mu.

In particular, all constraints above are Borel measurable and K(t)K(t) is compact. For δ>0\delta>0, put

Aδ(t)=Cp∩{μ:d(μ,K(t))≤δ}.A_\delta(t)=C_p\cap\{\mu:d(\mu,K(t))\le\delta\}.

Then An(t)⊆Aδ(t)A_n(t)\subseteq A_\delta(t) for all sufficiently large nn. Indeed, any sequence contradicting this inclusion has a convergent subsequence in CpC_p; the two moving lower bounds and bn→btb_n\to b_t place its limit in K(t)K(t), a contradiction.

Let YiY_i be independent with law ρ\rho and write LnYL_n^Y for their empirical measures. The change of measure in (27) gives

Wn(t)=E[enm1(LnY)1{LnY∈An(t)}].W_n(t)=\mathbb{E}\left[e^{nm_1(L_n^Y)}\mathbf{1}_{\{L_n^Y\in A_n(t)\}}\right].

Consequently, for fixed R,δ>0R,\delta>0 and all sufficiently large nn,

Wn(t)≤enR1−pE[enm1,R(LnY)1{LnY∈Aδ(t)}].W_n(t)\le e^{nR^{1-p}}\mathbb{E}\left[e^{nm_{1,R}(L_n^Y)}\mathbf{1}_{\{L_n^Y\in A_\delta(t)\}}\right].

Sanov’s theorem and the restricted Laplace upper bound used in the proof of Proposition 3.1 yield

lim sup⁡n→∞1nlog⁡Wn(t)≤R1−p+sup⁡μ∈Aδ(t)(m1,R(μ)−H(μ∣ρ))≤R1−p+sup⁡μ∈Aδ(t)h(μ).\limsup_{n\to\infty}\frac{1}{n}\log W_n(t)\le R^{1-p}+\sup_{\mu\in A_\delta(t)}\left(m_{1,R}(\mu)-H(\mu\mid\rho)\right) \le R^{1-p}+\sup_{\mu\in A_\delta(t)}h(\mu).

Let R→∞R\to\infty and then δ↓0\delta\downarrow0. The compact sets Aδ(t)A_\delta(t) decrease to K(t)K(t), so upper semicontinuity of hh gives the upper bound in (58).

The lower bound. First let ν=f dx\nu=f\,dx have bounded support, finite entropy, and

E(ν)<1,Er,s(ν)<bts.\mathcal{E}(\nu)<1,\qquad\mathcal{E}_{r,s}(\nu)<b_t^s.

For independent ZiZ_i with law ν\nu, we have LnZ⇒νL_n^Z\Rightarrow\nu almost surely. Quantile convergence and dominated convergence give

En(LnZ)⟶E(ν),En;r,s(LnZ)⟶Er,s(ν)almost surely.\mathcal{E}_n(L_n^Z)\longrightarrow\mathcal{E}(\nu),\qquad \mathcal{E}_{n;r,s}(L_n^Z)\longrightarrow\mathcal{E}_{r,s}(\nu)\quad\text{almost surely}.

Here both weights ww and wr,sw_{r,s} are integrable. Thus LnZ∈An(t)L_n^Z\in A_n(t) eventually almost surely. Moreover, flog⁡f∈L1f\log f\in L^1, and the strong law gives −n−1∑i=1nlog⁡f(Zi)→h(ν)-n^{-1}\sum_{i=1}^n\log f(Z_i)\to h(\nu) almost surely. The density comparison from the lower bound in Proposition 3.1 now gives, for every ε>0\varepsilon>0,

Wn(t)≥en(h(ν)−ε)P(LnZ∈An(t),−1n∑i=1nlog⁡f(Zi)≥h(ν)−ε).W_n(t)\ge e^{n(h(\nu)-\varepsilon)}\mathbb{P}\left(L_n^Z\in A_n(t),-\frac{1}{n}\sum_{i=1}^n\log f(Z_i)\ge h(\nu)-\varepsilon\right).

The probability tends to one, hence lim inf⁡nn−1log⁡Wn(t)≥h(ν)\liminf_{n}n^{-1}\log W_n(t)\ge h(\nu).

For arbitrary μ∈K(t)\mu\in K(t) with finite entropy, let μM=μ( ⋅∣[0,M])\mu_M=\mu(\,\cdot\mid[0,M]) whenever μ([0,M])>0\mu([0,M])>0. As in the proof of Proposition 3.1, QμM≤QμQ_{\mu_M}\le Q_\mu and h(μM)→h(μ)h(\mu_M)\to h(\mu). For 0<c<10<c<1, the boundedly supported measure ν=cμM\nu=c\mu_M satisfies

E(ν)≤cp<1,Er,s(ν)≤csbts<bts,h(ν)=h(μM)+log⁡c.\mathcal{E}(\nu)\le c^p<1,\qquad \mathcal{E}_{r,s}(\nu)\le c^s b_t^s<b_t^s,\qquad h(\nu)=h(\mu_M)+\log c.

Apply the preceding lower bound, let M→∞M \to\infty and c↑1c \uparrow1, and take the supremum over μ∈K(t)\mu\in K(t) of finite entropy. Such measures exist: the uniform law on a sufficiently short interval [0,η][0,\eta] belongs to K(t)K(t) and has entropy log⁡η\log\eta. Since K(t)⊆CpK(t) \subseteq C_p, its entropy supremum is finite. This proves (6.12). Combining (3.7) and (6.11) now proves (6.10).

The set K(t)K(t) is compact, and the entropy is upper semicontinuous there. Since its supremum is finite and attained, uniqueness of the entropy maximizer on KK shows that J(t)=0J(t)=0 if and only if μ∗∈K(t)\mu_* \in K(t). By (1.13) and (1.14), this is equivalent to t≥τq,p;r,st \ge\tau_{q,p;r,s}. This also proves strict positivity of J(t)J(t) below the threshold.

In particular, if 1<s<r<∞1 < s < r < \infty and (r,s)≠(q,p)(r,s) \ne(q,p), then Corollary 1.2 gives τq,p;r,s>1\tau_{q,p;r,s} > 1 and

vol⁡n(B^q,pn∩B^r,sn)=exp⁡{−nJ(1)+o(n)},J(1)>0.\operatorname{vol}_n\left(\widehat{B}_{q,p}^n \cap\widehat{B}_{r,s}^n\right)=\exp\{-nJ(1)+o(n)\},\qquad J(1)>0.

Acknowledgement

The authors were supported by the German Research Foundation under DFG project 516672205 Limit theorems for the volume of random projections of ℓp\ell_p-balls. A considerable part of this work was carried out during a research semester of JP at Graz University of Technology. JP gratefully acknowledges the excellent working conditions and the hospitality provided during his stay. This work is intended to form part of SM’s doctoral dissertation.

Declaration of AI use

During the preparation of this work, the authors used Claude (Anthropic) and ChatGPT (OpenAI) as tools for literature search as well as technical and editorial assistance. It was also used to produce Figure 1. The mathematical framework and proof strategy were fully developed by the authors (and partly rely on [12]).

References

  1. [1]D. Alonso-Gutiérrez, J. Prochno, and C. Thäle. Large deviations for high-dimensional random projections of ℓ_p^n-balls. Advances in Applied Mathematics, 99:1–35, 2018.
  2. [2]F. Barthe and P. Wolff. Volume properties of high-dimensional Orlicz balls. In High dimensional probability IX—the ethereal volume, volume 80 of Progr. Probab., pages 75–95. Birkhäuser/Springer, Cham, 2023.arxiv.org/abs/2106.01675
  3. [3]A. Dembo and O. Zeitouni. Large Deviations Techniques and Applications, volume 38 of Applications of Mathematics. Springer, New York, second edition, 1998.
  4. [4]R. S. Ellis. Entropy, Large Deviations, and Statistical Mechanics. Classics in Mathematics. Springer, Berlin, 2006. Reprint of the 1985 edition.
  5. [5]D. H. Fremlin. Measure Theory. Vol. 2: Broad Foundations. Torres Fremlin, Colchester, second edition, 2010.
  6. [6]L. Frühwirth and J. Prochno. Sanov-type large deviations and conditional limit theorems for high-dimensional Orlicz balls. Journal of Mathematical Analysis and Applications, 536(1):128169, 2024.arxiv.org/abs/2111.04691
  7. [7]L. Frühwirth and J. Prochno. Sharp concentration phenomena in high-dimensional Orlicz balls. J. Funct. Anal., 290(6):Paper No. 111322, 45, 2026.arxiv.org/abs/2407.15579
  8. [8]N. Gantert, S. S. Kim, and K. Ramanan. Large deviations for random projections of ℓ^p balls. The Annals of Probability, 45(6B):4419–4476, 2017.arxiv.org/abs/1512.04988
  9. [9]M. Juhos, Z. Kabluchko, and J. Prochno. Limit theorems for mixed-norm sequence spaces with applications to volume distribution. Electron. J. Probab., 29:Paper No. 94, 44, 2024.DOI
  10. [10]M. Juhos and J. Prochno. Spectral flatness and the volume of intersections of p-ellipsoids. J. Complexity, 70:Paper No. 101617, 21, 2022.DOI
  11. [11]Z. Kabluchko and J. Prochno. The maximum entropy principle and volumetric properties of Orlicz balls. Journal of Mathematical Analysis and Applications, 495(1):124687, 2021.arxiv.org/abs/2007.05247
  12. [12]Z. Kabluchko, J. Prochno, and M. Sonnleitner. A probabilistic approach to Lorentz balls ℓ^n_{q,1}. Journal of Functional Analysis, 288(1):110682, 2025.
  13. [13]Z. Kabluchko, J. Prochno, and C. Thäle. High-dimensional limit theorems for random vectors in ℓ_p^n-balls. Communications in Contemporary Mathematics, 21(1):1750092, 2019.
  14. [14]Z. Kabluchko, J. Prochno, and C. Thäle. Intersection of unit balls in classical matrix ensembles. Israel J. Math., 239(1):129–172, 2020.arxiv.org/abs/1804.03466
  15. [15]Z. Kabluchko, J. Prochno, and C. Thäle. Sanov-type large deviations in Schatten classes. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 56(2):928–953, 2020.arxiv.org/abs/1808.04862
  16. [16]Z. Kabluchko, J. Prochno, and C. Thäle. High-dimensional limit theorems for random vectors in ℓ_p^n-balls. II. Communications in Contemporary Mathematics, 23(3):1950073, 2021.
  17. [17]S. S. Kim and K. Ramanan. A conditional limit theorem for high-dimensional ℓ^p-spheres. Journal of Applied Probability, 55(4):1060–1077, 2018.
  18. [18]J. Prochno. The large and moderate deviations approach in geometric functional analysis. In R. Adamczak, N. Gozlan, K. Lounici, M. Madiman, F. Merlevède, and E. Werner, editors, High Dimensional Probability X: The Będlewo Volume, volume 82 of Progress in Probability, pages 347–421. Birkhäuser, Cham, 2026.arxiv.org/abs/2403.03940
  19. [19]S. T. Rachev and L. Rüschendorf. Approximate independence of distributions on spheres and their stability properties. The Annals of Probability, 19(3):1311–1337, 1991.DOI
  20. [20]G. Schechtman and M. Schmuckenschläger. Another remark on the volume of the intersection of two L^n_p balls. In J. Lindenstrauss and V. D. Milman, editors, Geometric Aspects of Functional Analysis: Israel Seminar (GAFA) 1989–90, volume 1469 of Lecture Notes in Mathematics, pages 174–178. Springer, Berlin, Heidelberg, 1991.DOI
  21. [21]G. Schechtman and J. Zinn. On the volume of the intersection of two L^n_p balls. Proceedings of the American Mathematical Society, 110(1):217–224, 1990.

Paper details

vlnc:W_jwj37d2f
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

Claude, ChatGPT

View statement
Looking for related papers…
Contents