Skip to main content
Bounty Ended with 100 reputation awarded by Arbuja
added 792 characters in body
Source Link
Iosif Pinelis
  • 127.7k
  • 8
  • 107
  • 229

Here is another possible approach, perhaps closer to what the OP had in mind.

Let $S:=[0,1]^2$ be the unit square. "Partition" $S$ naturally into four congruent squares $S_{1,j}$ (with side length $1/2$ each), where $j=1,\dots,4$; the quotation marks are used here because the $S_{1,j}$'s will have some common boundary points. Next, "partition" each $S_{1,j}$ naturally into four congruent squares (with side length $1/2^2$ each), so that we get $4^2$ squares $S_{2,j}$ for $j=1,\dots,4^2$. Continue doing so, so that at the $k$th step we get $4^k$ squares $S_{k,j}$ for $j=1,\dots,4^k$, for each $k=1,2,\dots$.

Take any subset $A$ of $S$. For each $k=1,2,\dots$ and each $j=1,\dots,4^k$, let $$A_{k,j}:=(A\cap S_{k,j})-s_{k,j},$$ where $s_{k,j}$ is the southwest vertex of the square $S_{k,j}$, so that $A_{k,j}\subseteq S_k:=2^{-k}S$.

Suppose that for each $k$ we have a "measure" $D_k$ of dissimilarity for subsets of $S_k$, so that for any two subsets $B$ and $C$ of $S_k$ we have a nonnegative real number $D_k(B,C)$, which is the greater the more "dissimilar" $B$ and $C$ are (and, say, is $0$ if $B=C$); here the term "measure" is used in the general sense, not necessarily in the sense of measure theory. For instance, $D_k(B,C)$ may depend on the Hausdorff distance between $B$ and $C$ or on some "measure" of the symmetric difference of the sets $B$ and $C$ or on some combination thereof.

Then the distance of the set $A$ from uniformity can be defined by the formula $$D(A):=\sum_{k=1}^\infty\frac1{L^k}\sum_{j=1}^{4^k}\sum_{m=1}^{4^k} \frac{D_k(A_{k,j},A_{k,m})}{1+D_k(A_{k,j},A_{k,m})},$$ where $L$ is a real number $>16$ (to ensure the convergence of the series). Then $D(A)$ will be small if, for "most" levels $k$ of "zooming", "most" of the intersections of the set $A$ with all the "$k$-level" small squares $S_{k,j}$ "look similar" to one another. (Of course, $D(A)$ will depend on the choices of $L$ and the dissimilarity "measures" $D_k$.)

For instance, for any $L$ and any $D_k$'s we have $D(S)=0$ -- of course, the unit square $S$ is at distance $0$ from uniformity (in itself).

As another example, for the uniform grid $G_n$ (defined in the previous answer) with $n=2^K$ for a natural $K$, any real $L>16$, and any $D_k$'s we have $$D(G_n)\le \sum_{k=K+1}^\infty\frac1{L^k}\,16^k =Cn^{-p}\to0$$ as $n=2^K\to\infty$, where $C:=\dfrac{16}{L-16}$ and $p:=\log_2\dfrac L{16}$. So, we see that $G_n$ is close to uniformity for large $n$, even though the Hausdorff dimension of $G_n$ is $0$ for all $n$ (in big contrast with the Hausdorff dimension of $S$, which is $2$). Thus, again we see that the Hausdorff dimension can hardly have anything to do with the idea of uniformity.

Here is another possible approach, perhaps closer to what the OP had in mind.

Let $S:=[0,1]^2$ be the unit square. "Partition" $S$ naturally into four congruent squares $S_{1,j}$ (with side length $1/2$ each), where $j=1,\dots,4$; the quotation marks are used here because the $S_{1,j}$'s will have some common boundary points. Next, "partition" each $S_{1,j}$ naturally into four congruent squares (with side length $1/2^2$ each), so that we get $4^2$ squares $S_{2,j}$ for $j=1,\dots,4^2$. Continue doing so, so that at the $k$th step we get $4^k$ squares $S_{k,j}$ for $j=1,\dots,4^k$, for each $k=1,2,\dots$.

Take any subset $A$ of $S$. For each $k=1,2,\dots$ and each $j=1,\dots,4^k$, let $$A_{k,j}:=(A\cap S_{k,j})-s_{k,j},$$ where $s_{k,j}$ is the southwest vertex of the square $S_{k,j}$, so that $A_{k,j}\subseteq S_k:=2^{-k}S$.

Suppose that for each $k$ we have a "measure" $D_k$ of dissimilarity for subsets of $S_k$, so that for any two subsets $B$ and $C$ of $S_k$ we have a nonnegative real number $D_k(B,C)$, which is the greater the more "dissimilar" $B$ and $C$ are (and, say, is $0$ if $B=C$); here the term "measure" is used in the general sense, not necessarily in the sense of measure theory. For instance, $D_k(B,C)$ may depend on the Hausdorff distance between $B$ and $C$ or on some "measure" of the symmetric difference of the sets $B$ and $C$ or on some combination thereof.

Then the distance of the set $A$ from uniformity can be defined by the formula $$D(A):=\sum_{k=1}^\infty\frac1{L^k}\sum_{j=1}^{4^k}\sum_{m=1}^{4^k} \frac{D_k(A_{k,j},A_{k,m})}{1+D_k(A_{k,j},A_{k,m})},$$ where $L$ is a real number $>16$ (to ensure the convergence of the series). Then $D(A)$ will be small if, for "most" levels $k$ of "zooming", "most" of the intersections of the set $A$ with all the "$k$-level" small squares $S_{k,j}$ "look similar" to one another. (Of course, $D(A)$ will depend on the choices of $L$ and the dissimilarity "measures" $D_k$.)

Here is another possible approach, perhaps closer to what the OP had in mind.

Let $S:=[0,1]^2$ be the unit square. "Partition" $S$ naturally into four congruent squares $S_{1,j}$ (with side length $1/2$ each), where $j=1,\dots,4$; the quotation marks are used here because the $S_{1,j}$'s will have some common boundary points. Next, "partition" each $S_{1,j}$ naturally into four congruent squares (with side length $1/2^2$ each), so that we get $4^2$ squares $S_{2,j}$ for $j=1,\dots,4^2$. Continue doing so, so that at the $k$th step we get $4^k$ squares $S_{k,j}$ for $j=1,\dots,4^k$, for each $k=1,2,\dots$.

Take any subset $A$ of $S$. For each $k=1,2,\dots$ and each $j=1,\dots,4^k$, let $$A_{k,j}:=(A\cap S_{k,j})-s_{k,j},$$ where $s_{k,j}$ is the southwest vertex of the square $S_{k,j}$, so that $A_{k,j}\subseteq S_k:=2^{-k}S$.

Suppose that for each $k$ we have a "measure" $D_k$ of dissimilarity for subsets of $S_k$, so that for any two subsets $B$ and $C$ of $S_k$ we have a nonnegative real number $D_k(B,C)$, which is the greater the more "dissimilar" $B$ and $C$ are (and is $0$ if $B=C$); here the term "measure" is used in the general sense, not necessarily in the sense of measure theory. For instance, $D_k(B,C)$ may depend on the Hausdorff distance between $B$ and $C$ or on some "measure" of the symmetric difference of the sets $B$ and $C$ or on some combination thereof.

Then the distance of the set $A$ from uniformity can be defined by the formula $$D(A):=\sum_{k=1}^\infty\frac1{L^k}\sum_{j=1}^{4^k}\sum_{m=1}^{4^k} \frac{D_k(A_{k,j},A_{k,m})}{1+D_k(A_{k,j},A_{k,m})},$$ where $L$ is a real number $>16$ (to ensure the convergence of the series). Then $D(A)$ will be small if, for "most" levels $k$ of "zooming", "most" of the intersections of the set $A$ with all the "$k$-level" small squares $S_{k,j}$ "look similar" to one another. (Of course, $D(A)$ will depend on the choices of $L$ and the dissimilarity "measures" $D_k$.)

For instance, for any $L$ and any $D_k$'s we have $D(S)=0$ -- of course, the unit square $S$ is at distance $0$ from uniformity (in itself).

As another example, for the uniform grid $G_n$ (defined in the previous answer) with $n=2^K$ for a natural $K$, any real $L>16$, and any $D_k$'s we have $$D(G_n)\le \sum_{k=K+1}^\infty\frac1{L^k}\,16^k =Cn^{-p}\to0$$ as $n=2^K\to\infty$, where $C:=\dfrac{16}{L-16}$ and $p:=\log_2\dfrac L{16}$. So, we see that $G_n$ is close to uniformity for large $n$, even though the Hausdorff dimension of $G_n$ is $0$ for all $n$ (in big contrast with the Hausdorff dimension of $S$, which is $2$). Thus, again we see that the Hausdorff dimension can hardly have anything to do with the idea of uniformity.

Source Link
Iosif Pinelis
  • 127.7k
  • 8
  • 107
  • 229

Here is another possible approach, perhaps closer to what the OP had in mind.

Let $S:=[0,1]^2$ be the unit square. "Partition" $S$ naturally into four congruent squares $S_{1,j}$ (with side length $1/2$ each), where $j=1,\dots,4$; the quotation marks are used here because the $S_{1,j}$'s will have some common boundary points. Next, "partition" each $S_{1,j}$ naturally into four congruent squares (with side length $1/2^2$ each), so that we get $4^2$ squares $S_{2,j}$ for $j=1,\dots,4^2$. Continue doing so, so that at the $k$th step we get $4^k$ squares $S_{k,j}$ for $j=1,\dots,4^k$, for each $k=1,2,\dots$.

Take any subset $A$ of $S$. For each $k=1,2,\dots$ and each $j=1,\dots,4^k$, let $$A_{k,j}:=(A\cap S_{k,j})-s_{k,j},$$ where $s_{k,j}$ is the southwest vertex of the square $S_{k,j}$, so that $A_{k,j}\subseteq S_k:=2^{-k}S$.

Suppose that for each $k$ we have a "measure" $D_k$ of dissimilarity for subsets of $S_k$, so that for any two subsets $B$ and $C$ of $S_k$ we have a nonnegative real number $D_k(B,C)$, which is the greater the more "dissimilar" $B$ and $C$ are (and, say, is $0$ if $B=C$); here the term "measure" is used in the general sense, not necessarily in the sense of measure theory. For instance, $D_k(B,C)$ may depend on the Hausdorff distance between $B$ and $C$ or on some "measure" of the symmetric difference of the sets $B$ and $C$ or on some combination thereof.

Then the distance of the set $A$ from uniformity can be defined by the formula $$D(A):=\sum_{k=1}^\infty\frac1{L^k}\sum_{j=1}^{4^k}\sum_{m=1}^{4^k} \frac{D_k(A_{k,j},A_{k,m})}{1+D_k(A_{k,j},A_{k,m})},$$ where $L$ is a real number $>16$ (to ensure the convergence of the series). Then $D(A)$ will be small if, for "most" levels $k$ of "zooming", "most" of the intersections of the set $A$ with all the "$k$-level" small squares $S_{k,j}$ "look similar" to one another. (Of course, $D(A)$ will depend on the choices of $L$ and the dissimilarity "measures" $D_k$.)