Asymptotics of Constrained Quantization for Compactly Supported Measures

Chenxing Qian

arXiv: 2505.10801v1 · PDF · 16 May 2025
Categories: math.MG, math.CA, math.FA, math.OC, math.PR
Licence: CC BY 4.0 The site shows the text.

Abstract

We investigated the asymptotics of high-rate constrained quantization errors for a compactly supported probability measure P on Euclidean spaces whose quantizers are confined to a closed set S. The key tool is the metric projection of K onto S that assigns each source point to its nearest neighbor in S, allowing the errors to be transferred to the projection, where K = supp P. For the upper estimate, we establish a projection pull-back inequality that bounds the errors by the classical covering radius of the projection. For the lower estimate, a weighted distance function enables us to perturb any quantizer element lying on the projection slightly into the complement in S without enlarging the error, provided the projection is nowhere dense (automatically true when S and K are disjoint). Under mild conditions on the pushforward measure of P by T, obtained via a measurable selector T, we derive a uniform lower bound. If this set is Ahlfors regular of dimension d, the error decays like the reciprocal of the d-th root of n and every constrained quantization dimension equals d. The two estimates coincide, giving the first complete dimension comparison formula for constrained quantization and closing the gap left by earlier self-similar examples by Pandey-Roychowdhury while extending classical unconstrained theory to closed constraints under mild geometric assumptions.

Items 70

Theorem

Theorem 1 (Upper Asymptotics)

4BKC

Suppose either the condition (?) or the condition (?) hold. Then,

let T be a measurable selector constructed in Lemma 11, for r\geq 1,
\begin{aligned} &\underline{\widehat{\dim}}_Q^{r}{(P;S)}\leq\underline{\dim}_B^{*}{(T_*P)}\leq\underline{\dim}_B^{}{(\pi_S(K))},\,\,\overline{\widehat{\dim}}_Q^{r}{(P;S)}\leq\overline{\dim}_B^{*}{(T_*P)}\leq\overline{\dim}_B^{}{(\pi_S(K))};\\ &\underline{\widetilde{\dim}}_Q^{r}{(P;S)}\leq\underline{\dim}_B^{*}{(T_*P)}\leq\underline{\dim}_B^{}{(\pi_S(K))},\,\,\overline{\widetilde{\dim}}_Q^{r}{(P;S)}\leq\overline{\dim}_B^{*}{(T_*P)}\leq\overline{\dim}_B^{}{(\pi_S(K))}; \end{aligned}
for r>0, if \pi_S(K) is Ahlfors regular of dimension d, then,
\begin{aligned} \dim_H(\pi_S(K))&\geq \overline{\widetilde{\dim}}_Q^{r}{(P;S)}\geq \underline{\widetilde{\dim}}_Q^{r}{(P;S)},\\\dim_H(\pi_S(K))&\geq \overline{\widetilde{\dim}}_Q^{\infty}{(P;S)}\geq \underline{\widetilde{\dim}}_Q^{\infty}{(P;S)},\\ \dim_H(\pi_S(K))&\geq\overline{\widehat{\dim}}_Q^{1}{(P;S)}\geq\underline{\widehat{\dim}}_Q^{1}{(P;S)}. \end{aligned}

Theorem

Theorem 2 (Lower Asymptotics)

ZXCG

Suppose the condition (?) is true. Let r\geq 1. In addition, suppose the pushforward measure as constructed in Lemma 11 is upper Ahlfors regular of dimension d. Then,

\underline{\widehat{\dim}}_Q^{r}{(P;S)}\geq \dim_H^*({T_*P}) and \underline{\widetilde{\dim}}_Q^{r}{(P;S)}\ge \dim_H^*({T_*P});
if the condition (?) is true, then \dim_H(T(K))=\dim_H^*({T_*P}), \dim_H(T(K))\le \underline{\widehat{\dim}}_Q^{r}{(P;S)} and \dim_H(T(K))\le \underline{\widehat{\dim}}_Q^{r}{(P;S)}

Corollary

Corollary 3

L1PK

Suppose conditions (?) and (?) are true. Suppose T constructed in Lemma 11 is surjective and T_*P is Ahlfors regular of dimension d. Let r\geq 1. Then we have the sharp dimension bounds,

\dim_H(\pi_S(K))=\underline{\widetilde{\dim}}_Q^{r}{(P;S)}=\underline{\widetilde{\dim}}_Q^{\infty}{(P;S)}=\underline{\widehat{\dim}}_Q^{1}{(P;S)}=\overline{\widetilde{\dim}}_Q^{r}{(P;S)}=\overline{\widetilde{\dim}}_Q^{\infty}{(P;S)}=\overline{\widehat{\dim}}_Q^{1}{(P;S)}.

Proposition

Proposition 4

I8VT

In [graf], if P is a compactly supported probability on K and 1\leq r\leq s\leq\infty, then \underline{\dim}_B^{}{(K)}=\underline{\dim}_Q^{\infty}{(P)}=\underline{\dim}_Q^{\infty}{(K)}\geq \underline{\dim}_Q^{s}{(P)}\geq\underline{\dim}_Q^{r}{(P)} and \overline{\dim}_B^{}{(K)}=\overline{\dim}_Q^{\infty}{(P)}=\overline{\dim}_Q^{\infty}{(K)}\geq \overline{\dim}_Q^{s}{(P)}\geq\overline{\dim}_Q^{r}{(P)};
In [graf], let P be a probability, for r\geq 1, \dim_H^*({P})\leq \underline{\dim}_Q^{r}{(P)};
In [dai2008quantization], if P is a probability on \bigtimes\limits_{i=1}^D [0,1] and 0<r<\infty, then \dim_P^*({P})\leq \overline{\dim}_Q^{^r}{(P)} and \dim_H^*({P})\leq \underline{\dim}_Q^{r}{(P)};
In [kloeckner2012], if P Ahlfors regular of dimension d and compactly supported on a stable Riemannian manifold, let \infty>r\geq 1, then 0<\liminf\limits_{n\rightarrow\infty}ne_{n,r}^{d}(P)\leq \limsup\limits_{n\rightarrow\infty}ne_{n,r}^{d}(P)<\infty, that is, \underline{\dim}_Q^{r}{(P)}=\overline{\dim}_Q^{r}{(P)}=d

Used by Proof

Remark

Remark 1

R3J7

In [dai2008quantization], the author established the proof of \dim_H^*({P})\leq \underline{\dim}_Q^{r}{(P)} when P is supported in \bigtimes\limits_{i=1}^D [0,1]. This proof can be further extended to scenarios where the quantizers are constrained within the support , say \dim_H^*({P})\leq \underline{\dim}_Q^{r}{(P;S)}, provided that P is compactly supported. The proof can also be generalized to the cases where the quantizers are constrained in the support K=\text{\normalfont supp} {(P)}, for P compactly supported. In addition [graf, Section 11.3], the author proposed the proof for \underline{\dim}_B(K)=\underline{\dim}_Q^{\infty}{(P)} where \text{\normalfont supp} {(P)}=K\Subset\mathbb R^D under the assumption of unconstrained quantization. This result can also be generalized to the case of constrained quantization when the constraint set itself is K, as the proof does not rely on the specific geometry of the quantizers.

Used by Proof

Lemma

Lemma 5 (Existence of C_{n,r}(P;S))

XJHI

Let 1\leq r<\infty. Suppose P is a Borel probability on \mathbb R^D, and S\subseteq \mathbb R^D is a closed set. Then

C_{1,r}(P;S) is nonempty;
If there exists at least n elements in S whose Voronoi regions have positive probability, then C_{n,r}(P;S) is nonempty and |\alpha|=n for any \alpha\in C_{n,r}(P;S);
C_{n,\infty}(P;S) is nonempty if \text{\normalfont supp} {(P)}\Subset \mathbb R^D is compact.

Proof of Lemma 5

The first argument was proven in [pandey2023, Proposition 1.2] with a setting of 0<r<\infty. The second argument was proven in [pandey2023, Proposition 1.3] based on the first argument. There was also a proof of the second argument on the unconstrained case in [graf, Section 4.1]. The third argument under the unconstrained case was proven in [graf, Section 10.3], which can be genuinely generalized by Lemma 6.

Uses Lemma 6

Lemma

Lemma 6

8CUX

Let n\in\mathbb N, S\subseteq\text{\normalfont supp} {(P)} \subseteq\mathbb R^D. Suppose 1\leq r\leq s\leq\infty, then

e_{n,r}(P;S)\leq e_{n,s}(P;S)

Furthermore, if S\Subset \mathbb R^D then

\lim\limits_{r\rightarrow \infty}e_{n,r}(P;S)=e_{n,\infty}(P;S)

Used by Proof

Proof of Lemma 6

For the first argument, using the Jensen’s inequality:

\begin{aligned} e_{n,r}(P;S)&=\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}\bigg(\int 0.9355(\alpha,x)^rdP(x)\bigg)^{\frac{1}{r}}=\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}\bigg(\int 0.9355(\alpha,x)^{s\cdot\frac{r}{s}}dP(x)\bigg)^{\frac{1}{s}\cdot\frac{s}{r}}\\ &\leq \inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}\bigg(\int 0.9355(\alpha,x)^sdP(x)\bigg)^{\frac{1}{s}}=e_{n,s}(P;S) \end{aligned}

If \text{\normalfont supp} {(P)} is compact, then \sup\limits_{x\in\text{\normalfont supp} {(P)}}0.9355(x,0)<\infty and therefore \lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)\leq e_{n,\infty}(P;S)<\infty. For the second argument, as the mapping (a_1,a_2,\cdots,a_n)\mapsto \{a_1,a_2,\cdots,a_n\} is d_H continuous for a_i\in \mathbb R^D, \{\alpha\subseteq \mathbb R^D:1\leq|\alpha|\leq n,\alpha\in S\} is thus d_H compact. Consider \alpha_r to be the n-optimal set for e_{n,r}(P;S), there exists a subsequence \{\alpha_{r_k}\}_k with the d_H cluster point \alpha\subseteq S, then:

\bigg({\int (0.9355(\alpha_{r_k},x)-0.9355(\alpha,x))^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\leq \sup\limits_{x\in\text{\normalfont supp} {(P)}}|0.9355(\alpha_{r_k},x)-0.9355(\alpha,x)|=d_H(\alpha_{r_k},\alpha)

with \alpha_{r_k}\subseteq S and \alpha\subseteq S. In addition:

\begin{aligned} \bigg({\int 0.9355(\alpha_{r_k},x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}} &\geq \bigg({\int 0.9355(\alpha,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}-\bigg({\int \bigg|{0.9355(\alpha_{r_k},x)-0.9355(\alpha,x)}\bigg|^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\\ &\geq \bigg({\int 0.9355(\alpha,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}-d_H(\alpha_{r_k},S) \end{aligned}

from above we can observe

\begin{aligned} \lim\limits_{k\rightarrow\infty}e_{n,r_k}(P;S)&=\lim\limits_{k\rightarrow\infty}\bigg({\int 0.9355(\alpha_{r_k},x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\geq \lim\limits_{k\rightarrow\infty}\bigg({\int 0.9355(\alpha,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}=\sup\limits_{x\in\text{\normalfont supp} {(P)}}0.9355(x,\alpha)\\ &\geq \inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}\sup\limits_{x\in\text{\normalfont supp} {(P)}}0.9355(x,\alpha)=e_{n,\infty}(P;S) \end{aligned}

Lemma

Lemma 7

TBI9

Suppose K\Subset \mathbb R^D is compact, then C_{n,\infty}(K) is nonempty, plus

e_{n,\infty}(K)=\inf\limits_{\substack{\alpha\subseteq \mathbb R^D\\ |\alpha|\leq n}}d_H(\alpha,K)

In addition, suppose S\subseteq \mathbb R^D is closed (and therefore \pi_S(K) is compact) then C_{n,\infty}(K;S) is nonempty with

e_{n,\infty}(K;S)=\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}d_H(\alpha,K)

Used by Proof

Proof of Lemma 7

The first argument is due to [graf, Section 10.3 and 10.4]. The second argument will be a genuine generalization of the proof in [graf, Section 10.4] to the constrained case, based on the nonemptiness results in Lemma 16.

Uses Lemma 16

Lemma

Lemma 8

ASJI

Suppose \pi_S^{-1}(E) is defined in (?) for E\in\pi_S(K). \pi_S^{-1}(E) is closed if E is closed.

Proof of Lemma 8

For F\subseteq \pi_S(K) a closed set, let \{s_n\}_n be a converging sequence in \pi_S^{-1}(F) with cluster point \overline{s}. Assuming \overline{s}\not\in \pi_S^{-1}(F), which means \pi_S(\overline{s})\cap F=\varnothing, by the continuity of the function f(x)=\inf\limits_{z\in S}0.9355(x,z), we have \lim\limits_{n\rightarrow\infty}f(s_n)=f(\overline{s})=0.9355(\overline{s},\pi_S(K)). On the other hand \lim\limits_{n\rightarrow\infty}f(s_n)=\lim\limits_{n\rightarrow\infty}\inf\limits_{z\in S}0.9355(s_n,z)=0.9355(\overline{s},F). Therefore 0.9355(\overline{s},F)=0.9355(\overline{s},\pi_S(K)), and thus \overline{s}\in \pi_S^{-1}(F), contradicting \overline{s}\not\in \pi_S^{-1}(F).

Remark

Remark 2

82P4

For a multivalued function f:X\rightarrow 2^Y, where f(x) is a set for x\in X, we say that f is upper semi-continuous if for any closed set G\subseteq S, the set \{y\in X: f(y)\subseteq G, f(y)\neq\varnothing\} is closed, by Kuratowski’s definition in [kuratowski2014topology], and therefore \pi_S:K\rightarrow 2^S is upper semi-continuous.

Lemma

Lemma 9

C2N1

Suppose S is a closed set and K=\text{\normalfont supp} {(P)} is a compact set for a probability measure P. \pi_S(K) and \pi_S^{-1}(E) is defined in (?) and (?) for E\subseteq \pi_S(K). Then \pi_S(K) is compact.

Used by Proof

Proof of Lemma 9

As the function f(x)=\inf\limits_{z\in S}0.9355(x,z) is continuous and uniformly continuous on K, let g=f|_K, g(K) is compact \mathbb R^{+}\cup\{0\} and g(K)=\bigcup\limits_{i=1}^N F_i where F_i is closed. Denote G_i=g^{-1}(F_i) and we have \bigcup\limits_{i=1}^N G_i=K, as g^{-1}(g(K))=K by the fact that g=f|_K. \pi_S(G_i) is bounded and closed, therefore \pi_S(K)=\bigcup\limits_{i=1}^N \pi_S(G_i) is compact.

Lemma

Lemma 10 (Closed graph property of \pi_S(K))

9O04

For S\in\mathbb R^D a closed set, and K\Subset\mathbb R^D a compact set. The graph of the projection \pi_S, denoted,

\mathrm{{Gr } }\pi_S:\{(x,y)\in K\times S\}:0.9355(x,y)=0.9355(x,S)\},

is compact.

Used by Proof

Proof of Lemma 10

Let f(x,y)=0.9355(x,y)-0.9355(x,S), then f:K\times S\rightarrow\mathbb R is continuous. We can rewrite \text{Gr}\pi_S as \text{Gr}\pi_S=f^{-1}(\{0\})\cap (K\times S), therefore \text{Gr}\pi_S is closed. For (x,y)\in\text{Gr}\pi_S, we can show that 0.9355(y,0) is bounded, and therefore \text{Gr}\pi_S is compact as a closed subset of the product of two compact sets.

Lemma

Lemma 11 (Meaurable selector)

4QWO

For a compact set K, if K=\text{\normalfont supp} {(P)} for some Borel measure P, there exists a single-valued map T:K\rightarrow \pi_S(K), where T(x)\in \pi_S(x), such that T is measurable with respect to P.

Used by Proof, Proof

Proof of Lemma 11

In Lemma 10, we showed \mathrm{Gr}\pi_S is compact. In addition, \pi_S(x) is closed. Then we can use the Kuratwoski-Ryll-Nardzewski theorem in [jureczko2023remarks] and [srivastava2008course, Page 189] to guarantee the existence of a measuable selector.

Uses Lemma 10

Lemma

Lemma 12 (Nowhere density of \pi_S(K))

8YZT

Suppose S\cap K=\varnothing, S is closed in \mathbb R^D and K is compact in \mathbb R^D, then \pi_S(K) is nowhere dense.

Used by Proof

Proof of Lemma 12

Suppose \pi_S(K) is not nowhere dense, then (\overline{\pi_S(K)})^\circ=({\pi_S(K)})^\circ\neq\varnothing, that is, there exists y\in\pi_S(K) and some r_0>0 such that B(y,r_0)\subseteq \pi_S(K). Let x\in\pi_S^{-1}(y), and select z\in B(y,r_0) such that 0.9355(y,z)=r_0/2. We claim that 0.9355(x,z)\geq 0.9355(x,y), because otherwise x\not\in\pi_S^{-1}(y). Thereafter, 0.9355(x,z)\geq 0.9355(x,y)\geq 0.9355(x,z)+0.9355(y,z)=0.9355(x,z)+r_0/2, there is then a contradiction.

Lemma

Lemma 13 (Stability)

IX6F

Let \{x_n\}_n\subseteq K be a sequence converging to \overline{x}, then \lim\limits_{n\rightarrow\infty}0.9355(\pi_S(\overline{x}),\pi_S(x_n))=0.

Used by Proof

Proof of Lemma 13

Assume the statement is false; then there exist some a>0 and some subsequences \{x_{n_l}\}_l such that 0.9355(\pi_S(\overline{x}),\pi_S(x_{n_l}))>a. Denote \{x_{n_l}\}_l as \{x_m\}_m. For each m\in\mathbb N^{+}, if 0.9355(\overline{x},\pi_S(\overline{x}))=0.9355(\overline{x},\pi_S(x_n)), then \pi_S(\overline{x})\cap \pi_S(x_n)\neq \varnothing, contradicting 0.9355(\pi_S(\overline{x}),\pi_S(x_{m}))>a. Therefore 0.9355(\overline{x},\pi_S(\overline{x}))<0.9355(\overline{x},\pi_S(x_n)). By the continuity of the function f(x)=\inf\limits_{z\in S}0.9355(x,z), we know that \lim\limits_{m\rightarrow\infty}0.9355(x_m,\pi_S(x_m))=\lim\limits_{m\rightarrow\infty}\inf\limits_{y\in S}0.9355(x_m,y)=\inf\limits_{y\in S}0.9355(\overline{x},y)=0.9355(\overline{x},\pi_S(\overline{x})), and moreover \lim\limits_{m\rightarrow\infty}0.9355(\overline{x},\pi_S(x_m))=0.9355(\overline{x},\pi_S(\overline{x})) by the fact that

\begin{aligned} \lim\limits_{m\rightarrow\infty}0.9355(\overline{x},\pi_S(x_m))&=\lim\limits_{m\rightarrow\infty}\inf\limits_{y\in\pi_S(x_m)}0.9355(\overline{x},y)\leq \lim\limits_{m\rightarrow\infty}\inf\limits_{y\in\pi_S(x_m)}\bigg({0.9355(\overline{x},x_m)+0.9355(x_m,y)}\bigg)\\ &=\lim\limits_{m\rightarrow\infty}0.9355(x_m,\pi_S(x_m))=\lim\limits_{m\rightarrow\infty}\inf\limits_{y\in S}0.9355(x_m,y)\\ &\leq \lim\limits_{m\rightarrow\infty}\inf\limits_{y\in S}\bigg({0.9355(\overline{x},y)+0.9355(\overline{x},x_m)}\bigg)=\lim\limits_{m\rightarrow\infty}0.9355(\overline{x},\pi_S(x_m)) \end{aligned}

Assume there exists \{y_k\}_k\subseteq\bigcup\limits_{n=1}^\infty \pi_S(x_n) such that \lim\limits_{k\rightarrow\infty} 0.9355(\overline{x},y_k)=0.9355(\overline{x},\pi_S(\overline{x})), by the compactness of \pi_S(K) from Lemma 9, there exists some subsequence \{y_{k_l}\}_l\subseteq \{y_k\}_k converging to some \overline{y}\in \pi_S(\overline{x}), contradicting 0.9355(\pi_S(\overline{x}),\pi_S(x_{m}))>a. Therefore, there exists some b>0 such that 0.9355(\overline{x},\pi_S(\overline{x}))+b<0.9355(\overline{x},y) for y\in \bigg(\bigcup\limits_{z\in \pi_S(\overline{x})}\{t:t=z+s,s\in B(0,a/2)\}\bigg)^c, and therefore \lim\limits_{m\rightarrow\infty}0.9355(\overline{x},\pi_S(x_m))>b+0.9355(\overline{x},\pi_S(\overline{x})), contradicting the previous established equality that \lim\limits_{m\rightarrow\infty}0.9355(\overline{x},\pi_S(x_m))=0.9355(\overline{x},\pi_S(\overline{x})).

Uses Lemma 9

Remark

Remark 3

VRW8

If for any x\in K, for any sequence \{x_n\}_n converging to x, there exists p\in\pi_S(x) such that \lim\limits_{n\rightarrow\infty}0.9355(p,\pi_S(x_n))=0, then there exist a measurable selector T:K\rightarrow \pi_S(K) which is not only measurable but also a finite combination of continuous functions supported on Borel sets.

Proof (Proof of the remark.) of Remark 3

First, we claim that for \overline{x}\in K, there exists G(\overline{x}) a closed set containing \overline{x}, and there exists a continuous function f:G(\overline{x})\rightarrow S such that f(x)\in \pi_S(x) for x\in G(\overline{x}). To prove the first claim, it suffices to show that f is continuous on \overline{x}, which is guaranteed by the Lemma 13, by the compactness of G(\overline{x}). For each \overline{x}, we can generate such a continuous function at \overline{x} which is also continuous on a closed neighborhood G(\overline{x}). Therefore, \bigcup\limits_{\overline{x}\in K}G(\overline{x}) forms a cover of K and there exists a finite subcover \bigcup\limits_{i=1}^k G(x_i), with corresponding continuous function f_i on x_i and G(x_i). By the axiom of choice, there exists a map T:K\rightarrow S with T(x)\in \pi_S(x) such that T(x)=f_i(x) for some 1\leq i\leq k. We would like to show that T^{-1}(U) is measurable where U is open in the sense of subspace topology on S. Let M_i be a integer only dependent to i, then we have

\begin{aligned} T^{-1}(U)=\bigcup\limits_{i=1}^k \bigg({G(x_i)\cap T^{-1}(U)}\bigg)=\bigcup\limits_{i=1}^k\bigcup\limits_{l=1}^{M_i}f_i^{-1}(E_l) \end{aligned}

where \{E_l\}_{l\leq M_i} are Borel sets and are generated by finite set operations by how we define T by f_i previously. We can argue from above that T^{-1}(U) is Borel, and therefore measurable.

Uses Lemma 13

Remark

Remark 4 (Convex constraint)

BH6J

It is a well-known result in [boyd2004convex, Chapter 8] that if S is convex, then \pi_S is not only single-valued, but also 1-Lipschitz, following the discussion in Remark 5, we can conclude \dim_H(T(K))\leq \dim_H(K).

Remark

Remark 5

HXKR

If the map T:K\rightarrow\pi_S(K) is bi-Lipschitz, say there exists C'>0 such that \dfrac{1}{C'}0.9355(x,y)\leq 0.9355(T(x),T(y))\leq C'0.9355(x,y) for x,y\in K, then \dim_H(K)=\dim_H(T(K)).

Proof (Proof of the remark.) of Remark 5

Let t=\dim_H(T(K)). First, we claim that if 0.9355(T(x),T(y))\leq C'0.9355(x,y), then \dim_H(T(K))\leq\dim_H(K). To show the claim, consider K\subset\bigcup\limits_{i=1}^\infty U_i where \text{\normalfont diam}(U_i)<\delta, then \{T(U_i)\}_{i=1}^\infty is a cover of T(K) and

\text{\normalfont diam}(T(U_i))=\sup\limits_{x,y\in T(U_i)}0.9355(x,y)=\sup\limits_{x,y\in U_i}0.9355(T(x),T(y))\leq C'\text{\normalfont diam}(U_i)<C'\delta

which implies that

\begin{aligned} \mathcal{H}_\delta^t(K)&=\inf\limits_{\substack{K\subseteq\bigcup\limits_{i=1}^\infty U_i\\\text{diam}(U _i)<\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(U_i)^t\geq \dfrac{1}{C'}\inf\limits_{\substack{T(K)\subseteq\bigcup\limits_{i=1}^\infty T(U_i)\\ \text{\normalfont diam}(T(U _i))<C'\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(U_i)^t \\&\geq\dfrac{1}{C'}\inf\limits_{\substack{T(K)\subseteq\bigcup\limits_{i=1}^\infty U_i\\ \text{\normalfont diam}(U _i)<C'\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(U_i)^t =\mathcal{H}_{C'\delta}^t(K) \end{aligned}

and therefore \mathcal{H}^t(K)\geq\mathcal{H}^t(T(K)) which implies \dim_H(T(K))\leq\dim_H(K). On the other hand, if 0.9355(T(x),T(y))\geq \dfrac{1}{C'}0.9355(x,y) then

\delta>\text{\normalfont diam}(U_i)=\sup\limits_{x,y\in U_i}0.9355(x,y)=\sup\limits_{T(x),T(y)}0.9355(T(x),T(y))\geq \dfrac{1}{C'}\sup\limits_{T(x),T(y)\in U_i}0.9355(x,y)=\dfrac{1}{C'}\text{\normalfont diam}(T^{-1}(U_i))

and if we consider \{U_i\}_{i=1}^\infty covers T(K) with \text{\normalfont diam}(U_i)<\delta then T(K)\subseteq \bigcup\limits_{i=1}^\infty U_i and K=T^{-1}(T(K))=\bigcup\limits_{i=1}^\infty T^{-1}(U _i)

\begin{aligned} \mathcal{H}_\delta^t(K)&=\inf\limits_{\substack{K\subseteq \bigcup\limits_{i=1}^\infty U_i\\ \text{\normalfont diam}(U_i)<\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(U_i)^t\leq\inf\limits_{\substack{K\subseteq \bigcup\limits_{i=1}^\infty T^{-1}(U_i)\\ \text{\normalfont diam}(T^{-1}(U_i))<C'\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(T^{-1}(U_i))^t\leq \\ & \leq \inf\limits_{\substack{T(K)\subset\bigcup\limits_{i=1}^\infty U_i\\\text{\normalfont diam}(U_i)<C'\delta}}\sum\limits_{i=1}^\infty\text{\normalfont diam}(U_i)^t=\mathcal{H}_{C'\delta}^t(T(K)) \end{aligned}

and therefore \mathcal{H}^t(K)\leq\mathcal{H}^t(T(K)) which implies \dim_H(K)\leq\dim_H(T(K))

Lemma

Lemma 14

DIAC

(The support of the push-forward measure T_*P) For P a probability measure compactly supported on \text{\normalfont supp} {(P)}=K, with the mapping T:K\rightarrow \pi_S(K) constructed in Lemma 11, and assume conditions (?) or (?) are true, then for the push-forward measure T_*P we have \text{\normalfont supp} {(T_*P)}=\overline{T(K)}.

Used by Proof, Proof

Proof of Lemma 14

It suffices to assume condition (?) to be true. For any x\in\pi_S(K), and for any \varepsilon>0, by the definition in (?) we have

\begin{aligned} T^{-1}(B(x,\varepsilon)\cap S)&=\{y\in K:T(y)\in B(x,\varepsilon)\cap S\}\\ &\supset \{y\in K:\pi_S(y)\subseteq B(x,\varepsilon)\cap S\}=\pi_S^{-1}(B(x,\varepsilon)\cap S) \end{aligned}

and therefore

T_*P(B(x,\varepsilon)\cap S)=P(T^{-1}(B(x,\varepsilon)\cap S))\geq P(\pi_S^{-1}(B(x,\varepsilon)\cap S))>0

which implies that \overline{T(K)}\subseteq \text{\normalfont supp} {(T_*P)}, and on the other hand, \text{\normalfont supp} {(T_*P)}\subseteq \overline{T(K)} as T(K)\subseteq\overline{T(K)}.

Lemma

Lemma 15 (Lower Ahlfors regularity of the measure T_*P)

KWDL

For S\subseteq\mathbb R^D a closed set, P a probability measure supported on K which is a compact set, suppose the condition (?) is true, then there exists C>0 such that for any x\in \overline{T(K)} and for any r>0 we have T_*P(B(x,r)\cap S)\geq Cr^s where s=\dim_H(\overline{T(K)}).

Used by Proof

Proof of Lemma 15

In the proof of Lemma 14 we established that T_*P(B(x,\varepsilon))\geq P(\pi_S^{-1}(B(x,\varepsilon)\cap S)), then trivially the lemma is true by the condition (?).

Uses Lemma 14

Lemma

Lemma 16 (Nonemptiness of C_{n,r}(P;S))

TDRF

For 0<r<\infty, if \int 0.9355(x,0)^r dP(x)<\infty, then C_{n,r}(P;S)\neq\varnothing. Moreover, if P is compactly supported in \mathbb R^D, C_{n,\infty}(P;S)\neq\varnothing.

Used by Proof, Proof

Proof of Lemma 16

The following proof can proceed in a similar way of [pandey2023]. Construct a function

\psi_r:\bigtimes\limits_{i=1}^n S\rightarrow\mathbb R^+\cup \{0\},\quad (s_1,\cdots,s_n)\mapsto\int 0.9355(x,(s_1,\cdots,s_n))^r dP(x),\quad s_i\in S,\quad i=1,\cdots,n

We first show that the level set

\{\psi_r\leq c\}:=\{(s_1,\cdots,s_n)\subseteq\bigtimes\limits_{i=1}^n S:\int 0.9355(x,(s_1,\cdots,s_n))^rdP(x)\leq c\}

is compact for c>\inf\limits_{\beta\subseteq S,|\beta|= n}\int 0.9355(x,\beta)dP(x). If we construct a decreasing sequence \{c_n\}_n such that

c_m>\inf\limits_{\beta\subseteq S,|\beta|=n}\int 0.9355(x,\beta)dP(x)\quad\text{ and }\quad \lim\limits_{m\rightarrow\infty}c_m=\inf\limits_{\beta\subseteq S,|\beta|= n}\int 0.9355(x,\beta)dP(x)

We can show that \bigcap\limits_{m=1}^\infty \{\psi_r\leq c_m\}\neq\varnothing. If (a_1,\cdots,a_n)\in \bigcap\limits_{m=1}^\infty \{\psi_r\leq c_m\}\neq\varnothing, we have \inf\limits_{\beta\subseteq S,|\beta|\leq n}\int 0.9355(x,\beta)dP(x)\leq \int 0.9355(x,(a_1,\cdots,a_n))^rdP(x)<c_m for any m\in\mathbb N^+ and therefore

\inf\limits_{\beta\subseteq S,|\beta|= n}\int 0.9355(x,\beta)dP(x)=\int 0.9355(x,(a_1,\cdots,a_n))^rdP(x).

As in [pandey2023], the authors established the nonemptiness of C_{1,r}(P;S), we can argue the nonemptiness of C_{n,r}(P;S) by an inductive argument. The proof of the second argument can proceed in a similar way in [graf]. As \pi_S(K) is compact, provided the compactness of K, the set \{\alpha\subseteq \pi_S(K):1\leq |\alpha|\leq n\} is d_H compact by the d_H continuity of the map (a_1,\cdots,a_n)\mapsto \{a_1,\cdots,a_n\}. Provided the nonemptiness of C_{n,r}(P;S), let \alpha_r\in C_{n,r}(P;S) and \{r_k\}_k a sequence with \lim\limits_{k\rightarrow\infty}r_k=\infty. We claim that the d_H cluster points of \{\alpha_{r_k}\}_k is a nonempty set of C_{n,\infty}(K;S), where K=\text{\normalfont supp} {(P)}. Let \alpha be one d_H cluster point of \{\alpha_{r_k}\}_k, as \alpha\subseteq \pi_S(K) by the compactness of \pi_S(K), we have:

\begin{aligned} \bigg({\int \bigg({0.9355(x,\alpha_{r_k})-0.9355(\alpha,x)}\bigg)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\leq \sup\limits_{x\in K}|0.9355(\alpha_{r_k},x)-0.9355(\alpha,x)|=d_H(\alpha_{r_k},\alpha) \end{aligned}

and therefore

\begin{aligned} \bigg({\int 0.9355(\alpha_{r_k},x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}&\geq \bigg({\int 0.9355(\alpha,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}-\bigg({\int \bigg({0.9355(\alpha_{r_k},x)-0.9355(\alpha,x)}\bigg)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\\ &\geq \bigg({\int 0.9355(\alpha,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}-d_H(\alpha_{r_k},\alpha) \end{aligned}

As by Lemma 18, \lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)=e_{n,\infty}(P;S), we have the following estimate:

\begin{aligned} e_{n,\infty}(P;S)&\geq \lim\limits_{k\rightarrow\infty}e_{n,r_k}(P;S)=\lim\limits_{k\rightarrow\infty}\bigg({\int 0.9355(\alpha_{r_k},x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\\ &\geq \lim\limits_{k\rightarrow\infty}\bigg({\int 0.9355(\alpha,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}=\sup\limits_{x\in K}0.9355(\alpha,x)\geq e_{n,\infty}(P;S) \end{aligned}

therefore \alpha\in C_{n,\infty}(K;S)=C_{n,\infty}(P;S)

Uses Lemma 18

Lemma

Lemma 17

PX3C

Suppose P is a compactly supported Borel probability measure with supp(P)=K, S\subseteq \mathbb R^D is a closed set. Let \{\beta_n\}_n be the sequence that minimizes e_{n,r}(K;S). d_H(A,B) stands for the Hausdorff distance between A and B.

If S\subseteq \subseteq K\Subset \mathbb R^D, then \lim\limits_{n\rightarrow\infty}d_H(\beta_n,S)=0. More specifically, if \gamma_n\subseteq S satisfies |\gamma_n|\leq n and \lim\limits_{n\rightarrow\infty}\int 0.9355(x,\gamma_n) dP(x)=e_{\infty,r}(K;S), then \lim\limits_{n\rightarrow\infty}d_H(\gamma_n,S)=0
Similarly, if the condition (?) are satisfied, then the condition (?) is automatically true, i.e., \lim\limits_{n\rightarrow\infty}d_H(\beta_n,\pi_S(K))=0

Used by Proof, Proof, Proof, Proof

Proof of Lemma 17

We first prove the first argument. The existence of e_{n,r}(P;S) is justified by Lemma 16 for 1\leq r\leq\infty. First, consider the case 1\leq r<\infty. Suppose \{\beta_n\}_n is the sequence of minimizers of e_{n,r}(P;S). By contradiction, assume \lim\limits_{n\rightarrow\infty}d_H(\beta_n,S)=\lim\limits_{n\rightarrow\infty}\max\limits_{x\in S}\min\limits_{b\in\beta_n}0.9355(b,x) is not zero, then there exists some \varepsilon>0, y\in S and some subsequences \{\beta_{n_k}\}_k such that B(y,\varepsilon)\cap \beta_{n_k}=\varnothing. Let \{A_b:b\in \beta_{n_k}\} be the Voronoi partition of K with respect to \beta_{n_k}, then for any c>2,
\begin{aligned} e_{n_k,r}^r(P;S)&=\sum\limits_{b\in \beta_{n_k}}\int_{A_b}0.9355(x,b)^rdP(x)=\sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus B(y,\frac{\varepsilon}{c})}0.9355(x,b)^rdP(x)+\int_{B(y,\frac{\varepsilon}{c})}\min\limits_{b\in\beta_{n_k}}0.9355(b,x)^rdP(x)\\ &\geq \sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus B(y,\frac{\varepsilon}{c})}0.9355(x,b)^rdP(x)+\int_{B(y,\frac{\varepsilon}{c})}0.9355(y,x)^rdP(x)+\bigg(\frac{\varepsilon}{2}\bigg)^r\cdot P(B(y,\frac{\varepsilon}{c}))\\ &\geq \sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus B(y,\frac{\varepsilon}{2})}0.9355(x,b)^rdP(x)+\int_{B(y,\frac{\varepsilon}{2})}0.9355(y,x)^rdP(x)+\bigg(\frac{\varepsilon}{2}\bigg)^r\cdot P(B(y,\frac{\varepsilon}{c})) \end{aligned}
If we consider the Voronoi partition \{E_b:b\in \beta_{n_k}\cup\{y\}\} with respect to \beta_{n_k}\cup\{y\}, then B(y,\frac{\varepsilon}{2})\subseteq E_y and thus
\begin{aligned} &\sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus B(y,\frac{\varepsilon}{2})}0.9355(x,b)^rdP(x)+\int_{B(y,\frac{\varepsilon}{2})}0.9355(y,x)^rdP(x)\\ &\geq \sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus E_y}0.9355(x,b)^rdP(x)+\int_{E_y}0.9355(y,x)^rdP(x) \geq \sum\limits_{b\in\beta_{n_k}\cup\{y\}}\int_{E_b}0.9355(b,x)^rdP(x)\geq e_{n_{k}+1,r}^r(P;S) \end{aligned}
Therefore
\begin{aligned} \lim\limits_{k\rightarrow\infty}e_{n_k,r}^r(P;S)&\geq \lim\limits_{k\rightarrow\infty}e_{n_k+1,r}^r(P;S)+\bigg(\frac{\varepsilon}{2}\bigg)^r\cdot P(B(y,\frac{\varepsilon}{c}))\\ &=e_{\infty,r}^r(P;S)+\bigg(\frac{\varepsilon}{2}\bigg)^r\cdot P(B(y,\frac{\varepsilon}{c}))>e_{\infty,r}^r(P;S) \end{aligned}
The limit of e_{n,r}(P;S) is larger than e_{\infty,r}(P;S), there is an contradiction. For the case that r=\infty, for \{\gamma_n\}_n a sequence of minimizers of e_{n,\infty}(K;S), we assume that there exist some \delta>0 and z\in S and a subsequence \{\gamma_{n_k}\}_k such that B(z,\delta)\cap \gamma_{n_k}=\varnothing. Denote V_{n_k,r}:=\bigg({\int 0.9355(\gamma_{n_k},x)^rdP(x)}\bigg)^{\frac{1}{r}}, and we further observe from the previous proof that V_{n_k,r}\geq e_{n_k+1,r}^{r}(P;S)+(\frac{\varepsilon}{2})^{r}P(B(y,\frac{\varepsilon}{10})) that
\frac{V_{n_k,r}}{2^{r}}\geq \bigg({\frac{e_{n_k+1,r}^{r}(P;S)+(\frac{\varepsilon}{2})^{r}P(B(y,\frac{\varepsilon}{10}))}{2}}\bigg)^{\frac{1}{r}}\geq e_{n_k+1,r}(P;S)+(\frac{\varepsilon}{2})P(B(y,\frac{\varepsilon}{10}))^{\frac{1}{r}}
We can then take the limit r\rightarrow\infty and k\rightarrow\infty and then
0=\lim\limits_{\substack{ k\rightarrow\infty\\ r\rightarrow\infty}}\frac{V_{n_k,r}}{2^{r}}\geq\lim\limits_{\substack{k\rightarrow\infty}}\bigg({e_{n_k+1,r}(P;S)+(\frac{\varepsilon}{2})P(B(y,\frac{\varepsilon}{10}))^{\frac{1}{r}}}\bigg)=e_{\infty,\infty}(P;S)+\frac{\varepsilon}{2}>0
There is then a contradiction.
For the second argument, suppose \{\beta_n\}_n is the sequence of minimizers of e_{n,r}(P;S) and assume \lim\limits_{n\rightarrow\infty}d_H(\beta_n,\pi_S(K)) is not zero, then there exists some \varepsilon>0, y\in \pi_S(K) and some subsequences \{\beta_{n_k}\}_k such that B(y,\varepsilon)\cap \beta_{n_k}=\varnothing. In addition, let 1>c>0 and for z\in \pi_S^{-1}(B(y,c\varepsilon)) then 0.9355(z,B(y,c\varepsilon))\geq 0.9355(z,y)-c\varepsilon. For \zeta\in \pi_S^{-1}(B(y,\varepsilon)) we also have 0.9355(\zeta,\beta_{n_k})-2c\varepsilon>0.9355(\zeta,B(y,\varepsilon)), if we choose proper c. Therefore for z\in \pi_S^{-1}(B(y,c\varepsilon)) we have 0.9355(z,y)-c\varepsilon\leq 0.9355(z,B(y,c\varepsilon))=0.9355(z,B(y,\varepsilon))<0.9355(z,\beta_{n_k})-2c\varepsilon which gives 0.9355(z,y)+c\varepsilon<0.9355(z,\beta_{n_k}). Hence,
\begin{aligned} e_{n_k,r}^r(P;S)&=\sum\limits_{b\in \beta_{n_k}}\int_{A_b}0.9355(x,b)^rdP(x)\\ &=\sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus \pi_S^{-1}(B(y,c\varepsilon))}0.9355(x,b)^rdP(x)+\int_{\pi_S^{-1}(B(y,c\varepsilon))}\min\limits_{b\in\beta_{n_k}}0.9355(b,x)^rdP(x)\\ &\geq \sum\limits_{b\in \beta_{n_k}}\int_{A_b\setminus \pi_S^{-1}(B(y,c\varepsilon))}0.9355(x,b)^rdP(x)+\int_{\pi_S^{-1}(B(y,c\varepsilon))}0.9355(y,x)^rdP(x)+(c\varepsilon)^r\cdot P(\pi_S^{-1}(B(y,c\varepsilon))) \end{aligned}
Similarly to the previous part of the proof, we can have a contradiction for r<\infty and r=\infty respectively.

Uses Lemma 16

Lemma

Lemma 18 (Monotonicity and limit for e_{n,r}(P;S))

EX8U

Let n\in\mathbb N, and S is a closed constraint. Suppose 0< r\leq s\leq\infty and the existence of e_{n,r}(P;S), then

e_{n,r}(P;S)\leq e_{n,s}(P;S)

Furthermore, if \text{\normalfont supp} {(P)} is compact (\pi_S(K) is hence compact) then

\lim\limits_{r\rightarrow \infty}e_{n,r}(P;S)=e_{n,\infty}(P;S)

Used by Proof, Proof

Proof of Lemma 18

For the first argument, using the Jensen’s inequality:

with \alpha_{r_k}\subseteq S and \alpha\subseteq S. In addition:

from above we can observe

Lemma

Lemma 19 (Characterization of e_{\infty,r}(P;S) by the projection map)

TM57

Given the conditions (?) or (?) hold, K=\text{\normalfont supp} {(P)} a compact set in \mathbb R^D, then e_{\infty,r}(P;S)=\bigg({\int 0.9355(x,\pi_S(x))^rdP(x)}\bigg)^{1/r}.

Used by Proof

Proof of Lemma 19

By the definition of d_H, if condition (?) or (?) hold, suppose \gamma_n to be a sequence of quantizers, then by Lemma 17, we know that \lim\limits_{n\rightarrow\infty}\sup\limits_{z\in\pi_S(K)}0.9355(z,\gamma_n)=0, this implies that for x\in K, we have that for some c\in \gamma_n, 0.9355(x,\gamma_n)^r=0.9355(x,c)^r\leq (0.9355(x,\pi_S(x))+0.9355(\pi_S(x),c))^r<0.9355(x,\pi_S(x))^r+\varepsilon. Therefore, for n large enough,

\begin{aligned} e_{n,r}(P;S)=\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}\bigg({\int 0.9355(x,\alpha)^rdP(x)}\bigg)^{1/r}=\bigg({\int 0.9355(x,\gamma_n)^rdP(x)}\bigg)^{1/r}<\bigg({\int 0.9355(x,\pi_S(x))^rdP(x)+\varepsilon}\bigg)^{1/r} \end{aligned}

The rest of the proof follows the limit argument.

Uses Lemma 17

Lemma

Lemma 20 (Monotonicity and limit for \tilde{e}_{n,r}(P;S) and \hat{e}_{n,r}(P;S))

O6LU

Let n\in\mathbb N, if conditions (?) or (?) are satisfied, or S\subseteq \text{\normalfont supp} {(P)}\subseteq \mathbb R^D. Suppose 0< r\leq s\leq\infty, then

\tilde{e}_{n,r}(P;S)\leq \tilde{e}_{n,s}(P;S)\quad\text{and}\quad \hat{e}_{n,r}(P;S)\leq \hat{e}_{n,s}(P;S)

Moreover,

\lim\limits_{r\rightarrow\infty}\tilde{e}_{n,r}(P;S)=e_{n,\infty}(P;S)-e_{\infty,\infty}(P;S)\quad\text{ and }\quad\lim\limits_{r\rightarrow\infty}\hat{e}_{n,r}(P;S)=e_{n,\infty}(P;S)

Used by Proof, Proof

Proof of Lemma 20

Suppose m>n with m,n\in\mathbb N and 0< r\leq s.

\begin{aligned} &e_{n,r}^r(P;S)-e_{m+n,r}^r(P;S)=\inf\limits_{\substack{\alpha,\beta,\gamma\subseteq S\\ |\alpha|,|\gamma|\leq n\\|\beta|\leq m}}\int \bigg({0.9355(\gamma,x)^r-0.9355(\alpha\cup\beta,x)^r}\bigg)dP(x)\\ \leq &\inf\limits_{\substack{\alpha,\beta\subseteq S\\ |\alpha|\leq n\\|\beta|\leq m}}\int \bigg({0.9355(\alpha,x)^r-0.9355(\alpha\cup\beta,x)^r}\bigg)dP(x)\leq \inf\limits_{\substack{\alpha,\beta\subseteq S\\ |\alpha|\leq n\\|\beta|\leq m}}\int \bigg({0.9355(\alpha,x)^s-0.9355(\alpha\cup\beta,x)^s}\bigg)^{\frac{r}{s}}dP(x)\\ \leq &\inf\limits_{\substack{\alpha,\beta\subseteq S\\ |\alpha|\leq n\\|\beta|\leq m}}\bigg[{\int \bigg({0.9355(\alpha,x)^s-0.9355(\alpha\cup\beta,x)^s}\bigg)dP(x)}\bigg]^{\frac{r}{s}}\leq \inf\limits_{\substack{\alpha\subseteq \gamma_{m+n}\\ |\alpha|\leq n}}\bigg[{\int\bigg({0.9355(\alpha,x)^s-0.9355(\gamma_{n+m},x)^s}\bigg)dP(x)}\bigg]^{\frac{r}{s}} \end{aligned}

where \gamma_{n+m} above is the minimizer of e_{m+n,s}(P;S). The second inequality holds based on a fact that x^a\leq y^a+(x-y)^a if 0<a\leq 1 and x\geq y\geq 0, as well as the observation that 0.9355(\alpha,x)\geq 0.9355(\alpha\cup\beta,x). Lemma 17 informed us that \lim\limits_{m\rightarrow\infty}d_H(\gamma_{n+m},S)=\lim\limits_{m\rightarrow\infty}\max\limits_{s\in S}\min\limits_{g\in\gamma_{n+m}}0.9355(s,g)=0. Let \gamma_n be minimizing e_{n,s}(P;S), then we also have \lim\limits_{m\rightarrow\infty}d_H(\gamma_{n+m},\gamma_n)=0 and hence we are able to find some \gamma_n'\subseteq \gamma_{m+n} with |\gamma_n'|\leq n such that

\begin{aligned} \inf\limits_{\substack{\alpha\subseteq \gamma_{m+n}\\ |\alpha|\leq n}}\bigg[{\int\bigg({0.9355(\alpha,x)^s-0.9355(\gamma_{n+m},x)^s}\bigg)dP(x)}\bigg]^{\frac{r}{s}}\leq& \bigg[{\int 0.9355(\gamma_n',x)^sdP(x)-e_{m+n,s}^s(P;S)}\bigg]^{\frac{r}{s}} \\\leq \bigg[{\bigg({e_{n,s}(P;S)+\varepsilon}\bigg)^s-e_{m+n,s}^s(P;S)}\bigg]^{\frac{r}{s}} \end{aligned}

for sufficiently small \varepsilon>0 and sufficiently large m>M for some M\in\mathbb N. Consequently

\begin{aligned} &e_{n,r}^r(P;S)-e_{m+n,r}^r(P;S)\leq \bigg[{\bigg({e_{n,s}(P;S)+\varepsilon}\bigg)^s-e_{m+n,s}^s(P;S)}\bigg]^{\frac{r}{s}}\\=&\bigg[{\bigg({e_{n,s}(P;S)+\varepsilon}\bigg)^{r\cdot\frac{s}{r}}-e_{m+n,s}^{r\cdot\frac{s}{r}}(P;S)}\bigg]^{\frac{r}{s}}\\\leq& \bigg[{\bigg({e_{n,s}(P;S)+\varepsilon}\bigg)^r-e_{m+n,s}^r(P;S)}\bigg]^{\frac{r}{s}\cdot\frac{s}{r}}=\bigg({e_{n,s}(P;S)+\varepsilon}\bigg)^r-e_{m+n,s}^r(P;S). \end{aligned}

Therefore

e_{n,r}(P;S)-e_{m+n,r}(P;S)\leq e_{n,s}(P;S)+\varepsilon-e_{m+n,s}(P;S)

Then we can take the limit m\rightarrow\infty

\begin{aligned} e_{n,r}(P;S)-e_{\infty,r}(P;S)&=\lim\limits_{m\rightarrow\infty}\bigg({e_{n,r}(P;S)-e_{m+n,r}(P;S)}\bigg)\\ &\leq \lim\limits_{m\rightarrow\infty}\bigg({e_{n,s}(P;S)+\varepsilon-e_{m+n,s}(P;S)}\bigg)=e_{n,s}(P;S)+\varepsilon-e_{\infty,s}(P;S) \end{aligned}

for any \varepsilon>0, and therefore e_{n,r}(P;S)-e_{\infty,r}(P;S)\leq e_{n,s}(P;S)-e_{\infty,s}(P;S). Similarly,

\begin{aligned} \bigg({e_{n,r}^r(P;S)-e_{m+n,r}^r(P;S)}\bigg)^{\frac{1}{r}}&\leq\bigg({(e_{n,s}(P;S)+\varepsilon)^r-e_{m+n,s}^r(P;S)}\bigg)^{\frac{1}{r}}=\bigg({(e_{n,s}(P;S)+\varepsilon)^{s\cdot\frac{r}{s}}-e_{m+n,s}^{s\cdot\frac{r}{s}}(P;S)}\bigg)^{\frac{1}{r}}\\ &\leq \bigg({(e_{n,s}(P;S)+\varepsilon)^{s}-e_{m+n,s}^s(P;S)}\bigg)^{\frac{1}{s}} \end{aligned}

and therefore

\begin{aligned} \bigg({e_{n,r}^r(P;S)-e_{\infty,r}^r}\bigg)^{\frac{1}{r}}&=\lim\limits_{m\rightarrow\infty}\bigg({e_{n,r}^r(P;S)-e_{m+n,r}^r(P;S)}\bigg)^{\frac{1}{r}}\\ &\leq\lim\limits_{m\rightarrow\infty}\bigg({(e_{n,s}(P;S)+\varepsilon)^{s}-e_{m+n,s}^s(P;S)}\bigg)^{\frac{1}{s}}=\bigg({(e_{n,s}(P;S)+\varepsilon)^{s}-e_{\infty,s}^s(P;S)}\bigg)^{\frac{1}{s}} \end{aligned}

As \varepsilon>0 is arbitrary, we have that

\hat{e}_{n,r}(P;S)=\bigg({e_{n,r}^r(P;S)-e_{\infty,r}^r(P;S)}\bigg)^{\frac{1}{r}}\leq \hat{e}_{n,s}(P;S)=\bigg({e_{n,s}^s(P;S)-e_{\infty,s}^s(P;S)}\bigg)^{\frac{1}{s}}

For the second part, first, informed by Lemma 18 we know that \lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)=e_{n,\infty}(P;S) and thus it suffices to show that \lim\limits_{r\rightarrow\infty}e_{\infty,r}(P;S)\leq\lim\limits_{n\rightarrow\infty}e_{n,\infty}(P;S), as \lim\limits_{r\rightarrow\infty}\tilde{e}_{n,r}(P;S)\leq e_{n,\infty}(P;S)-e_{\infty,\infty}(P;S) by the monotonicity in the previous part of this proof. To show \lim\limits_{r\rightarrow\infty}e_{\infty,r}(P;S)\leq\lim\limits_{n\rightarrow\infty}e_{n,\infty}(P;S), by conditions (?) and (?) and Lemma 17, on the one hand:

\begin{aligned} \lim\limits_{r\rightarrow\infty}e_{\infty,r}(P;S)&=\lim\limits_{r\rightarrow\infty}\lim\limits_{n\rightarrow\infty}\bigg({\int\inf\limits_{\substack{\gamma\subseteq S\\ |\gamma|\leq n}}0.9355(x,\gamma)^rdP(x)}\bigg)^{\frac{1}{r}}=\lim\limits_{r\rightarrow\infty}\bigg({\int 0.9355(x,\pi_S(x))^rdP(x)}\bigg)^{\frac{1}{r}}=\sup\limits_{x\in K}\inf\limits_{y\in\pi_S(K)}0.9355(x,y) \end{aligned}

where K=\text{\normalfont supp} {(P)}. On the other hand, if denote \{\tau_{n,r}\} to be one quantizer of e_{n,r}(P;S), with \tau_n to be the d_H cluster point of one subsequence \{\tau_{n,r_k}\}, then similar to the proof of Lemma 18:

\begin{aligned} \lim\limits_{n\rightarrow\infty}e_{n,\infty}(P;S)&=\lim\limits_{n\rightarrow\infty}\lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)=\lim\limits_{n\rightarrow\infty}\lim\limits_{r\rightarrow\infty}\bigg({\int 0.9355(x,\tau_{n,r})^r dP(x)}\bigg)^{\frac{1}{r}}\geq \lim\limits_{n\rightarrow\infty}\lim\limits_{k\rightarrow\infty}\bigg({\int 0.9355(\tau_n,x)^{r_k}dP(x)}\bigg)^{\frac{1}{r_k}}\\ &=\lim\limits_{n\rightarrow\infty}\sup\limits_{x\in\text{\normalfont supp} {(P)}}0.9355(\tau_n,x)\geq \sup\limits_{x\in\text{\normalfont supp} {(P)}}\inf\limits_{y\in\pi_S(\text{\normalfont supp} {(P)})}0.9355(x,y) \end{aligned}

To show \lim\limits_{r\rightarrow\infty}\hat{e}_{n,r}(P;S)=e_{n,\infty}(P;S), if we denote L=\limsup\limits_{r\rightarrow\infty}\frac{e_{\infty,r}(P;S)}{e_{n,r}(P;S)} and with loss of generality consider L<1, then on the one hand for \varepsilon>0 sufficiently small, we can observe that:

\begin{aligned} \lim\limits_{r\rightarrow\infty}\hat{e}_{n,r}(P;S)&= \lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)\bigg({1-\bigg({\frac{e_{\infty,r}(P;S)}{e_{n,r}(P;S)}}\bigg)^r}\bigg)^{\frac{1}{r}}\geq \lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)\bigg({1-(L+\varepsilon)^r}\bigg)^{\frac{1}{r}}=\lim\limits_{r\rightarrow\infty}e_{n,r}(P;S) \end{aligned}

If L=1, we know that

\lim\limits_{r\rightarrow\infty}\hat{e}_{n,r}(P;S)\geq \lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)\bigg({1-(L+\varepsilon)^r}\bigg)^{\frac{1}{r}}=\lim\limits_{r\rightarrow\infty}e_{n,r}(P;S)\cdot(1+\varepsilon)

where \varepsilon>0 is arbitrary. On the other hand, \lim\limits_{r\rightarrow\infty}\hat{e}_{n,r}(P;S)\leq e_{n,\infty}(P;S) by the previously proven facts.

Uses Lemma 17, Lemma 18

Lemma

Lemma 21

QUYZ

Suppose S\Subset \mathbb R^D, K\Subset \mathbb R^D with a metric d on Euclidean space. Define \pi_S(x):=\{y\in S:0.9355(x,y)=0.9355(x,S)\}. Let e_{\infty,\infty}(K;S)=\lim\limits_{n\rightarrow\infty}e_{n,\infty}(X;S). If the following condition satisfies:

\pi_S(K)=\{s\in S:s=\pi_S(x) \text{ for some }x\in K\} exists as a bounded set;
d_H(\pi_S(K),K)\leq \lim\limits_{n\rightarrow\infty}\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}d_H(\alpha,K)=e_{\infty,\infty}(K;S)\quad\quad

then

e_{n,\infty}(K;S)-e_{\infty,\infty}(K;S)\leq e_{n,\infty}(\pi_S(K);\pi_S(K))

In particular, if \pi_S(K)=S, then e_{n,\infty}(K;S)-e_{\infty,\infty}(K;S)\leq e_{n,\infty}(S).

Used by Proof

Proof of Lemma 21

By Lemma 7, e_{n,\infty}(K;S)=\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}d_H(\alpha,K) and e_{\infty,\infty}(K;S)=\lim\limits_{n\rightarrow \infty}\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}d_H(\alpha,K). Define a new metric r(x,y):=0.9355(\pi_S(x),\pi_S(y)) on K, and denote the covering radius under r as e_{n,\infty}(X;S;r), then:

e_{n,\infty}(K;S;r)=\inf\limits_{\substack{\alpha\subseteq S \\ |\alpha|\leq n}}\sup\limits_{x\in K}\inf\limits_{a\in\alpha}0.9355(a,\pi_S(x))=\inf\limits_{\substack{\alpha\subseteq S \\ |\alpha|\leq n}}\sup\limits_{y\in \pi_S(K)}\inf\limits_{a\in\alpha}0.9355(a,y)=e_{n,\infty}(\pi_S(K))

then

\begin{aligned} e_{n,\infty}(K;S)-e_{\infty,\infty}(K;S)&\leq\inf\limits_{\substack{\alpha\in S\\|\alpha|\leq n}}\sup\limits_{x\in K}\inf\limits_{a\in\alpha}\bigg({\inf\limits_{p\in\pi_S(x)}\bigg({0.9355(p,x)+0.9355(a,p)}\bigg)}\bigg)-e_{\infty,\infty}(K;S)\\ &\leq \inf\limits_{\substack{\alpha\subseteq S \\ |\alpha|\leq n}}\sup\limits_{x\in K}\inf\limits_{a\in\alpha}\bigg({0.9355(\pi_S(x),x)+0.9355(a,\pi_S(x))}\bigg)-e_{\infty,\infty}(K;S)\\ &\leq \sup\limits_{x\in K}0.9355(\pi_S(x),x)+\inf\limits_{\substack{\alpha\subseteq S \\ |\alpha|\leq n}}\sup\limits_{y\in K}\inf\limits_{a\in\alpha}0.9355(a,\pi_S(y))-e_{\infty,\infty}(K;S)\\ &=\sup\limits_{x\in K}\inf\limits_{a\subseteq\pi_S(K)}0.9355(a,x)+e_{n,\infty}(\pi_S(K);S)-e_{\infty,\infty}(K;S)\\ &\leq d_H(\pi_S(K),K)-\lim\limits_{n\rightarrow \infty}\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}d_H(\alpha,K)+e_{n,\infty}(\pi_S(K);S)\\ &\leq e_{n,\infty}(\pi_S(K);S) \leq e_{n,\infty}(\pi_S(K);\pi_S(K)) \end{aligned}

Uses Lemma 7

Lemma

Lemma 22

AJAB

If a compact set K\Subset \mathbb R^D is lower regular in dimension d with respect to some finite measure \eta, say cr^{d}\leq\eta(B(a,r)) for each a\in K and each r\in(0,r_0) for some r_0>0. Then

\liminf\limits_{n\rightarrow \infty}ne_{n,\infty}(K;K)^{d}\leq\limsup\limits_{n\rightarrow\infty}ne_{n,\infty}(K;K)^{d}<\infty

Used by Proof

Proof of Lemma 22

The proof will be in lines of [graf, Section 12.17]. The proof will be trivial if e_{n_0,\infty}(K;K)=0 for some n_0\in\mathbb N. Consider the case that e_{n,\infty}(K;K)>0 for all n\in\mathbb N. We can find a set \alpha_n\subseteq K with the maximal cardinality (guaranteed by the compactness of K) satisfying 0.9355(x,y)\geq e_{n,\infty}(K;K) for all x,y\in\alpha_n such that x\neq y. We claim that |\alpha_n|>n. Assume the contrary that |\alpha_n|\leq n, by definition e_{n,\infty}(K;K)\leq \sup\limits_{x\in K}0.9355(\alpha_n,x) and hence by compactness there exists some z\notin \alpha_n with 0.9355(z,a)\geq e_{n,\infty}(K;K) for all a\in \alpha_n, contradicting the maximality. For each x,y\in\alpha_n we know that B(x,\dfrac{e_{n,\infty}(K;K)}{2})\cap B(y,\dfrac{e_{n,\infty}(K;K)}{2})=\varnothing and therefore

\begin{aligned} \infty>\eta(K)&\geq \eta\bigg({\bigcup\limits_{a\in\alpha_n}B(a,\frac{e_{n,\infty}(K;K)}{2})}\bigg)=\sum\limits_{a\in\alpha_n}\eta(B(a,\frac{e_{n,\infty}(K;K)}{2}))\\ &\geq \frac{c}{2^{d}}|\alpha_n|e_{n,\infty}(K;K)^{d}>\frac{c}{2^{d}} n e_{n,\infty}(K;K)^{d} \end{aligned}

Here we finish the proof.

Corollary

Corollary 23 (Upper asymptotics in terms of Ahlfors dimensions)

KTYR

Suppose S\subseteq\mathbb R^D is closed, K\Subset\mathbb R^D is compact, where \pi_S(K) is regular of dimension d. Suppose conditions (?) or (?) are true. Let P be a probability measure compactly supported on \text{\normalfont supp} {(P)}=K. Denote e_{\infty,\infty} (K)=\lim\limits_{n\rightarrow\infty}e_{n,\infty}(K) and e_{\infty,r}(P;S)=\lim\limits_{n\rightarrow\infty}e_{n,r}(P;S). We have the following

\limsup\limits_{n\rightarrow\infty}\tilde{e}_{n,r}^d(P;S)\leq\limsup\limits_{n\rightarrow\infty}n\tilde{e}_{n,\infty}^d(P;S)<\infty,

which implies that for 0<r<\infty,

\begin{aligned} \dim_H(\pi_S(K))&\geq \overline{\widetilde{\dim}}_Q^{r}{(P;S)}\geq \underline{\widetilde{\dim}}_Q^{r}{(P;S)},\\\dim_H(\pi_S(K))&\geq \overline{\widetilde{\dim}}_Q^{\infty}{(P;S)}\geq \underline{\widetilde{\dim}}_Q^{\infty}{(P;S)},\\ \dim_H(\pi_S(K))&\geq\overline{\widehat{\dim}}_Q^{1}{(P;S)}\geq\underline{\widehat{\dim}}_Q^{1}{(P;S)}. \end{aligned}

Used by Example 26

Proof of Corollary 23

Lemma 21, Lemma 35 and Lemma 17, as well as Lemma 22, gives

\limsup\limits_{n\rightarrow\infty}n\bigg({e_{n,\infty}(K;S)-e_{\infty,\infty}(K;S)}\bigg)^{d}<\infty

We further have that

0<\limsup\limits_{n\rightarrow\infty}n\bigg({e_{n,r}(K;S)-e_{\infty,r}(K;S)}\bigg)^{d}\leq \limsup\limits_{n\rightarrow\infty}n\bigg({e_{n,\infty}(K;S)-e_{\infty,\infty}(K;S)}\bigg)^{d}

guaranteed by Lemma 20. By Lemma 34, \dim_H(\pi_S(K)=d). The proof of the case that \text{\normalfont supp} {(P)} is not compact and S not a subset of K is from a similar argument.

Uses Lemma 21, Lemma 35, Lemma 17, Lemma 22, Lemma 20, Lemma 34

Proposition

Proposition 24 (Upper asymptotics in terms of box dimensions)

RT1J

Suppose T:K\rightarrow S is constructed as Lemma 11 and either conditions (?) or (?) hold. Let P be a compactly supported probability measure on K. Then for r\geq 1 we have

\begin{aligned} &\underline{\widehat{\dim}}_Q^{r}{(P;S)}\leq\underline{\dim}_B^{*}{(T_*P)}\leq\underline{\dim}_B^{}{(\pi_S(K))},\,\,\overline{\widehat{\dim}}_Q^{r}{(P;S)}\leq\overline{\dim}_B^{*}{(T_*P)}\leq\overline{\dim}_B^{}{(\pi_S(K))};\\ &\underline{\widetilde{\dim}}_Q^{r}{(P;S)}\leq\underline{\dim}_B^{*}{(T_*P)}\leq\underline{\dim}_B^{}{(\pi_S(K))},\,\,\overline{\widetilde{\dim}}_Q^{r}{(P;S)}\leq\overline{\dim}_B^{*}{(T_*P)}\leq\overline{\dim}_B^{}{(\pi_S(K))}. \end{aligned}

Proof of Proposition 24

Given conditions (?) and (?) and by Lemma 17, and Lemma 19, \hat{e}_{n,r}^r(P;S)=\inf\limits_{\substack{{\alpha\subseteq S}\\ |\alpha|\leq n}}\int 0.9355(x,\alpha)^r - 0.9355(x,\pi_S(x))^r dP(x). Let \{A_i\}_{i\leq n} be a constrained Voronoi’s partition with respect to P and S, with \beta being one optimal quantizer, then we have:

\begin{aligned} n^{\frac{1}{l}}\hat{e}_{n,r}(\mu;S)=n^{\frac{1}{l}}\bigg({\inf\limits_{\substack{{\alpha\subseteq S}\\ |\alpha|\leq n}}\int 0.9355(x,\alpha)^r -0.9355(x,\pi_S(x))^r dP(x)}\bigg)^{\frac{1}{r}}=n^{\frac{1}{l}}\bigg({\sum\limits_{i=1}^n\int_{A_i} 0.9355(x,\alpha)^r - 0.9355(x,\pi_S(x))^r dP(x)}\bigg)^{\frac{1}{r}} \end{aligned}

Lemma 11 indicates that T is measurable, and it is a classical result that the push-forward measure T_*P, where T_*P(E)=P(T^{-1}(E)), is a Borel measure on \mathcal{B}(\pi_S(K)). Consider the constrained n-quantizer \gamma=\{a_i\}_{i\leq n} for e_{n,r}(T_*P;S) with corresponding to Voronoi’s partition \{B_i\}_{i\leq n}. Furthermore, let a sequence of Borel sets \{C_i\}_{i\leq n} be a partition of K such that for x\in C_i, \pi_S(x)\in B_i. Let g_i:B_i\rightarrow[0,\infty) satisfy g:x\mapsto 0.9355(a_i,x) and consider T constructed in Lemma 11 we then have the following estimate:

\begin{aligned} n^{\frac{1}{l}}\tilde{e}_{n,r}(P;S)&\leq n^{\frac{1}{l}}\hat{e}_{n,r}(P;S)= n^{\frac{1}{l}}\bigg({\sum\limits_{i=1}^n\int_{A_i} \bigg[{0.9355(x,\beta)^r - 0.9355(x,\pi_S(x))^r}\bigg] dP(x)}\bigg)^{\frac{1}{r}}\\&\leq n^{\frac{1}{l}}\bigg({\sum\limits_{i=1}^n\int_{C_i} \bigg[{0.9355(x,\gamma)^r - 0.9355(x,\pi_S(x))^r}\bigg] dP(x)}\bigg)^{\frac{1}{r}}\leq n^{\frac{1}{l}}\bigg({\sum\limits_{i=1}^n\int_{C_i} 0.9355(a_i,\pi_S(x))^r dP(x)}\bigg)^{\frac{1}{r}} \\&=n^{\frac{1}{l}}\bigg({\sum\limits_{i=1}^n\int_{B_i} 0.9355(a_i,y)^r 0.9355(T_*P(y))}\bigg)^{\frac{1}{r}}=n^{\frac{1}{l}}e_{n,r}(T_*P;\pi_S(K)) \end{aligned}

and for l=\overline{\widehat{\dim}}_Q^{r}{(T_*P)} and t=\underline{\widehat{\dim}}_Q^{r}{(T_*P)} we have

\begin{aligned} &0\leq\limsup\limits_{n\rightarrow\infty}n^{\frac{1}{l}}\hat{e}_{n,r}(P;S)\leq \limsup\limits_{n\rightarrow\infty}n^{\frac{1}{l}}e_{n,r}(T_*P;S)<\infty\\ &0\leq\liminf\limits_{n\rightarrow\infty}n^{\frac{1}{t}}\hat{e}_{n,r}(P;S)\leq \liminf\limits_{n\rightarrow\infty}n^{\frac{1}{t}}e_{n,r}(T_*P;S)<\infty \end{aligned}

therefore \overline{\widehat{\dim}}_Q^{r}{(P;S)}\leq \overline{\widehat{\dim}}_Q^{r}{(T_*P;S)} and \underline{\widehat{\dim}}_Q^{r}{(P;S)}\leq \underline{\widehat{\dim}}_Q^{r}{(T_*P;S)}. Moreover, by Remark 1, the proof of the third argument of Proposition 4 in [dai2008quantization] and [graf, Section 11.3] informed us that \underline{\widehat{\dim}}_Q^{r}{(P;S)}\leq\underline{\dim}_B^{}{(\pi_S(K))}, \overline{\widehat{\dim}}_Q^{r}{(T_*P;S)}\leq \overline{\dim}_B^{*}{(T_*P)}\leq\overline{\dim}_B^{}{(\pi_S(K))}. We can establish the results for \overline{\widetilde{\dim}}_Q^{r}{(P;S)} and \underline{\widetilde{\dim}}_Q^{r}{(P;S)} in a similar way as \tilde{e}_{n,r}(P;S)\leq\hat{e}_{n,r}(P;S).

Uses Lemma 17, Lemma 19, Lemma 11, Remark 1, Proposition 4

Lemma

Lemma 25 (Lower dimensions when quantizers lie out of \pi_S(K))

J39L

Let S be a closed constraint, P be a compactly supported probability with \text{\normalfont supp} {(P)}=K. Suppose that the condition (?) is true. In addition, suppose that for the measurable selection T:K\rightarrow \pi_S(K) constructed in Lemma 11, the pushforward measure T_*P is upper Ahlfors regular of dimension d, then for r\geq 1, \underline{\widehat{\dim}}_Q^{r}{(P;S)}\geq d, and \underline{\widetilde{\dim}}_Q^{r}{(P;S)}\geq d.

Used by Example 26, Proof

Proof of Lemma 25

First, follwing the same notation of (?), (?) and by (?), we have,

\sum\limits_{i=1}^{|\alpha_n|}\int_{V_i}w_\lambda(x,a_i)0.9355(a_i,\pi_S(x))dP(x)-\lambda\leq \hat{e}_{n,1}(P;S),

and,

\hat{e}_{n,1}(P;S)\leq \sum\limits_{i=1}^{|\alpha_n|}\int_{V_i}w_\lambda(x,a_i)0.9355(a_i,\pi_S(x))dP(x)+\lambda.

In the following, we focus on \sum\limits_{i=1}^{|\alpha_n|}\int_{V_i}w_\lambda(x,a_i)0.9355(x,\pi_S(x))dP(x) as \lambda can be chosen arbitrarily small. Let

A_\delta^\lambda(y)=\{x\in K:\omega_\lambda(x,y)\geq \delta\}.

We choose \delta_0=\delta, guaranteed by Lemma 36 and the fact that if w_\lambda(x,y)\geq w_1(x,y)\geq \delta, for

F_\delta=\bigcap\limits_{y\in S\setminus\pi_S(K)}\{x\in K:w_1(x,y)\geq \delta\},\quad \lambda=1,

and for the measurable selector constructed in Lemma 11,

\begin{aligned} \hat{e}_{n,1}(P;S)+\lambda\geq&\sum\limits_{i=1}^{|\alpha_n|}\int_{V_i}w_\lambda(x,a_i)0.9355(a_i,\pi_S(x))dP(x)\\\geq& \sum\limits_{i=1}^{|\alpha_n|}\delta_0\int_{V_i\cap A_{\delta_0}^\lambda(a_i)}0.9355(a_i,\pi_S(x))dP(x)+\sum\limits_{i=1}^{|\alpha_n|}\int_{V_i\setminus A_{\delta_0}^\lambda(a_i)}w_\lambda(x,\pi_S(x))0.9355(a_i,\pi_S(x))dP(x)\\ \geq &\sum\limits_{i=1}^{|\alpha_n|}\delta_0\int_{V_i\cap A_{\delta_0}^\lambda(a_i)}0.9355(a_i,\pi_S(x))dP(x)=\sum\limits_{i=1}^{|\alpha_n|}\delta_0\int_{V_i\cap A_{\delta_0}^\lambda(a_i)}0.9355(a_i,T(x))dP(x). \end{aligned}

In addition,

\begin{aligned} \delta_0\int_{V_i\cap A_{\delta_0}^\lambda(a_i)}0.9355(a_i,T(x))dP(x)&=\int 0.9355(a_i,T(x))dP|_{V_i\cap A_{\delta_0}^\lambda(a_i)}(x)=\int_{T(V_i\cap A_{\delta_0}^\lambda(a_i))}d(a_i,y)dT_*(P|_{V_i\cap A_{\delta_0}^\lambda(a_i)})(y). \end{aligned}

By Lemma 37, there exists some constant C>0 only dependent to T, P such that,

\begin{aligned} \int_{T(V_i\cap A_{\delta_0}^\lambda(a_i))}d(a_i,y)dT_*(P|_{V_i\cap A_{\delta_0}^\lambda(a_i)})(y)&\geq CT_{*}(P|_{V_i\cap A_{\delta_0}^\lambda(a_i)})(T(V_i\cap A_{\delta_0}^\lambda(a_i)))^{1+1/d}\\&\geq CP|_{V_i\cap A_{\delta_0}^\lambda(a_i)}(V_i\cap A_{\delta_0}^\lambda(a_i))^{1+1/d}=CP(V_i\cap A_{\delta_0}^\lambda(a_i))^{1+1/d}. \end{aligned}

By Hölder’s inequality and Lemma 36,

\begin{aligned} n^{1/d}\sum\limits_{i=1}^{|\alpha_n|} P(V_i\cap A_{\delta_0}^\lambda(a_i))^{1+1/d}&\geq |\alpha_n|^{1/d}\sum\limits_{i=1}^{|\alpha_n|} P(V_i\cap A_{\delta_0}^\lambda(a_i))^{1+1/d}\geq \bigg({\sum\limits_{i=1}^{|\alpha_n|}P(V_i\cap A_{\delta_0}^\lambda(a_i))}\bigg)^{1+1/d}\\ &\geq P(\bigcap\limits_{i=1}^{|\alpha_n|}A_{\delta_0}^\lambda(a_i))^{1+1/d}\geq P(\bigcap\limits_{i=1}^{|\alpha_n|}A_{\delta_0}^1(a_i))^{1+1/d}\geq P(F_{\delta_0})^{1+1/d}>0, \end{aligned}

and therefore, for r\geq 1,

\begin{aligned} \liminf\limits_{n\rightarrow\infty} n^{1/d}\hat{e}_{n,r}(P;S)\geq \liminf\limits_{n\rightarrow\infty} n^{1/d}\hat{e}_{n,1}(P;S)>0. \end{aligned}

As \hat{e}_{n,1}(P;S)=\tilde{e}_{n,r}(P;S), and by the monotonicity results in Lemma 20, we can also conculde that,

\begin{aligned} \liminf\limits_{n\rightarrow\infty} n^{1/d}\tilde{e}_{n,r}(P;S)\geq \liminf\limits_{n\rightarrow\infty} n^{1/d}\tilde{e}_{n,1}(P;S)>0. \end{aligned}

Uses Lemma 36, Lemma 11, Lemma 37, Lemma 20

Example

Example 26 (Quantizers outside \pi_S(K))

QPO4

Consider S=\overline{B(0,1)} and K=\partial B(0,1) in \mathbb R^2, then there exist a measurable selector T:K\rightarrow S, x\mapsto x, which is exactly the identity map. Let P=U_{\partial B(0,1)} be the uniform distribution on \partial B(0,1). \pi_S(K)=K has dimension of 1 in the sense of Hausdorff, box, and Ahlfors and quantization. We will then find the lower constrained quantization dimension. By Corollary 23 and Proposition 24, \overline{\widehat{\dim}}_Q^{r}{(P;S)}=\overline{\widetilde{\dim}}_Q^{r}{(P;S)}=1 for r\geq 1. By symmetry the n codepoints form a regular n–gon

\alpha_n=\{(\theta_i,0.9355):\theta_i=\tfrac{2\pi i}{n},\; i=0,\dots,n-1\}, \qquad 0\le0.9355\le1 .

For a point (1,t) on the unit circle, the nearest center is (0.9355,0) after rotation. The squared distance is 1+0.9355^{2}-20.9355\cos t . Therefore

\hat{e}_{n,2}(P;S)=\frac{n}{2\pi}\int_{-\pi/n}^{\pi/n}1+0.9355^{2}-20.9355\cos t \,dt =1+0.9355^{2}-\frac{20.9355 n}{\pi}\sin\!\Bigl(\frac{\pi}{n}\Bigr).

Observe that, \dfrac{d}{d0.9355} \hat{e}_{n,2}(P;S)=20.9355-\dfrac{2n}{\pi}\sin\bigl(\dfrac{\pi}{n}\bigr), and setting \dfrac{d}{d0.9355} \hat{e}_{n,2}(P;S)=0 gives

0.9355=\frac{n}{\pi}\sin\!\Bigl(\frac{\pi}{n}\Bigr),

which coincide with the result in [rosenblatt2023uniform]. Notice that 0.9355<1 for any n as \sin(\pi/n)<\pi/n for n\geq 2, thus there exists a sequence of n-quantizers which lie strictly inside the unit disc as required. This result also coincide with Lemma 17 that under conditions (?) and (?), the quantizers will approximate \pi_S(K). By Lemma 25, \underline{\widehat{\dim}}_Q^{r}{(P;S)}\geq 1 and \underline{\widetilde{\dim}}_Q^{r}{(P;S)}\geq 1, and therefore \widehat{\dim}_Q^{r}{(P;S)}=\widetilde{\dim}_Q^{r}{(P;S)}=1. Our results in Section ? and Section ? coincide with the quantization dimension as presented in [rosenblatt2023uniform].

Uses Corollary 23, Lemma 25

Proposition

Proposition 27 (Lower asymptotics if S\cap K=\varnothing)

TIE1

Suppse S is closed, P is a compactly supported probability with \text{\normalfont supp} {(P)}=K. Let r\geq 1. Suppose A and B are constructed as (?) does. Suppose the following,

either S\cap K=\varnothing, or the nowhere density of \pi_S(K);
there is a sequence of quantizers \{\alpha_n\}_n of e_{n,r}(P;S) such that \alpha_n\cap \overline{B}=\varnothing for all n\in\mathbb N^+;
for the measurable selector T:K\rightarrow\pi_S(K) constructed in Lemma 11, the pushforward measure T_*P is upper Ahlfors regular of dimension d;
for all p\in \pi_S(K)\setminus\overline{B}, there exists some \delta>0, such that (B(p,\delta)\cap A)\setminus\pi_S(K)\neq\varnothing.

Then \underline{\widehat{\dim}}_Q^{r}{(P;S)}\geq d and \underline{\widetilde{\dim}}_Q^{r}{(P;S)}\geq d.

Proof of Proposition 27

The proof of Proposition 27 can be illustrated in two major steps: to construct quasi-quantizers and then reduce the problem to the case in Lemma 25. We denote \{\alpha_n\}_n as a sequence of constrained quantizers of e_{n,r}(P;S). Let S=A\cup B where B is the set of isolated points as Lemma (?). Step 1. Constructing Epsilon-n-Quantizers for \alpha_n\subset \pi_S(K)\setminus \overline{B}.For each \alpha_n, denote each element by \{a_j\}_{j\leq |\alpha_n|}, with the corresponding Voronoi partitions denoted by \{V_j\}_{j\leq |\alpha_n|}. Let \beta_n=\{a\in \alpha_n,a\in\pi_S(K)\}, then for each b\in\beta_n, by the nowhere density of \pi_S(K) guaranteed by Lemma 12, we can perturb each b=a_j inside the ball B(b,\dfrac{\varepsilon}{n^{1/d}}), and then choose \tilde{b}=\tilde{a}_j\not\in\pi_S(K). Guaranteed by (?), \tilde{b}=\tilde{a}_j can be inside S. We can replace each b=a_j\in\beta_n by \tilde{b}=\tilde{a}_j in \alpha_n to form another set \tilde{\alpha}_n. We denote \{c_j\}_{j\leq |\tilde{\alpha}_j|} as elements of \tilde{\alpha}_n. The set \tilde{\alpha}_n is called a epsilon-n-quantizer, because for,

\begin{aligned} e_{n,1}(P;S)=\sum\limits_{\substack{j\leq |\alpha_n|\\ a_j\in\beta_n}}\int_{V_j}0.9355(x,a_j)dP(x)+\sum\limits_{\substack{j\leq |\alpha_n|\\ a_j\not\in\beta_n}}\int_{V_j}0.9355(x,a_j)dP(x), \end{aligned}

we have,

\begin{aligned} \sum\limits_{\substack{j\leq |\alpha_n|\\ a_j\in\beta_n}}\int_{V_j}\bigg({0.9355(x,\tilde{a}_j)-\dfrac{\varepsilon}{ n^{1/d}}}\bigg)dP(x)\leq \sum\limits_{\substack{j\leq |\alpha_n|\\ a_j\in\beta_n}}\int_{V_j}0.9355(x,a_j)dP(x)\leq \sum\limits_{\substack{j\leq |\alpha_n|\\ a_j\in\beta_n}}\int_{V_j}\bigg({0.9355(x,\tilde{a}_j)+\dfrac{\varepsilon}{n^{1/d}}}\bigg)dP(x), \end{aligned}

and therefore for arbitrarily small \varepsilon>0,

\begin{aligned} e_{n,1}(P;S)-e_{\infty,1}(P;S)-\varepsilon/n^{1/d}\leq \sum\limits_{j\leq |\tilde{\alpha}_n|}0.9355(x,c_j)dP(x)-e_{\infty,1}(P;S)\leq e_{n,1}(P;S)-e_{\infty,1}(P;S)+\varepsilon/n^{1/d} \end{aligned}

Step 2. Reducing to Lemma 25 for \alpha_n\subseteq \pi_S(K)\setminus \overline{B}. As \tilde{\alpha}_n\subseteq S\setminus\pi_S(K), we can repeat the argument in Lemma 25, to produce the result that,

\begin{aligned} \liminf\limits_{n\rightarrow\infty}n^{1/d}\bigg({\sum\limits_{j\leq |\tilde{\alpha}_n|}0.9355(x,c_j)dP(x)-e_{\infty,1}(P;S)}\bigg)+\varepsilon\geq \liminf\limits_{n\rightarrow\infty}n^{1/d}\hat{e}_{n,1}(P;S)=\liminf\limits_{n\rightarrow\infty}n^{1/d}\tilde{e}_{n,1}(P;S), \end{aligned}

and,

\begin{aligned} 0<\liminf\limits_{n\rightarrow\infty}n^{1/d}\bigg({\sum\limits_{j\leq |\tilde{\alpha}_n|}0.9355(x,c_j)dP(x)-e_{\infty,1}(P;S)}\bigg)\leq \liminf\limits_{n\rightarrow\infty}n^{1/d}\hat{e}_{n,1}(P;S)+\varepsilon. \end{aligned}

By the squeezing properties we can conclude the lower quantization dimensions when \alpha_n are not confined in S\setminus \pi_S(K).

Uses Lemma 25, Lemma 12

Remark

Remark 6

C13A

The condition (?) is not absolutely necessary to construct the admissible epsilon-n-quantizers. Let a be an element of a n-quantizer on \pi_S(K), recall the construction in (?), we only need the perturbed element \tilde{a} to satisfy 0.9355(\tilde{a},x)>0.9355(x,\pi_S(x)) for any x\in K to produce a uniform lower bound as (?) demonstrated. To be concrete, in Example 29 and Figure ?, the condition (?) does not hold. There exists a sequence of n-quantizers lying on \pi_S(K). There are two sides separated by the left (right) branch of S. Suppose S_\delta to be collection of points on the side without K in the closed \delta-neighborhood of \pi_S(K). Then if b is an element of a n-quantizer, we can obtain a perturbed \tilde{b}\in S_\delta with \tilde{b}\not\in \pi_S(K) and satisfying0.9355(\tilde{b},x)>0.9355(x,\pi_S(x)) for any x\in K.

Proposition

Proposition 28

TE1E

S\subseteq\mathbb R^D is closed and P is a probability compactly supported on K\subseteq\mathbb R^D. Suppose the measurable selector T constructed in Lemma 11 satisfies the Ahlfors regularity of T_*P in dimension d. Let r\geq 1. Suppose either condition (?) or (?) are true. In addition, suppose \underline{\widehat{\dim}}_Q^{r}{(P;S)}\geq d and \underline{\widetilde{\dim}}_Q^{r}{(P;S)}\geq d. Then,

\dim_H^*({T_*P})\leq \underline{\widehat{\dim}}_Q^{r}{(P;S)} and \dim_H^*({T_*P})\leq \underline{\widetilde{\dim}}_Q^{r}{(P;S)};
if condition (?) is true, then \dim_H(T(K))=\dim_H^*({T_*P});
if condition (?) is true and T:K\rightarrow\pi_S(K) is surjective, then \dim_H(\pi_S(K))\leq \underline{\widehat{\dim}}_Q^{r}{(P;S)}\leq \underline{\dim}_B^{}{(\pi_S(K))} and \dim_H(\pi_S(K))\leq \underline{\widetilde{\dim}}_Q^{r}{(P;S)}\leq \underline{\dim}_B^{}{(\pi_S(K))}.

Proof of Proposition 28

By Lemma 34, d=\dim_H(\text{\normalfont supp} {(T_*P)})\geq \dim_H^*({T_*P}). By Lemma 14, given conditions (?) or (?), \dim_H(T(K))\geq \dim_H^*({T_*P}). We now want to show that \dim_H(T(K))\leq \dim_H^*({T_*P}), so assume that \inf\limits_{\substack{E\subset\mathcal{B}(\overline{T(K)})\\T_*P(E)=1}}\dim_H(E)<\dim_H(T(K)), therefore, there exists E\in\mathcal{B}(\overline{T(K)}) such that T_*P(E)=1 and \dim_H(E)<\dim_H(T(K)). Hence \dim_H(\overline{T(K)}\setminus E)>0 and thus \dim_H(\overline{T(K)}\setminus E )=\dim_H(T(K)) by Lemma 33. Let F=\overline{T(K)}\setminus E and s=\dim_H(T(K)). For \delta>0, consider \{ B(x_i,3r_i)\}_{i=1}^n a Vitali covering of F with r_i<\delta such that \{B(x_i,r_i)\}_{1\leq i\leq n} are mutually disjoint. By the lower Ahlfors regularity of T_*P in Lemma 15, we have the estimate:

\begin{aligned} T_*P(F)=\sum\limits_{i=1}^nT_*P(B(x_i,r_i)\cap F)\geq C\sum\limits_{i=1}^n r_i^s=\frac{C}{3^s}\sum\limits_{i=1}^n (3r_i)^s\geq \frac{C}{3^s}\mathcal{H}_{3\delta}^s(F) \end{aligned}

Therefore T_*P(F)\geq \frac{C}{3^s}\mathcal{H}^s(F)>0 by the fact that \dim_H(F)>0, contradicting the fact that T_*P(F)=T_*P(\overline{T(K)}\setminus E)=0. Finally, \dim_H(\pi_S(K))=\dim_H{T(K)}=\dim_H^*(T_*P) if T is surjective. We can establish the results for \overline{\widetilde{\dim}}_Q^{r}{(P;S)} and \underline{\widetilde{\dim}}_Q^{r}{(P;S)} in a similar way as \tilde{e}_{n,r}(P;S)\leq\hat{e}_{n,r}(P;S).

Uses Lemma 34, Lemma 14, Lemma 33, Lemma 15

Example

Example 29

BDHK

Suppose

K=\{(x,y)\in\mathbb R^2:y=-|x|,|x|\leq 1\},

S=\{(x,y)\in\mathbb R^2:|y|=x-1\}\cup\{(x,y)\in\mathbb R^2:|y|=-x-1\}\cup\{(x,y)\in\mathbb R^2: y=1+|x|\}.

We can calculate that,

\pi_S(K)=\{(x,y)\in\mathbb R^2:|y|=-x-1,-3/2\leq x\leq -1\}\cup \{(x,y)\in\mathbb R^2:|y|=x-1,1\leq x\leq 3/2\}\cup\{(0,1)\}

We have \pi_S^{-1}(B((0,1),\varepsilon)\cap S)=\{(0,0)\}. Let P to be uniform on K, then P(\pi_S^{-1}(\{(0,0)\})=0.

Example

Example 30 (Dirac measure)

YWJK

Let P be a Dirac distribution on K=\{(0,0)\} where S=\{(x,y)\in\mathbb R^2:x^2+y^2=1\}. Then we can construct T:\{(0,0)\}\rightarrow \{(\cos\theta,\sin\theta)\} for any \theta\in[0,2\pi) with T(K)=\{(\cos\theta,\sin\theta)\}. T_*P is another Dirac distribution only defined on T(K), while \pi_ S(K)=S. \dim_H(T(K))=\dim_H(K)=0. We may verify that \pi_S^{-1}(B(x,\varepsilon)\cap S)=\{y\in \{(0,0)\}:\pi_S(y)\in B(x,\varepsilon)\cap S\}=\varnothing as \pi_S^{-1}(\{(0,0)\})=S, and therefore the condition (?) does not hold. The condition (?) does not hold either as, for a randomly selected \beta_n\subset S\text{ with }|\beta_n|=1 are a sequence of quantizers, which does not satisfy that \lim\limits_{n\rightarrow\infty}d_H(\beta_n, \pi_S(K))=0.

Example

Example 31 (Uniform on unit circle)

UTBF

Let P be a uniform distribution on K=\{(x,y)\in\mathbb R^2:x^2+y^2=1\} with S=\{(x,y)\in\mathbb R^2:x^2+y^2\leq \frac{1}{4}\}. Then \pi_S(K)=T(K)=\partial S=\{(x,y)\in \mathbb R^2:x^2+y^2=\frac{1}{4}\}. We can construct T:K\rightarrow \partial S,(x,y)\mapsto (x/2,y/2). T_*P will be the uniform distribution on \partial S. For sufficiently small \varepsilon>0 and x\in\pi_S(K) we can verify that \pi_S^{-1}(B(x,\varepsilon)\cap S)\supseteq B(y,c\varepsilon)\cap K for some c>0, and therefore conditions (?) and (?) hold. The condition (?) also holds as the uniform distribution on K is lower Ahlfors regular of dimension 1.

Example

Example 32 (Middle-third Cantor)

3YF8

This example was given in [pandey2024constrained]. Let P be the Cantor distribution on the Cantor set \mathcal{C} on \{(x,y)\in\mathbb R^2: x\in [0,1],y=0\}. Let S=\{(x,y)\in\mathbb R^2:y=1\}, then \pi_S(\mathcal{C})=\{(x,y)\in \mathbb R^2:(x,y-1)\in\mathcal{C}\}. There exists a surjective map T:\mathcal{C}\rightarrow\pi_S(\mathcal{C}), (x,y)\mapsto (x,y+1). Therefore \dim_H{\pi_S(\mathcal{C})}=\dim_H(T(\mathcal{C}))=\dim_H^*({T_*\mathcal{C}})=\dim_H(\mathcal{C})=\dfrac{\log 2}{\log 3}.

Lemma

Lemma 33 (Some properties on \dim_H)

GY31

For \{A_i\}_{i\geq1} a sequence of sets in \mathbb R^D, and A\subseteq B\subseteq\mathbb R^D:

\dim_H(\bigcup\limits_{i=1}^\infty A_i)=\sup\limits_{1\leq i<\infty}\dim_H(A_i);
If \dim_H(A)>0, then \text{\normalfont diam}(A)>0;
If \dim_H(A)<\dim_H(B), then \dim_H(A\setminus B)>0

Used by Proof

Lemma

Lemma 34

198T

Let X be a seperable metric space. Suppose for E\subset X is a closed set in X upon which there exists a positive measure \eta:\mathcal{B}(E)\rightarrow[0,\infty], then

Suppose \eta is lower Ahlfors regular of dimension s, that is, there exists c,r_0>0, for any x\in E and r<r_0 we have \eta(B(x,r)\cap S)\geq cr^s. Then s\geq\dim_H(E)
Suppose \eta is upper Ahlfors regular of dimension s, that is, there exists c,r_0>0, for any x\in E and r<r_0 we have \eta(B(x,r)\cap S)\leq cr^s. Then s\leq \dim_H(E).

Used by Proof, Proof

Proof of Lemma 34

For the first argument, by the general Vitali covering lemma, there exists a sequence of balls \{B(x_i,r_i)\}_{i=1}^\infty with r_i<r_0 such that \{B(x_i,5r_i)\}_{i=1}^\infty cover E while B(x_i,x_j) are mutually disjoint. Therefore

\begin{aligned} \eta(E)\geq \sum\limits_{i=1}^\infty \eta(B(x_i,r_i)\cap E)\geq c\sum\limits_{i=1}^\infty r_i^s \end{aligned}

and

\begin{aligned} \mathcal{H}_\delta^s(E)=\inf\limits_{\substack{E\subseteq \bigcup\limits_{i=1}^\infty U_i\\ \text{\normalfont diam}(U_i)<\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(U_i)^s\leq 5^s\inf\limits_{\substack{E\subseteq \bigcup\limits_{i=1}^\infty B(x_i,5r_i)\\5r_i<r}}\sum\limits_{i=1}^\infty r_i^s\leq 5^sc\eta(E) \end{aligned}

thus \mathcal{H}^s(E)\leq 5^sc\eta(E) which implies that s\geq \dim_H(E). For the second argument, for sufficiently small \delta>0, we have

\begin{aligned} \mathcal{H}_\delta^s(E)=\inf\limits_{\substack{E\subseteq \bigcup\limits_{i=1}^\infty U_i\\ \text{\normalfont diam}(U_i)<\delta}}\sum\limits_{i=1}^\infty \text{\normalfont diam}(U_i)^s\geq c\sum\limits_{i=1}^\infty\eta(B(x_i,\text{\normalfont diam}(U_i))\cap S)\geq c\sum\limits_{i=1}^\infty \eta(U_i\cap S)\geq c\eta(E) \end{aligned}

where x_i are arbitrarily selected from U_i. This implies that s\leq \dim_H(E).

Lemma

Lemma 35

YRMK

Suppose conditions (?) or (?) are satisfied and \text{\normalfont supp} {(P)}=K is compactly. Let \alpha_n\in C_{n,\infty}(K;\pi_S(K)) be the attained n-optimal set, if there exists a sequence of \{\alpha_n\}_n satisfies \lim\limits_{n\rightarrow\infty}d_H(\alpha_n,\pi_S(K))=0, then

d_H(\pi_S(K),K)\leq \lim\limits_{n\rightarrow\infty}\inf\limits_{\substack{\alpha\subseteq S\\ |\alpha|\leq n}}d_H(\alpha,K)

Used by Proof

Proof of Lemma 35

For any x\in K'\cup K we have |0.9355(x,A)-0.9355(x,B)|\leq d_H(A,B) for any nonempty sets A\subseteq K'\cup K and B\subseteq K'\cup K. Here, K' is compact with \pi_S(K)\subseteq K'. K' also satisfies that \alpha_n\subseteq K' for n\geq N for some N\in\mathbb N^{+} as \lim\limits_{n\rightarrow\infty}d_H(\alpha_n,\pi_S(K))=0. Furthermore, for some y\in K'\cup K the equality can be achieved for A=\alpha_n and B=K. Therefore,

\begin{aligned} d_H(\alpha_n,K)-d_H(\pi_S(K),K)&\geq|0.9355(y,\alpha_n)-0.9355(y,K)|-d_H(\pi_S(K),K)\\ &=|0.9355(y,\alpha_n)-0.9355(y,\pi_S(K))+0.9355(y,\pi_S(K))-0.9355(y,K)|-d_H(\pi_S(K),K)\\ &\geq-|0.9355(y,\alpha_n)-0.9355(y,\pi_S(K))|+|0.9355(y,\pi_S(K))-0.9355(y,K)|-d_H(\pi_S(K),K)\\ &= -|0.9355(y,\alpha_n)-0.9355(y,\pi_S(K))|\geq-\sup\limits_{x\in K}|0.9355(x,\alpha_n)-0.9355(x,\pi_S(K))|\\ &=-d_H(\alpha_n,\pi_S(K)) >-\varepsilon \end{aligned}

followed by a limiting argument. The second equality above is due to we set |0.9355(x,\pi_S(K))-0.9355(x,K)|=d_H(\pi_S(K),K)

Lemma

Lemma 36

3R16

For F_\delta constructed in (?), Suppose for a Borel probability P we have P(\bigcup\limits_{\delta>0}F_\delta)>0, then there exists some \delta_0>0 such that P(F_{\delta_0})>0.

Used by Proof

Lemma

Lemma 37

9CR3

Let \mu be a finite Borel measure on \mathbb R^D. Assume that there is some c>0 and some r_0>0 such that \mu(B(a,r))\leq cr^D for all a\in\text{\normalfont supp} {(\mu)} and all r\in (0,r_0), then there exists some constant c' dependent only to \mu and D, such that,

\int_B 0.9355(x,a)d\mu(x)\geq c'\mu(B)^{1+1/D},

for all a\in\mathbb R^D and all B\in\mathcal{B}(\mathbb R^D).

Used by Proof

Source and licence

The text on this page is from arXiv:2505.10801v1 by Chenxing Qian, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.

The record of each item is in analysis/objects/2505/2505.10801.jsonl.