\documentclass%[12pt] {amsart} \def\heute{1998-03-04} \usepackage{amsbsy} \usepackage{amssymb} \usepackage{amsmath} \title[Distribution of Subsequences] {Metric, fractal dimensional and Baire Results \\ on the \\ Distribution of Sequences and Subsequences} \author{Martin Goldstern} \email{Martin.Goldstern@tuwien.ac.at} \thanks{The first author was supported by an Erwin Schr\"odinger Fellowship from the Austrian Science Foundation (FWF)} \urladdr{http://info.tuwien.ac.at/goldstern/} \author{J\"org Schmeling} \email{shmeling@math.fu-berlin.de} \thanks{The second author was partially supported by the Leopoldina Forschungspreis} \author{Reinhard Winkler} \email{Reinhard.Winkler@oeaw.ac.at} \address{ % 2nd address for Winkler and current address for Goldstern Department of Algebra\\ Wiedner Hauptstr.~8-10/118\\ Technische Universit\"at\\ A-1040 Wien\\ Austria, Europe\\ } \address{ % address for Goldstern and Schmeling Fachbereich Mathematik \\ Freie Universit\"at Berlin\\ Arnimallee 3\\ D-14195 Berlin Germany} \address{ % first address for Winkler \"Osterreichische Akademie der Wissenschaften\\ Ignaz Seipel-Platz 2\\ A-1010 Wien\\ Austria, Europe } \subjclass{Primary 11K55; Secondary 28A33} %11Kxx Probabilistic theory: distribution modulo $1$; metric theory % %11K55 Metric theory of other algorithms and expansions; measure % and Hausdorff dimension, See also {11N99, 28Dxx} % %11K99 None of the above but in this section % %28A33 Spaces of measures, convergence of measures, See Also { % 46E27, 60Bxx} % % %60Bxx Probability theory on algebraic and topological structures %60B10 Convergence of probability measures % %28A78 Hausdorff measures \date{\heute} % \setlength{\textwidth}{140 mm} % \setlength{\textheight}{200 mm} \sloppy % % cases environment scheint es hier nicht zu geben? %\let\oldcases\cases \newcommand{\cJ}{{\cal J}} \newcommand{\on}{{\upharpoonright}} %\newenvironment{mycases}{\oldcases\bgroup\let\\\cr}{\egroup} %\let\cases\mycases %\let\endcases\endmycases %\newcommand{\text}{\mbox} \def\cases#1{ \left \{\,\vcenter {\normalbaselines \ialign {$##\hfil $&\quad ##\hfil \crcr #1\crcr }}\right .} \def\ka{j} \def\newsomething#1 {\newtheorem{#1}[Theo]{#1}} \theoremstyle{plain} \newtheorem{Theo}{Theorem} \newtheorem{Prop}[Theo]{Proposition} \newtheorem{Cor}[Theo]{Corollary} \newsomething Theorem \newsomething Corollary \newsomething Claim \newsomething Lemma \newsomething Fact \theoremstyle{definition} \newtheorem{Def}[Theo]{Definition} \newsomething Definition \theoremstyle{remark} \newsomething Remark \newsomething Examples \newsomething {Elementary examples} \newsomething{Open Problem} %\newcommand{\emph}[1]{{\em #1}} % \newcommand{\N}{{\mathbb N}} \newcommand{\R}{{\mathbb R}} \newcommand{\Z}{{\mathbb Z}} \newcommand{\xf}{{\bf x}} \newcommand{\yf}{{\bf y}} \newcommand{\zf}{{\bf z}} \newcommand{\ff}{{\bf f}} \newcommand{\tm}{\subseteq} \newcommand{\kn}{(k_n)_{n \in \N}} \newcommand{\xn}{(x_n)_{n \in \N}} \newcommand{\yn}{(y_n)_{n \in \N}} \newcommand{\an}{(a_n)_{n \in \N}} \newcommand{\bn}{(b_n)_{n \in \N}} \newcommand{\cn}{(c_n)_{n \in \N}} \newcommand{\qn}{(q_n)_{n \in \N}} \newcommand{\e}{\varepsilon} % \newcommand{\proof}{\par\noindent{\bf Proof: }} % \newcommand{\qed}{{\bf q.e.d.}} \newcommand{\ud}{\underline{d}} % \newcommand{\sup}{\text{sup}} \newcommand{\dimh}{\dim_{\rm H}} \newcommand{\impl}{\Longrightarrow} \newcommand{\implz}{\Longleftarrow} \newcommand{\ifff}{\Longleftrightarrow} \newcommand{\M}{{\cal M}} \newcommand{\K}{{\cal K}} \newcommand{\Mal}{\mbox{Mal}} \newcommand{\supp}{\mbox{supp}} \newcommand{\card}{{\rm card}} \newcommand{\limn}{\lim_{n \to \infty}} \newcommand{\limN}{\lim_{N \to \infty}} \newcommand{\forces}{\mathrel{\mathrel\|\joinrel\mathrel-}} \newcommand{\C}{{\cal C}} \newcommand{\hhh}{\par\centerline{\vrule height 0.5pt width 0.5\hsize}\par} \newcommand{\binomial}{\binom} \let\oldcal\cal % \renewcommand{\cal}[1]{{\oldcal #1}} \newcommand{\cal}[1]{{\mathcal{#1}}} \newcommand{\two}{\genfrac{}{}{0pt}{}} % \advance\textheight3cm % \advance\textwidth3cm % \advance\voffset-2cm % \advance\hoffset-2cm \begin{document} \begin{abstract} Let $X$ be a locally compact metric space. One important object connected with the distribution behavior of an arbitrary sequence $\xf$ on~$X$ is the set $M(\xf)$ of limit measures of~$\xf$. It is defined as the set of accumulation points of the sequence of the discrete measures induced by~$\xf$. Using binary representation of reals one gets a natural bijective correspondence between infinite subsets of the set $\N$ of positive integers and numbers in the unit interval $I = [0,1)$. Hence to each sequence $\xf = \xn \in X^{\N}$ and every $a \in I$ there corresponds a subsequence denoted by~$a\xf$. We investigate the set $M(a\xf)$ for given $\xf$ with emphasis on the behavior for ``typical'' $a$ in the sense of Baire category, Lebesgue measure and Hausdorff dimension. We also treat the special case of the sequence $(n \alpha) $ mod 1 and its subsequences. \end{abstract} \maketitle % % \include{not} \section{Introduction}\label{6} % \pagestyle{headings} % \markboth{\heute}{\heute} \subsection*{Notations} For the whole paper we fix the following abbreviations. Let $\N$ denote the set of positive integers $1,2,\ldots$ and let $\R/\Z$ be the one dimensional torus group, identified with the half open unit interval $I=[0,1)$ in the natural way. That is, we identify the remainder class $x+\Z$ with the real number $x - [x]$ (where $[x] $ denotes the the largest integer $k \in \Z$ with $k \le x$). So the canonical epimorphism $\R \to \R/\Z $ can be identified with the map $x \mapsto x - [x] $. For any set $X $ we write $X^\N $ for the set of all sequences $\xf=\xn $, $x_n\in X $. Sequences (both infinite and finite) are always denoted by boldface letters. If $\xf=\xn$ is an infinite sequence, we write $\xf\on N $ for the finite sequence $(x_1,\ldots, x_N) $. For a finite sequence $(\varepsilon_1, \ldots, \varepsilon_n) $ of elements of~$X $ we write $[\varepsilon_1,\ldots, \varepsilon_n] $ for the ``cylinder set'' \[ \{ \xn\in X^\N: x_1=\varepsilon_1, \ldots, x_n=\varepsilon_n\}. \] If $X $ is discrete (in particular if $X=\{0,1\}$), then we may write $C_n(\xf)$ for $[\varepsilon_1, \ldots, \varepsilon_n]$. In this case the family $(C_n(\xf))_{n\in\N}$ is a local basis at~$\xf$ in the product topology on~$X^\N$. Each number $a \in I$ has a representation $a = \sum_{n \in \N} a_n2^{-n}$ with $a_n \in \{0,1\}$. This representation is unique if one requires that the set $S_a = \{n \in \N \ | \ a_n=1\}$ is infinite. (Note that to~$0 \in I$, by identification, there corresponds the set $S_0 = S_1 = \N$.) Thus the correspondence $a \mapsto S_a$ is a bijection between $I$ and the system of all infinite subsets of~$\N$ or, equivalently, the set of all strictly increasing infinite sequences of~integers. If $S_a = \{k_1 < k_2 < \ldots\}$ and $\xf= \xn $, then we write $a \xf$ for the sequence $(x_{k_n})_{n \in \N} = (x_i)_{i \in S_a}$. If $\xf = (x_1, \ldots, x_n)$ is a finite sequence, then $a\xf$ will denote the subsequence $(x_i)_{i \in \{1,\ldots, n\} \cap S_a}$. We usually use the Euclidean metric and topology on $[0,1)$, in particular when we consider measure and Hausdorff dimension. For arguments dealing with Baire category, however, we identify $[0,1)$ with a subspace of $\N^\N$ and use the Tychonoff topology, which we call ``the $0$-dimensional topology.'' Note that the notions ``meager'' and ``residual'' are the same in the $0$-dimensional topology and the Euclidean topology. % For a given sequence $a$ we denote by~$C_n(A)$ the cylinder set % \[ % \{b \in [0,1] \ | \ b_1=a_1 \cdots b_n=a_n \}. % \] Each $p \in I$ induces the $p$-Bernoulli measure $\lambda_p$ on~$I$ which is the Borel measure uniquely defined by the requirement $\lambda_p(\{a \ | \ a_n=1\})=p$ for all $n \in \N$. If $p \in \{0,1\}$, $\lambda_p$ is concentrated on~$0=1 \in I$. Hence the case $0 < p < 1$ is the interesting one and will be assumed for the rest of the paper. Let $X$ be a locally compact Hausdorff space with countable topological base. Thus there is a complete metric $d_X$ inducing the topology. With $A(\xf)$ we denote the set of accumulation points of~$\xf$. This is the set of all $y \in X$ such that in each neighborhood $O$ of~$y$ the relation $x_n \in O$ holds for infinitely many $n \in \N$. Let now $\M(X)$ denote the set of all nonnegative Borel measures $\mu$ on~$X$ with $\mu(X)=1$ if $X$ is compact and $\mu(X)\le 1$ if $X$ is not compact. Let $\K(X)$ denote the space of all continuous $f: X \to \R$ with compact support \[ \supp(f) = \overline{\{x \in X \ | \ f(x) \neq 0 \}}. \] We consider the weak topology on~$\M(X)$ which is the weakest topology such that all functionals $\mu \mapsto \int_X f d\mu$, $f \in \K(X)$, are continuous. $\M(X)$ is a compact space with countable topological base, hence also induced by a metric $d_{\M(X)}$. In most cases it will cause no confusion to write $d$ for both $d_X$ and~$d_{\M(X)}$. We will need the fact that $d_{\M(X)}$ can be chosen to be compatible with convex combinations, i.e., \[ d_{\M(X)}(\mu_0, (1-t)\mu_0 + t \mu_1) = td_{\M(X)}(\mu_0,\mu_1) \] for $0 \le t \le 1$. (One possibility of defining $d$ is the following: Let $f_n:\R\to [0,1]$, $n \in \N$, be a countable family spanning a dense subspace of~$\K(X)$. Then the function $d:\M(X)^2 \to \R$, defined by \[ d(\mu_1, \mu_2) = \sum_{n \in \N} 2^{-n} \Bigl| \int f_n\, d\mu_1 - \int f_n \, d\mu_2\Bigr| \] satisfies the above requirements.) Each finite sequence $\xf = (x_1,\ldots, x_N)$ induces a probability measure $\mu_\xf = \frac{1}{N} \sum_{n=1}^N \delta_{x_n}$, where $\delta_x$ is the Dirac measure concentrate at~$x$. For an infinite sequence $\xf=\xn$ we let \[ \mu_{N,\xf} = \mu_{\xf\on N},\qquad \mu_{N,\xf}(B) = \frac{1}{N}\#\{i\le N: x_i\in B\} \] ($\# A$ denotes the cardinality of a set~$A$.) With $M(\xf)$ we denote the set of all accumulation points of the sequence $(\mu_{N,\xf})_{N \in \N}$ in~$\M(X)$. It follows from \cite{Si} (for the compact case) or \cite{W2} that for an arbitrary subset $M \tm \M(X)$ the following holds: There is a sequence $\xf$ with $M = M(\xf)$ if and only if $M$ is nonempty, closed and connected. $\xf$ is called uniformly distributed w.r.t.\ $\mu$ ($\mu$-u.d.) if $M(\xf) = \{ \mu \}$. If $X$ is a compact group and $\mu$ is the Haar measure on~$X$, we call $\xf$ simply u.d. In the other extreme case $M(\xf) = \M(X)$, the sequence $\xf$ is called maldistributed (md.). Thus we use the notation from \cite{W2} which is a slight modification of the one in~\cite{My}. If $A \tm X$ is a closed subset of~$X$, then the set $\M(A)$ can be considered as a subspace of~$\M(X)$ in an obvious way: Identify every $\mu\in \M(A)$ with its extension $\bar\mu\in \M(X)$, defined by~$\bar \mu(B) = \mu(B \cap A)$ for arbitrary Borel sets $B \tm X$. Recall that a subset of a topological space is called nowhere dense if its topological closure has empty interior. A set is of first category or meager if it is the union of at most countably many nowhere dense sets. All sets which are not meager are said to be of second category. A set is called co-meager or residual if its complement is meager. Complete metric spaces (as $I=\R/\Z$, $X$ and $\M(X)$) are Baire spaces, i.e., nonempty open subsets are of second category. \subsection*{Content of the paper} In section~\ref{5} we investigate the set $M(a\xf)$ for given $\xf$ with emphasis on the behavior for ``typical'' $a$ in the sense of Baire category, Lebesgue measure and Hausdorff dimension. In section~\ref{7} we consider sequences depending on an additional parameter. In section~\ref{8} we focus on the special case of $n\alpha$-sequences. We give Baire results on the typical distribution of~$(k_n \alpha )_{n\in \N}$ depending on the growth rate of $(k_n)_{n\in \N}$. For special sequences $(k_n)_{n\in \N}$ cf.\ also \cite{Be}. The reader who is interested in connections to ergodic theory may want to consult the survey paper~\cite{R-W}. The standard textbook on uniform distribution is~\cite{K-N}. Many interesting results and an up-to-date bibliography can found in \cite{D-T}. % % \include{sub} \section{Subsequences} \label{5} We fix the general assumption that $X$ be a locally compact and second countable (hence complete metric) space. Given a sequence $\xf$ on~$X$ we look, in dependence on~$a$, at its subsequences $a \xf$ and the corresponding sets $M(a\xf)$. The natural questions on this sets are: \begin{enumerate} \item Which $a$ guarantee $M(a\xf)=M(\xf)$ for arbitrary $\xf$? (density 1, theorem \ref{9}.) \item What can $M(a\xf)$ look like? (Anything reasonable, theorem \ref{1}.) \item What is the typical situation in the Baire sense? (Maldistribution, theorem \ref{2}.) \item What is the typical situation in the Lebesgue sense? ($M(a\xf)=M(x)$, theorem \ref{3}.) \item What is the (Hausdorff) dimension of all $a$ with $M(a \xf) \neq M(\xf)$? (Dimension 1, theorem \ref{4}.) \end{enumerate} For theorem \ref{9} we recall the notion of~density. For $S \tm \N$ and $n \in \N$ one considers the numbers $d(n,S) = \frac 1n \# S \cap (0,n]$. Since $0 \le d(n,S) \le 1$ the sequence $(d(n,S))_{n\in\N}$ has an upper and lower limit, the ``upper density $\overline d(S)$'' and the ``lower density $\underline d(S)$.'' If they coincide, their common value is called the density of~$S$, $d(S)$. \begin{Theo} \label{9}\ \begin{itemize} \item[a)] If $S_a$ has density $1$ then $M(a\xf)=M(\xf)$. \item[b)] If $\overline d(S) < 1$, and $\yf$ is an arbitrary sequence on~$X$, $\# X \ge 2$, then there exists a sequence $\xf$ such that $a\xf=\yf$ and $M(a\xf)= M(\yf) \not= M(\xf)$. \end{itemize} \end{Theo} \begin{proof} a): Use the fact that $ d( n, S_a) \ge 1-\varepsilon$ implies $|\mu_{N,\xf}(B) - \mu_{N,a\xf}(B) | \le 2 \varepsilon $ for any set~$B$. \\ b): First case: $M(\yf)=\M(X)$, i.e., $\yf$ is maldistributed. Take an arbitrary $x \in X$ and define $x_n=x$ for all $n \notin S_a$ and $x_{a_n}=y_n$ for $a_n \in S_a$ such that $a\xf=\yf$. Since $\# X \ge 2$ there exists a $\mu \in \M(X)$ with $\mu(\{x\})=0$. Note that \[ \liminf \mu_{N,\xf}(\{x\}) \ge 1 - \overline d > 0. \] Thus $\mu \notin M(\xf)$ proving $M(\xf) \neq \M(X) = M(\yf) = M(a\xf)$. Second case: If there is a $\mu_0 \in \M(X) \setminus M(\yf)$ consider $\N \setminus S_a = \{b_1 < b_2 < \ldots \}$. Let $\zf$ be $\mu_0$-u.d.\ and define $x_{b_n}=z_n$ and~$x_{a_n}=y_n$. $M(a\xf)$ is a nonempty compact set. Hence there is a $\mu_1 \in M(a\xf)$ such that $\e = d(\mu_0, \mu_1) > 0$ takes a the minimal value. It follows that there is a convex combination $\mu_t = t\mu_0 + (1-t)\mu_1$, $0 < t < 1$, such that $\mu_t \in M(\xf)$. Here we used $\overline d < 1$. Since $d(\mu_0, \mu_t) < d(\mu_0,\mu_1)$ we conclude $\mu_t \in M(\xf) \setminus M(a \xf)$. \end{proof} If lower and upper density $\underline d$ and $\overline d$ of~$S_a$ satisfy $\underline d < \overline d = 1$ general assertions as in theorem \ref{9} fail to be true. To illustrate the situation look at the following examples where we use the notations of theorem \ref{9}: If $\zf$ is $\mu_0$-u.d.\ and $\yf$ is $\mu_1$-u.d.\, $\mu_0 \neq \mu_1$, then in the second case of part b) one gets $M(\xf) = \{ \mu_t \ | \ t \in [\underline d,\overline d] \}$ with $\e = \underline d > 0$. Thus $M(\xf) = M(a\zf)$. On the other hand, for any $a$ with $\overline d(S_a)=1$, it is possible to construct a maldistributed sequence $\yf$ such that $a\xf=\yf$ implies $M\xf)=M(\yf)$. A more detailed analysis will be the object of a forthcoming paper. \begin{Theo}\label{1} Let $\xf \in X^{\N}$ and $M \tm \M(X)$. There exists an~$a \in I$ with $M(a\xf)=M$ if and only if $M$ is closed and connected with $\emptyset \neq M \tm \M(A(\xf))$. \end{Theo} \begin{proof} First assume $M=M(a\xf)$. Then, by theorem 3.1.1 in \cite{W2}, $M$ is nonempty, closed and connected. To derive the remaining relation $M \tm \M(A(\xf))$ it suffices to show that every $x \in X \setminus A(\xf)$ has a neighborhood $U$ with $\limN \mu_{N,a\xf}(U) = 0$. To show this take any neighborhood $U$ with compact closure $\overline U$ and with $\overline U \cap A(\xf) = \emptyset$. Such a $U$ exists since $X$ is a normal space. If $x_n \in \overline U$ for infinitely many~$n$, $\overline U$ would, by compactness, contain an accumulation point of~$\xf$, contradiction. Thus $x_n \notin U$ for all $n \ge N_0$. This proves \[ \limN \mu_{N,a\xf}(U) \le \limN \frac{N_0}N = 0. \] For the other direction assume that $M\not=\emptyset $ is closed and connected. By \cite[theorem 3.1.2]{W2}, there is sequence $\zf=\yn$ in $A(\xf)$ with $M(\zf)=M$. By the definition of $A(\xf)$ there is a subsequence $a\xf$ of $\xf$ with $a=(a_1,a_2,\ldots)$ such that $\limn d(y_n,x_{a_n}) = 0$. This subsequence, obviously, has the required properties. \end{proof} \begin{Theo}\label{2} $M(a\xf) = \M(A(\xf))$ holds for all $a \in R$ from a residual set $R \tm I$, i.e., most subsequences are maldistributed in~$A(\xf)$. \end{Theo} \proof Let $R$ denote the set of all $a$ with $M(a\xf) = \M(A(\xf))$, i.e., $a \xf$ is maldistributed in $A(\xf)$: \[ R = \{ a: M(a\xf) = \M(A(\xf)) \} \] The inclusion $\subseteq $ follows from theorem \ref{1}. For the other one we have to prove that for every $\mu \in \M(A(\xf))$ the set \[R_{\mu} = \{ a \in I: \mu \in M(a\xf) \} \] is residual. Since $\M(X)$ is second countable, one may choose a sequence $(\mu_n)_{n \in \N}$ which is dense in~$\M(A(\xf))$. We now check that \[ R = \bigcap_{n \in \N} R_{\mu_n} \] is a countable intersection of residual sets, hence residual. In the above equation the inclusion $\tm$ is trivial. But also the converse inclusion is clear: Let $a$ be an element from the right hand side, then the set $M(a\xf)$ contains the topological closure of the set of all $\mu_n$ which is $\M(A(\xf))$, hence $a \in R$. To show that $R_{\mu}$ is residual for $\mu \in \M(A(\xf))$, just observe \[ R_{\mu} = \bigcap_{n,k \in \N} O_{n,k} \] with the sets $O_{n,k}$ of all $a \in I$ such that there exists an $n' > n$ with $d(\mu_{n'a\xf},\mu) < \frac 1k$. Standard arguments show that each $O_{n,k}$ is dense and open. [$O_{n,k}$ is open since the map $a \mapsto d(\mu_{n,a\xf}, \mu)$ is continuous (with respect to the $0$-dimensional topology on~$I$). $O_{n,k}$ is dense: Take any $\mu$-u.d.\ $\yf = \yn$, $y_n\in A(\xf)$. Any subsequence $(x_{n_k})_{k\in \N}$ with $d(y_k, x_{n_k}) \to 0$ can be written as $a\xf$ with $a\in O_{n,k}$.] Thus the countable intersection $R_{\mu}$ of these sets is residual, and the theorem is proved. \qed \begin{Remark} A general principle containing all Baire arguments of the above type can be found in~\cite{Gr}. \end{Remark} % The following theorem is, in some sense, a generalization of results in % \cite{L-T} an other papers. \begin{Theo}\label{3} $M(a \xf) = M(\xf)$ for $\lambda_p$-almost all~$a$, $0 \tau(\delta)$ the $s$-dimensional Hausdorff measure of $A_0$ can be estimated by %\begin{align*} %\inf&\left\{\sum_{C_i\in \cal{C}^n} (diam C_i)^s : \cup_i C_i \supset A_{0} ; % \cal{C}^n\subset\cup_{m\ge n}\cal{C}_m \right\}\\& \le \sum_{n=1}^{\infty} % 2^{\tau(\delta) n +1} 2^{- s n}< \infty . %\end{align*} \[ \lim_{N \to \infty} \sum_{n=N}^{\infty} n 2^{-n\tau(\delta) } (2^{- n})^2 = 0 \] Hence $\dim_H(A_0) \le \tau(\delta)$. This completes the proof of part 1. \emph{\bf 2. ($\rho(M)=1 \impl \phi(M)=1$)} By theorem~\ref{9} we have that $M=M(\xf)$. Then theorem~\ref{3} implies that the set $ \{ a \ | \ M(a\xf)=M(\xf) \}$ has full Lebesgue measure and hence Hausdorff dimension 1. \emph{\bf 3.($\phi(M)=1 \impl \rho(M)=1$)} Let us assume that $\rho(M)<1$. In view of the definition of $\rho(M)$ there is a measure $\mu \in M(\xf)\triangle M$. Let $ \varepsilon = \frac{1}{10} d(\mu, M)$ if $\mu \in M(\xf)\setminus M$, and $ \varepsilon = \frac{1}{10} d(\mu, M(\xf))$ if $\mu \in M\setminus M(\xf)$. To estimate the Hausdorff dimension of \[ A_0 := \{ a: M(a\xf)=M\} \] we first claim that $a \in A_0$ implies that there are infinitely many $N\in \N$ such that $ d(\mu_{\xf\on N}, \mu_{a(\xf\on N)}) \le 8\varepsilon$. To check this claim we assume without loss of generality $\mu \in M (\xf ) \setminus M$ (the other case is similar). There are infinitely many $N$ such that $d(\mu, \mu_{\xf\on N}) \le \varepsilon $, and for all large enough $N$ we have $d(\mu_{a(\xf\on N)}, M(a\xf)) \le \varepsilon$, hence $d(\mu_{\xf\on N}, \mu_{a(\xf\on N)}) \ge 8 \varepsilon $. Hence we can find finitely many functions $f_1, \ldots, f_k$ such that the set $A_0$ is contained in \[ A_1:=\bigcup_{j=1}^k \{ a: \exists^\infty N\,\, |s(f_j,N) - s(f_j, a, N)| \ge 4\varepsilon \} \] It is therefore enough to show that for any continuous $f$ and any~$\varepsilon >0$ the set $ \{ a: \exists^\infty N\,\, \bigl| s(f,N) -s(f,N,a) \bigr| \ge 4\varepsilon \}$ has Hausdorff dimension~$<1$. So fix $f$ and $\varepsilon $. Let $m> 1/\varepsilon$, and define $T_1, \ldots, T_m$ as in the proof of~theorem~\ref{3}. Write $A_{N,{\varepsilon}}$ for the set \[ A_{N,{\varepsilon}} := \{a: |s(f,N) - s(f,N,a) | \ge {4\varepsilon} \} \] Note that $a\in A_{N,{\varepsilon}}$ implies that there is some $i\in \{1,\ldots, m\}$ such that \def\bigdiff#1{ \Bigl| \frac1N \sum_{\two{j\in T_i}{j\le N}} #1 \,\, - \,\, \frac{1}{\#S_a\cap \{1,\ldots, N\}} \sum _{\two{j\in T_i\cap S_a}{j\le N}} #1 \Bigr| } \[ \bigdiff{f(x_j)} \ge 4{\varepsilon} \frac{\#T_i}{N} \] Replacing $f(x_j)$ by $\frac {i}{m}$ introduces an error $\le\frac {1}{m}< \varepsilon $ in each term, hence \[ \bigdiff{\frac{i}{m}} \ge 3 \varepsilon \frac{\#T_i}{N} \] Employing the functions $ \gamma $ and $ \delta$ that we defined in the proof of theorem~\ref{3} and using $\frac{i}{m}\le 1$ we conclude that $a \in A_{N, \varepsilon}$ implies \[ 1 - \delta(N,a)\cdot \gamma(i,N,a) \ge3 \varepsilon \qquad \mbox{ for some $i$} \] so $ \delta(N,a) \le 2- \varepsilon $ or $\gamma(i,N,a) \ge \frac 12 - \varepsilon $. Hence \[ A_1 = \bigcap_{n=1}^\infty \bigcup_{N=n}^\infty A_{N,\varepsilon } \subseteq \bigcap_{n=1}^\infty \bigcup_{N=n}^\infty \{a: \delta(N,a)<2-\varepsilon \} \cup \bigcap_{n=1}^\infty \bigcup_{N=n}^\infty \{a: \gamma(i,N,a)< \frac12 -\varepsilon \} \] We will only estimate the Hausdorff dimension of the sets \[ \bigcap_{n=1}^\infty \bigcup_{N=n}^\infty \{ a: \gamma(i,N,a) \le \frac12 - {\varepsilon} \} % \qquad % \mbox{ and } % \qquad % \bigcap_{n=1}^\infty \bigcup_{k=n}^\infty % \{ a: \delta (N,a) \le 2- {4\varepsilon} \} \] leaving the second (similar) estimate to the reader. As in the first step of the proof of theorem~\ref{3}, let \[ \C_N({\varepsilon}):= \{ [\varepsilon _1,\ldots, \varepsilon_n]: \frac{1}{N_i} \sum_{\two{ j \le n}{ j \in T_i}} \varepsilon_j \le \frac12 - {\varepsilon}\} \qquad \mbox{(where $N_i:= \#(T_i \cap [1,\ldots,N])$),} \] and let $C_N(\varepsilon)$ be the union of all the cylinders in~$\C_N( \varepsilon)$. Again, $\C_N({\varepsilon})$ contains at most $ \displaystyle \binomial{N_i}{(\frac 12-{\varepsilon}) N_i} = o\bigl(N2^{-N\tau( \frac12 - {\varepsilon})}\bigr)$ many cylinders, % [where $N_i:=\# T_i\cap \{1,\ldots, N\}$] each of them of diameter $2^{-N_i}$. Since $A_1$ is covered by $ \bigcup_{n=N}^\infty C_n(\varepsilon)$ for any $N\in \N$, we can as before bound the the $s$-dimensional Hausdorff measure of $A_1$ for any $s> \tau(\frac12-{\varepsilon})$ by \[ \lim_{N \to \infty} \sum_{n=N}^\infty n2^{n\tau(\frac12-{\varepsilon})}2^{-ns} = 0 \] Hence the Hausdorff dimension of $A_0$ is $\le \tau(\frac12-\varepsilon ) < 1$. % % Hence, there is % a $\varepsilon >0$ such that for $a \in \{ a \ | \ M(a\xf)=M \}$ we can % find a continuous function $f\colon X\to [0,1)$ with % \[ % \limsup_{N \to \infty} \vert s(f,N) - s(f,N,a) \vert > \varepsilon. % \] % Following the proof of theorem~\ref{3} we conclude that for % some $m$ large, a subsequence~$N_k$, $\varepsilon_1 >0$, $\delta >0$ and % some $i$ % \[ % \gamma(i,N_k,a)\delta(N_k,a)\le 1- \varepsilon % \] % and % \[ % \frac{\#(T_i\cap (0,N_k])}{N_k}\ge\delta. % \] % Otherwise, the values of $f$ in the interval $\frac{i-1}{m}$ would be obtained % with the right frequency. % Let % \[ % A = \{a \ | \ \liminf \gamma(i,N_k,a) \le \frac12 - \sqrt{\varepsilon} \} % \] % and % \[ % B = \{a \ | \ \liminf \delta(N_k,a) \le 2 - \sqrt{\varepsilon} \}. % \] % By the above arguments % \begin{equation}\label{AB} % \{ a \ | \ M(a\xf)=M \} \subset A \cup B. % \end{equation} % As in the proof of part 1 we consider special coverings % of $A$ (respectively $B$). % Let ${\cal D}_k$ be the set of cylinders $[\epsilon_1 \cdots % \epsilon_{N_k}]$ such that each cylinder contains at most $(\frac12 % -\sqrt{\varepsilon}) \#(T_i\cap (0,N_k])$ 1's in $T_i\cap (0,N_k]$ and is % arbitrary outside$ T_i\cap (0,N_k]$. There are at % most $2^{\tau(\frac12 -\sqrt{\varepsilon_1}) \delta % N_k} 2^{(1-\delta) N_k}$ of them. Again we can estimate the Hausdorff sum for % $1>s>1-(1-\tau(\frac12 -\sqrt{\varepsilon_1})\delta)$ % \begin{align*} % \inf&\left\{\sum_{D_i\in {\cal D}^k} (diam D_i)^s : \cup_i D_i \supset A ; % {\cal D}^k \subset\cup_{m\ge k} {\cal % D}_m \right\}\\& \le \sum_{k=1}^{\infty} % 2^{\tau(\frac12 -\sqrt{\varepsilon_1})\delta % N_k} 2^{(1-\delta) N_k} 2^{- s n}< \infty . % \end{align*} % This gives part 3. % % % % We start by proving the following claim: % % % % % \begin{Claim} % Fix a sequence $\xf$ and a function~$f$. % Then for any ${\varepsilon}>0$ the Hausdorff dimension of the set % $$ \bigcap_{n=1}^\infty \bigcup_{k=n}^\infty % \{ a: \gamma(i,N,a) \le \frac12 - {\varepsilon} \} % $$ % is at most ????. A similar estimate for the set % $ \{ a: \gamma(i,N,a) \le \frac12 - {\varepsilon} \}$ % is left to the reader. % % \end{Claim} % \proof As in ???, let % \[ % \C_N({\varepsilon}):= \{ [\varepsilon _1,\ldots, \varepsilon_n): % \sum_{\two{ % j \le n}{ j \in T_i}} \varepsilon_j / \# T_i \cap % \{1,\ldots,N\} \le \frac12 - \frac {\varepsilon}4\} % \] % Again, $\C_N({\varepsilon})$ contains at most $ % \displaystyle % \binomial{N_i}{(\frac 12-{\varepsilon}) N_i} = o\bigl(N2^{-N\tau( % \frac12 - {\varepsilon})}\bigr)$ many cylinders [where $N_i:=\# T_i\cap % \{1,\ldots, N\}$], each of them of diameter $2^{-N_i}$... % % ..... . .... .......... ............. % % % % % \hhh \emph{\bf 4. ($\phi(M)=0 \impl \rho(M)=0$)} Let us assume that $\rho(M)=d>0$. In view of the definition of $\rho(M)$ there is a sequence $a$ with $M(a\xf)=M$ and positive lower density ${\underline d(a)}>d - \delta$. We are going to consider the subset in $[0,1]$ corresponding to all subsequences of~$a$. For this we consider for $N \in \N$ the union of cylinder sets $\cJ_N$ defined by \[ [\varepsilon_1 \cdots \varepsilon_N] \in \cJ_N \quad \ifff \quad \varepsilon_j =0 \text{ unless } j=a_k \text{ for } j \in [0,N] \text{ and some } k. \] We set \[ \Lambda_a = \bigcap_{N=1}^\infty \cJ_N = \{b: S_b \subseteq S_a\}. \] There is a natural identification of $\Lambda_a$ with the space $\{0,1\}^\N$ of all $\{0,1\}$-sequences. % Namely, for $\underline\N = \omega_1 \cdots %\omega_k \cdots \in \Sigma_2$ we set $S_{b(\omega)}=\{k_i\omega_i : k_i\in %S_a\}$. We define a measure $\mu_a$ supported on $\Lambda_a$ as the push--forward of the Bernoulli measure on $\{0,1\}^\N$ generated by the probability vector $(\frac12,\frac12)$ \[ \mu_a (C_N(b))= \cases{ \left(\frac12 \right)^{\#(S_a \cap [1,N])} &\text{if $b \in \Lambda_a$}\cr 0 &\text{if $C_N(b) \cap \Lambda_a =\emptyset$} } \] This measure is a probability measure supported on $\Lambda_a$ and has the following property \[ \liminf_{N \to \infty} -\frac{\log \mu_a(C_N(b))}{N \log 2} = \liminf_{N \to \infty} \frac{\#(S_a \cap [0,N])}{N}= {\underline d(a)} >0 \qquad \text{for $\mu_a$--a.e. $b$} \] Frostman's Lemma (see~\cite[p.221]{Fal}) tells us that \[ \dimh \mu_a = {\underline d(a)}. \] Now we want to apply theorem~\ref{3} to the case $\yf = a\xf$. In order to do this we identify the subsequences of $a$ with points in the interval $[0,1)$ in the standard way. Under this identification the measure $\mu_a$ transforms into the normalized Lebesgue measure on~$[0,1)$. Hence, theorem~\ref{3} implies that $\mu_a$--a.e.\ subsequence $b$ of $a$ has the same limit behavior as $a$ --- i.e., $M(a\xf) = M(b\xf) = M$. By the above considerations the set of these subsequences $b$ has Hausdorff dimension~${\underline d(a)}$. This yields $\phi(M) \ge {\underline d(a)} > 0$. \end{proof} We actually proved slightly stronger results than stated in theorem~\ref{HDTheo1}. These results are stated below. \begin{Cor}\label{HDCor1} \[\dimh \{b\ | \ M(b\xf) = M(a\xf) \} \ge \ud (a).\] \end{Cor} \begin{Remark} The estimate in corollary~\ref{HDCor1} is not very effective for $\ud (a) \le \frac12$. In this case the estimate \[\dimh \{b\ | \ M(b\xf) = M(a\xf) \} \ge \tau(\ud (a)).\] is better. One can derive this estimate along similar lines as in the proof of part 4 of the theorem. \end{Remark} % For two sequences $S_a$ and $S_b$ we can define a non--symmetric % semi--distance via % \[ % \delta'(S_a,S_b)=\lim_{n\to\infty}\frac{\#\{k : k\in S_a\setminus S_b ; k\le % n\}}{\#\{k : k\in S_a ; k\le n\}}. % \] % % JOERG FRAGEN: SOLL DER FOLGENDE ABSATZ SO BLEIBEN? % % % % Let $\delta(S_a,S_b):=\delta'(S_a,S_b)+\delta'(S_a,S_b)$ be its % symmetric semi--distance. We can consider the % equivalence relation $a\sim b$ if % and only if % $\delta(S_a,S_b)=0$. % Then we have shown that % equivalent sequences have equivalent % asymptotics: $M(a\xf)=M(b\xf)$ if % $a\sim b$. % Moreover, if % we define an ordering on the equivalence % classes by~$a\le b$ if % and only % if $\delta'(S_b,S_a)=0$ and $M(a\xf)=M(b\xf)$, % then the above corollary gives an estimate of the % Hausdorff dimension of maximal chains. To get a complete picture one % would need to analyze the structure of this ordering, especially to % estimate the cardinality of maximal chains with given asymptotic % behavior. \bigskip We also want to state another consequence of theorem~\ref{HDTheo1}: \begin{Theo}\label{4} Let $\xf$ be an arbitrary sequence on~$X$. Assume that $M(\xf)$ contains a non-atomic measure. Then the set of all $a$ such that $M(a\xf) \neq M(\xf)$ has Hausdorff dimension~$1$. \end{Theo} \begin{proof} By corollary~\ref{DefrhoCor1} \[ \{a \ | \ M(a\xf) \neq M(\xf) \} \supset \{a \ | \ \rho (M(a\xf)) < 1 \}. \] Let $\mu \in M(\xf)$ be a non-atomic measure. We fix $\varepsilon >0$ and choose an open subset $U \subset X$ with measure $\mu(U)= \varepsilon$. We note that such subsets $U$ exist since $\mu$ is non-atomic. Let $a$ be the subsequence defined by \[ k \in a \quad \ifff \quad x_k \not \in U. \] By the definition of the lower density and of~$M(a\xf)$ this sequence has a subsequence $a'$ with density $\underline d(a')=1-\varepsilon$ and $\mu \not \in M(a'\xf)$. By corollary~\ref{HDCor1} we get \[ \dimh \{b \ | \ \rho (M(b\xf)) < 1 \} \ge \dimh \{b \ | \ M(b\xf) = M(a'\xf) \} \ge \ud (a') = 1-\varepsilon \] Since $\varepsilon$ was arbitrary this concludes the proof. \end{proof} % % \include{seq} \section{Sequences of maps} \label{7} Let now $\ff$ be a sequence of maps $f_n: X \to Y$. Thus $\ff: \N \times X \to Y$, $\ff(n,x) = f_n(x)$. For notational convenience we also write $\ff(x)$ for the sequence $(f_n(x))_{n \in \N}$. To $\ff$ there corresponds a map $\Phi = \Phi_\ff: I^2 \to X^{\N}$, $(a,x) \mapsto a \ff(x)$ where $a\ff(x)$ is the subsequence $(f_{k_n}(x))_{n \in \N}$ of~$(f_n(x))_{n \in \N}$ which corresponds to~$a$ in the sense described in the first section. This point of view is similar to that in \cite{W1}, where several related metric results can be found. Let $T \tm I \times X$, $a \in I$, $x \in X$, then $_aT = \{x \in X \ | \ (a,x) \in T\}$ and $T_x = \{a \in I \ | \ (a,x) \in T\}$ denote the vertical and horizontal cross-section of~$T$. Important informations about $T$ and its cross-sections are their Hausdorff dimensions $\dim_H$, their $d$-dimensional Hausdorff measures $m_d$ and their Baire categories as subsets of $I \times X$ resp.\ $X$ or~$I$. We turn to the Baire aspect. % LIEBER MARTIN! DER FOLGENDE SATZ IST DIR GEWIDMET. DOCH KEINE EHRE % OHNE GEGENLEISTUNG: KOENNTEST DU NACHPRUEFEN; OB DER BEWEIS O.K. IST? % INSBESONDERE BIN ICH MIR NICHT SICHER, OB KURATOWSKI ULAM IN DER % UNTERSTELLTEN ALLGEMEINHEIT GILT. \begin{Theo}\label{10} Let $X,Y$ be complete metric spaces, $Y$ second countable, and $\ff$ a sequence of continuous maps $f_n: X \to Y$. Suppose $D \tm Y$ is a dense set such that for every $y \in D$ the sets \[ D_y = \bigcup_{n \in \N} f_n^{(-1)}(y) \] are dense in~$X$ in a strong sense, meaning that every nonempty open set in~$X$ has an infinite intersection with~$D_y$. Then the following statements hold: \newline (i) There is a residual set $R \tm I \times X$ such that $a\ff(x)$ is maldistributed for all $(a,x) \in R$. \newline (ii) There is a residual set $R \tm I$ such that for all $a \in R$ the sequence $a\ff(x)$ is maldistributed for all $x$ from a residual set $R(a) \tm X$. \newline (iii) There is a residual set $R \tm X$ such that for all $x \in X$ $a\ff(x)$ is maldistributed for all $a$ from a residual set $R(x) \tm I$. \end{Theo} \begin{proof} By the theorem of Kuratowski-Ulam (cf.~\cite{Ox}) the three statements of the theorem are equivalent. Hence it suffices to prove (iii). For an arbitrary nonempty open set $O \tm Y$ define \[ R(O) = \bigcup_{n \in \N} f_n^{-1}(O). \] By assumption $R(O)$ is open and dense. Since $Y$ is second countable it has a countable topological base of nonempty open sets~$O_k$, $k \in \N$. Since $X$ is a complete metric space this implies that \[ R = \bigcap_{k \in \N} R(O_k) \] is residual in~$X$. If $x \in R$ and $U$ is a nonempty open set in~$Y$ we have $O_k \tm U$ for some $k$ and, since $x \in R(O_k)$, $x \in f_n^{-1}(O_k)$ for some $n \in \N$, i.e.\ $f_n(x) \in O_k \tm U$. This proves that $(f_n(x))_{n \in \N}$ is dense in~$Y$, i.e., $A(\ff(x))=Y$. Thus (iii) follows from theorem \ref{2}. \end{proof} \begin{Remark} To get an analogous metric result [with ``maldistributed'' replaced by ``uniformly distributed''] one has to use Fubini's theorem instead of Kuratowski-Ulam's. A sufficient condition on the sequence of functions for this case is mentioned in Koksma's metric theorem (cf.~\cite[p.34]{K-N}). \end{Remark} \noindent \begin{Elementary examples} If the maps $f_n$ are expanding in an appropriate sense the condition of theorem \ref{10} might be fulfilled, maybe under additional hypotheses on~$X$ and~$Y$. Concrete examples are: $X= T \tm \R$, $Y=I$, $f_n(x) = a_n x$, $f_n(x) = x^{a_n}$, $f_n(x) = a_n^x$ etc., or similar sequences on the $s$-dimensional torus $X=Y=I^s$. \end{Elementary examples} % LIEBER JOERG! AN DIESER STELLE SCHIESST MIR IN DEN SINN, DASS DU % JA SPEZIALIST DARIN BIST, HAUSDORFFDIMENSIONEN ZU BERECHNEN. % HIER WIRST DU SICHER LUST HABEN, DEN ABBILDUNGEN GEEIGNETE % DIFFERENZIERBARKEITEIGENSCHAFTEN UND UNTERE SCHRANKEN IHRER % ABLEITUNGEN ZU UNTERSTELLEN, UM INTERESSANTE ERGEBNISSE ZU ERHALTEN. % AN DIESE WERDEN SICH WAHRSCHEINLICH BEISPIELE OBIGER GESTALT ANSCHLIESSEN. % DESHALB HABE ICH MICH MIT DIESEN BEISPIELEN KURZ GEFASST, WEIL % ICH DIER NICHT INS HANDWERK PFUSCHEN WOLLTE. ICH BITTE DICH DAHER % AN DIESER STELLE UM DAS ENTSPRECHENDE THEOREM. SOLLTEN ES GLEICH MEHRERE % SEIN - NUR ZU! % \begin {Examples} The most natural application is to consider subsequences of ergodic iterations--i.e. sequences $T^{n_k}$ where $T\colon X\to X$ is a map. The most interesting case (from the point of view of uniform distributions) would be if $T$ is ergodic. This means \[ \lim_{n\to\infty}\frac1n \sum_{k=0}^{n-1}\mu(A\cap T^{-k} B)=\mu(A)\mu(B) \] for all measurable sets $A$ and~$B$. This property (for subsequences) obviously implies the above assumptions for the support of the measure. Unfortunately, this property is not inherited by subsequences. (This is one reason why subsequences of~$n\alpha$ are so interesting.) Therefore it might be better to consider systems which are topologically mixing. These are maps such that for all pairs of open sets $U$ and $V$ there is a number $N$ such that for all $n\ge N$ the intersection $U\cap T^N(V)$ is not empty. This condition follows if for all measurable sets $A$ and $B$ \[ \lim_{n\to\infty}\mu(A\cap T^{-n}B)=\mu(A)\mu(B) \] holds. Maps satisfying the latter condition are called mixing. There are many systems which are known as mixing. For instance any transitive expanding map of the n--torus is topologically mixing. Also a broad class of systems decomposes into mixing components. Unfortunately, group translations, and hence $n\alpha$--sequences, are not mixing. \end {Examples} % % \include{res} % oct 14 -- still not done % oct 13, 1997 % april 7, 1997 \def\fin{{\rm fin}} \def\itm#1 {\item[{(#1)}]} \def\dalpha{\dot{\alpha}} \def\dbalpha{\dot{\boldsymbol\alpha}} \def\balpha{\boldsymbol{\alpha}} \def\dx{\dot{x}} \section{Results on~$n\alpha$-sequences} \label{8} We now turn to a special case of the sequences $(f_n(x))_{n \in \N}$ discussed in the previous section, namely, the sequence $(n \alpha)_{n\in \N}$ and its subsequences. If $a \in [0,1)$, $S_a = \{ k_1 < k_2 < \cdots \,\}$, then we will identify $a$ with the sequence $(k_1, k_2, \ldots)$. For any $\alpha \in \R/\Z$ we let $\balpha$ denote the sequence $(\alpha , 2 \alpha , 3 \alpha , \ldots\,)$, so $a \balpha $ will be the subsequence $(k_1 \alpha , k_2 \alpha, \ldots\,)$. In this section we will use the ``forcing'' notation, which is a {\it fa\c{c}on de parler} that allows us to talk about ``the typical element'' of a Polish (i.e., complete separable metric) space. (We use this notation only for proofs, not for the theorems themselves. The reader who does not want to learn the forcing calculus will have no difficulty translating our proofs into more conventional ones.) \begin{Definition} \begin{enumerate} \item Let $X$ be a Polish space. We call the nonempty open set of~$X$ ``conditions'' and we view them as giving us information about the ``typical element $\dx\in p$''. If $p, q$ are conditions with $p \supseteq q$, we write $p \le q$ to emphasize that ``$\dx \in q$'' contains more information than ``$\dx \in p$.'' For any $B \subseteq X$ we write $p \forces B$ (``$p$ forces~$B$'') as an abbreviation for ``except for a meager set, every element of~$p$ is in~$B$'' or more formally: ``$p \setminus B$ is meager in $X$''. \item For each space $X$ we reserve a special variable~$\dot x$ to denote the ``generic'' or ``typical'' element of~$X$. The ``typical'' real number will be denoted $\dalpha$. \\ For any formula (or property) $ \varphi $ we write $$ p \forces \varphi(\dx)$$ iff $p \forces \{ x: \varphi(x) \}$, i.e., if ``most'' elements of~$p$ have the property $ \varphi $ (or equivalently, if most elements $x$ of~$X$ have the property: ``if $x\in p$, then $ \varphi(x)$.'') \end{enumerate} \end{Definition} \begin{Fact}\label{4.1} If $B$ is a nonmeager Borel set, then there is a condition (i.e, a nonempty open set) $p$ with $p \forces B$. \end{Fact} \begin{proof} All Borel sets have the Baire property (cf.~\cite[chapter 4]{Ox}). \end{proof} \begin{Fact}\label{f-} Assume that $ \varphi(x) $ and $\psi(x)$ are formulas such that the corresponding sets $\{x: \varphi (x) \}$ and $\{x: \psi (x) \}$ are Borel sets. \begin{enumerate} \item \label{f-cl} % cl=clopen If $p \Delta q$ is meager, then $p \forces \varphi(\dx) $ iff $ q \forces \varphi(\dx) $. (We will not need the converse, which is also true: If ``$p \forces \varphi(\dx) $ implies $q \forces \varphi(\dx) $'' for every Borel formula $ \varphi(\dx) $, then $ p \setminus q $ is meager.) \item \label{f-a} % a=and $p \forces\varphi(\dx) \& \psi(\dx) $ iff: $p \forces \varphi(\dx) $ and $p \forces \psi(\dx) $. \item \label{f-n} % n=negation ``$p \forces \varphi(\dx)$'' is false iff there is~$q \ge p$ such that $q \forces \lnot \varphi(\dx) $. (Apply fact~\ref{4.1} to $p \setminus \{ x: \varphi(x)\}$.) \item \label{f-e} % e=existential quantifier If $ p \forces \exists k\in \N\, \varphi(\dx, k)$ then there is~$q \ge p$ and $k_0\in \N$ such that $ q \forces \varphi(\dx, k_0)$. (This is a thinly disguised version of Baire's theorem.) \item \label{f-i} % i=inconsistency It is impossible that $p \forces \varphi(\dx) $ and $p \forces \lnot \varphi(\dx) $. \item \label{f-s} % s=stronger If $p \forces \varphi(\dx) $, $q \ge p$, then also $q \forces \varphi(\dx) $. % \item % \label{f-o} % o=open % For every $p$ there is~$q \ge p$, $q\not=\emptyset $ open, $q % \setminus p$ meager. (See \cite[???]{Ox}) \end{enumerate} \end{Fact} \bigskip \begin{Theorem} \label{25-1} Let ${a} = ({k}_0, {k}_1, {k}_2, \ldots\,)$ be a sequence of natural numbers. If\/ $\liminf_n ({k}_{n+1}/{k}_n) > 1 $, then the set $$ \{ \alpha: {a} \dbalpha \hbox { is u.d. } \} $$ is meager. \end{Theorem} This theorem answers a question of \v{S}al\'at \cite{Sa}, who has proved a similar result under the assumption $\liminf_n ({k}_{n+1}/{k}_n) \ge 2 $. We prepare ourselves for the proof of theorem~\ref{25-1} with a definition and a lemma. \begin{Definition} We say that a family of functions $(f_1, f_2, \ldots)$ is $ \varepsilon $-mixing if: for all intervals~$J$ of length $ \varepsilon $ and all $k\in \N$: $$ \bigcap_{n=1}^k f^{-1}_n(J) \ \not= \ \emptyset $$ More generally we say that $(f_1, f_2, \ldots)$ is $(\varepsilon,{\delta})$-mixing if: for all intervals $J$ of length $\varepsilon $, and all $k\in N$, and all intervals~$J'$ of length $ {\delta} $: \[ J' \cap \bigcap_{n=1}^k f^{-1}_n(J) \ \not= \ \emptyset \] \end{Definition} \begin{Remark} Although we work in~$I=[0,1)$, we identify elements in~$I$ with their equivalence classes in~$\R/\Z$, so an (open) interval can either be of the form $(a,b)$ or of the form $[0,a) \cup (b,1)$ (for $0\le a \frac{2}{ \varepsilon } {k}_n$ for all~$n$. \item $ {k}_1 > \frac{\varepsilon } {{\delta} }$. \end{enumerate} Then $(f_1, f_2, \ldots\, )$ is $(\varepsilon, {\delta})$-mixing. \end{Lemma} \begin{proof} Let~$J$ be an interval of length $ \varepsilon $, $ J' $ an interval of length $ {\delta} $. % (We assume $ \varepsilon, {\delta} \in % (0,1 )$, and by ``interval of length $ {\delta} $ me mean either a set % $(a, a+ {\delta} )$ with $0 \le a < a+ {\delta} \le 1$ or a set % $[0, {\delta} - a) \cup (1-a, 1)$, with $0 \le a< {\delta} < 1$. We will show (by induction on~$m$) that each set $$ J' \cap \bigcap_{i=1}^m f^{-1}_i(J)$$ contains in fact an interval $I_m$ of length $ \varepsilon/{k}_m$. This is clear for~$m=0$, as the length of~$J'$ is $ {\delta} > \varepsilon/{k}_0$. Consider~$m>0$. Note that $f^{-1}_m(J) = \bigcup_{j=1}^{{k}_m} \{\alpha : {k}_m\cdot \alpha - j \in J \}$ is a union of~${k}_m$ many disjoint intervals, each of length $ \varepsilon /{k}_m$. By inductive assumption, the set $ J' \cap \bigcap_{i=1}^{m-1} f^{-1}_i(J)$ contains an interval $I_{m-1} $ of length $ \varepsilon/{k}_{m-1} $. Since $ \varepsilon / {k}_{m-1} > 2/{k}_m$, we can find a natural number~$j $ such that the interval $$ [\frac{j}{{k}_m}, \frac{ j+1}{{k}_m} ] = \{ \alpha : {k}_m \cdot \alpha - j \in [0,1]\}$$ is contained in~$I_{m-1} $. Hence the set \[I_m:= \{\alpha : {k}_m \cdot \alpha - j \in J \}, \] an interval of length $ \varepsilon / {k}_m$, is also contained in~$I_m$. \end{proof} \bigskip \noindent {\bf Proof of theorem \ref{25-1}} Let $1 < q < \liminf_n ({k}_{n+1}/{k}_n)$. Without loss of generality we have ${k}_{n+1} / {k}_n>q$ for {\em all}~$ k$. Let $\dalpha $ be the ``generic'' real number. Towards a contradiction, assume that ``${a} \dbalpha $ is not u.d.'' is not forced, then by fact~\ref{f-}(\ref{f-n}) there is a condition~$p$ such that \[ p \forces {a} \dbalpha \hbox{ is u.d.} \] Pick an integer $c>1$ such that $q ^ c \ge 9 c$, and let $ \varepsilon = \frac{1}{3c}$. Let $J$ be any interval of length $ \varepsilon $. So we have $$ p \forces \frac {1}{n} \#\{j \frac{\varepsilon}{{\delta}} $. Define for $i=0, 1,2,\ldots, 100 n_0$: $$ f_i(x) = {k}_{(n_0+i)c}x. $$ Note that $ {k}_{(n_0+i+1)c}/ {k}_{(n_0+i)c} \ge q^c \ge 9c > 6c = 2/\varepsilon $. Hence (by lemma \ref{17}) the functions $f_0, f_1, \ldots$ are $(\varepsilon, {\delta})$-mixing, so there is an interval~$ q \subseteq p $ of positive length such that \[ q \subseteq \bigcap_{i=0}^{100n_0} f_i^{-1}(J) \] and hence: \[ q \forces \forall i \in \{0, \ldots, 100n_0\}: {k}_{(n_0+i)c} \dalpha \in J.\] But this means that \[ q \forces \#\{j\le 100n_0: {k}_j \alpha \in J \} \ge \#\{n_0 c, (n_0+1)c,\ldots, 100n_0c\} \ge 99n_0>\frac{1}{2}(100n_0)\] a contradiction to~$(*)$. \begin{Theorem}\label{explicit} If ${a}=\kn$ with $\lim k_{n+1}/k_n = \infty$, then the set $\{ \alpha: a\balpha \mbox{m.d.}\}$ is residual. \end{Theorem} \begin{proof} Similar to the above (but easier). \end{proof} \bigskip % DAS FOLGENDE PASST JEDENFALLS NICHT HIERHER. VIELLEICHT KOENNTE MAN % ES ZU EINER BEMERKUNG KONDENSIEREN UND IN DIE EINLEITUNG VERLEGEN. Obviously, ``$\liminf$'' is not an appropriate concept for questions dealing with uniform distribution, since for example a modification of a very thin subsequence of a sequence of reals can change the limes inferior without affecting the distribution behavior of the sequence. A more appropriate concept would be the following: \begin{Definition} Let $\cJ$ be an ideal of subsets of~$\N$, $x=\xn$ a sequence of real numbers. We define the lower limit (with respect to~$\cJ$) of this sequence as follows: \[ \liminf\nolimits^\cJ x := \sup \{ t\in \R \mid \{n:x_n < t \} \in \cJ \},\] or equivalently, \[ \liminf\nolimits^\cJ x := \sup \{ t\in \R \mid \{n:x_n \le t \} \in \cJ \}.\] We define $\limsup^\cJ$ dually: $ \limsup\nolimits^\cJ x := \inf \{ t\in \R \mid \{n:x_n \ge t \} \in \cJ \}$. \end{Definition} Thus, the usual notion of~$\liminf$ is $\liminf^{\text{fin}}$, where $\text{fin}$ is the ideal of finite subsets of~$\N$. In our context the ideal of sets of density~$0$ is the more interesting notion. For example, we can refine the previous theorem to the following: \begin{Remark} \label{26-1} Let $\cJ$ be the ideal of sets with density 0. Let ${a} = ( {k}_1, {k}_2, \ldots \,)$, ${k}_1 < {k}_2 < \cdots $. If $\liminf^\cJ_n ({k}_{n+1}/{k}_n) > 1 $, then the set \[ \{ \alpha: a \alpha \hbox { is u.d. } \} \] is meager. \end{Remark} We leave the (easy) proof of this fact, as well as the formulation of similar generalizations for other theorems, to the reader. % % \proof Let $1 < q < \liminf^\cJ_n ({k}_{n+1}/{k}_n) > 1 $. Define a % sequence $(z_n)_{n\in \N}$ by $z_1={k}_1$, $ z_{n+1} = % \min \{ {k}_j: {k}_j > z_n, {k}_j > q\cdot {k}_z\}$. Then we have % \[ \#\{ {k}_n:nn$ and a sequence $(k_{n+1}, \ldots, k_m)$ such that \begin{itemize} \item $(k_1, \ldots, k_m) \in X_c^\fin$ \item $k_m\ge 2R\cdot m$ \end{itemize} \item Let~$n$, $R$, $m$ be as above, and assume that $k_{m+\ell+1}-k_{m+\ell} \in [R, 2R]$ for all $\ell\ge 0$. Then \begin{itemize} \itm a for all~$i$: $k_{m+i} \ge R(m+i)$ \itm b $(k_1, k_1, \ldots) \in X_c$ \end{itemize} \end{enumerate} \end{Lemma} \noindent \begin{proof}[Proof of (1)] %%% ??? !!! We may assume $n \ge 4R$ [otherwise we start by replacing $(k_1, \ldots, k_n)$ by~$(k_1, \ldots, k_n, k_n+1, k_n+2, \ldots, k_n+(4R-n))$]. Let $m= n^3$, and define \begin{itemize} \item $k_{n+\ell} = k_n+2\ell$ for $n+\ell\le n^2$. \item $k_{j n^2 + \ell} = k_{k n^2}+(k+1)\ell$ for $j=1,\ldots, n-1$, $\ell=1,\ldots, n^2$. \end{itemize} % % We first notice $k_i \ge i $ for all~$i$, hence \[ \frac{k_{n+\ell+1}}{k_{n+\ell}} \le 1+ \frac{2}{k_{n+\ell}} \le 1+ \frac{2}{n+\ell} \] for all $\ell < n^2-n$. Next, we see by induction that \[ k_{\ka n^2} \ge \frac{\ka(\ka+1)}{2} n^2 \ \ \ \ \mbox{for $\ka=1, \ldots, n$} . \] so \begin{align*} \frac{ k_{\ka n^2 + \ell + 1}} { k_{\ka n^2 + \ell }} &\le 1 + \frac{ \ka+1 } { k_{\ka n^2 + \ell }} = 1 + \frac{ \ka+1 } { k_{\ka n^2} + \ell(\ka+1) } \le \\ &\le 1 + \frac{ \ka+1 } { \frac {1}{2} \ka(\ka+1)n^2 + \ell(\ka+1)} = 1 + \frac{ 2 } { \ka n^2 + 2\ell} \le 1+\frac{2}{\ka n^2 + \ell}. \end{align*} Finally, we have $k_{n^3} \ge \frac{n(n+1)}{2} n^2 > n^3\cdot \frac {n}{2} \ge Rn^3$. \smallskip The proof of (2) is trivial. \end{proof} \begin{Fact} \label{continuous} % \begin{enumerate} %\item The map $\xf \mapsto M(\xf)$ is continuous (if we use the % box topology % on~$X^\N$ and the Hausdorff metric on the closed % subsets of~$\M(X)$. %\item Let $U \subseteq \M(X)$ be open, $\mu \in U$. Then the closed %set~$\{\mu\}$ is an interior point of the set %$\{ F \subseteq U: F \mbox{ closed }\}$. %\item If $\mu_{N,\xf} \to \mu_0$ (weakly, i.e., $\xf$ is $\mu_0$-u.d.), $U \subseteq \M(X)$ an open set containing~$\mu_0$, then there is $ \varepsilon>0$ such that for all $\yf \in X^ \N$: \[ \mbox{ If\/ $ \| \xf - \yf \|_\infty < \varepsilon$, \ then $ M(\yf) \subseteq U$} \] %\end{enumerate} \end{Fact} % \proof MARTIN!!! \qed \bigskip \noindent {\bf Proof of theorem \ref{26-2}:} Fix an irrational number $ \alpha $, and assume that the set \[ \{ {\xfora} \in X_{ c }: {\xfora} \balpha \hbox{ is m.d.}\}\] is not residual. So there is an open set~$p$ such that \[ p \forces \mbox{ $ \dxf \balpha $ is not m.d.} \] Now ``$\dxf \balpha $ is not m.d.'' is equivalent to ``$M(\dxf \balpha ) \not= \M(X)$'', so since $M(\dxf \balpha )$ must be closed this means \[ p \forces \mbox{there is a basic open set~$U$ such that $U\cap M(\dxf \balpha ) =\emptyset$.}\] By \ref{f-}(\ref{f-e}) can we can find a basic open set $U \subseteq \M(I)$ and a condition $p' \ge p$ such that \[ p' \forces U \cap M(\dxforda\balpha) = \emptyset\] By strengthening $p'$ to a condition $q$ and by shrinking~$U$ we can even get \[ q \forces \forall n: \mu_{n, {\dxforda}\balpha} \notin U\] Now we will use the fact that $q$ gives only finitely much information about $\dxforda$. Let $\mu_0\in U$, and let $\yf = \yn$ be a sequence with $\mu_{n,\yf}\to \mu_0$. Choose $ \varepsilon $ as in fact~\ref{continuous}. The condition $q$ is a cylinder ~$q = [q_1, \ldots, q_n]$. % basic neighborhood in~$X_c$, defined by $(q_1, \ldots, q_n)\in % X_c^\fin$. Without loss of generality % $n \ge k_1$ and $1/n < \varepsilon $. Choose $R$ so large that the set $\{ \alpha, 2\alpha, \ldots, R\alpha\}$ meets every interval of length $\varepsilon $ (mod 1). By lemma~\ref{7-1} we may extend the sequence $(q_1, \ldots, q_n)$ to a sequence $(q_1, \ldots, q_m)$ satisfying $q_m \ge 2R m$. For all $i > m $ we can now choose \[q_i \in [q_{i-1}+R, q_{i-1}+2R]\] such that $ \alpha q_i \in [y_i-\varepsilon, y_i+\varepsilon]$ (mod 1), and $\qn \in X_c$. Let ${\bf q} = (q_n)_{n\in \N}$. We have $M({\bf q}) \subseteq U$, so for large enough $n$ we have $\mu_{n,{\bf q}\balpha} \in U$. Since trivially $(q_1, \ldots, q_n) \forces \dot {k}_1 = q_1, \ldots, \dot {k}_n=q_n$, we have \[ (q_1, \ldots, q_n) \forces \mu_{n,\dxf \balpha} \in U,\] a contradiction. % Let $r:= (q_1, \ldots, q_{m^2})$. Then % \[ r \forces \{ i \le m^2: k_i \alpha \in I \} \supseteq \{ i\le m^2: % i > m \} \] % hence % \[ r \forces card \{ i \le m^2: k_i \alpha \in I \} \ge m^2-m > % m^2(1- \varepsilon ),\] % a contradiction. % \bigskip We now deal with the (easier) case where $c<2$. It turns out that in this case the space $X_c$ has a trivial structure: It contains a countable dense set of isolated points. \begin{Theorem}\label{26-3} Let $c < 2$. Then for every irrational number $ \alpha $ the set \[ \{ {a} = \kn \in X_{ c }: {\xfora} \balpha \hbox{ is u.d.}\}\] contains a countable dense open subset of~$X_{ c}$. \end{Theorem} \begin{proof} Let \[X^*_c:= \{ {\xfora} \in X_c: \exists \ka \, \exists N\, \forall n\ge N \ k_n = n+\ka\}\] Clearly $ X^*_c$ is a countable dense subset of~$X_c$, and for any irrational $ \alpha $ we know that $ {\xfora} \dalpha $ is u.d.\ for every $ {\xfora} \in X^*_c$. It remains to show that $X^*_c$ is open. We will show that all points of~$ X^*_c$ are isolated in~$ X_c$. Assume that for all $ n> N$ we have $k_n=n+\ka$. Without loss of generality we have $N\cdot (\frac{2}{c} -1) > \ka$, so~$(n+\ka)\cdot \frac{c}{n} < 2$ for all~$n\ge N$. Let \[ U:=\{ \yf= \yn\in X_c: y_1=k_1, \ldots, y_N=k_N\}\] Then $U$ is a basic open neighborhood of~${\xfora}$, and ${\xfora} $ is the only point in~$U$. \end{proof} Consider the following question: given $a$, what is the behavior of~$a \balpha$ for ``typical'' $\alpha$? \relax From theorem~\ref{2} we know that for a residual set of~$a$'s the typical behavior of~$a \balpha$ is maldistribution. (An explicit example of such an~$a$ is mentioned in theorem~\ref{explicit}.) For some $a$'s (e.g., $a= (1,2,3,\ldots\,)$) the typical behavior is uniform distribution. In general, the following dichotomy holds: \begin{Theorem} For any $a = \kn$ the set \[ A = \{ \alpha \in \R/\Z: \mbox{$a \balpha$ is u.d.}\} \] is either meager or residual. \end{Theorem} \begin{proof} Assume that $A$ is not meager, then there is an interval $[a,b]$ such that $A \cap [a,b]$ is residual in $[a,b]$. Clearly, $ \alpha \in A$ implies $n \alpha \in A$ for any $n \in \N$. Now for $n > \frac{1}{b-a}$ the set $n\cdot A$ is residual in $n \cdot [a,b] = \R/\Z$. \end{proof} \begin{Open Problem} Is it true that \begin{quote} for all $a$, one of the sets \[ \{ \alpha: a\balpha \mbox{ is u.d.}\} \qquad \{ \alpha: a\balpha \mbox{ is m.d.}\} \] is residual? \end{quote} The proof of theorem~\ref{25-1} suggests a negative answer. \end{Open Problem} \bigskip % \bibliographystyle{../biblio/lit1plain} % \bibliography{other,winkler} \begin{thebibliography}{R-W} \makeatletter \renewcommand{\@biblabel}[1]{[#1]} \makeatother \bibitem[Be]{Be}D.~Berend. \newblock IP-sets on the circle. \newblock {\em Can. J. Math.}, {\bf XLII}:575--589, 1990. \bibitem[D-T]{D-T}M.~Drmota and R.~F. Tichy. \newblock {\em Sequences, Discrepancies and Applications}, volume 1651 of {\em Lecture Notes in Mathematics}. \newblock Springer, 1997. \bibitem[E-T]{E-T}P.~Erd{\H o}s and S.~J. Taylor. \newblock On the set of points of convergence of a lacunary trigonometric series and the equidistribution properties of related sequences. \newblock {\em Proc. Lond. Math. Soc.}, {\bf 7}:598--615, 1957. \bibitem[Fal]{Fal}Falconer. \newblock {\em Fractal Geometry}. \newblock Mathematical Foundations and Applications. John Wiley \& Sons, Chichester, 1990. \bibitem[Gr]{Gr}P.~M. Gruber. \newblock In most cases approximation is irregular. \newblock {\em Rend. Sem. Mat. Univers. Politecn. Torino}, {\bf 41.1}, 1983. \bibitem[K-N]{K-N}L.~Kuipers and H.~Niederreiter. \newblock {\em Uniform distribution of sequences}. \newblock Wiley, New York, 1974. \bibitem[L-T]{L-T}V.~Losert and R.~F. Tichy. \newblock On uniform distribution of subsequences. \newblock {\em Prob. Th. Rel. Fields}, {\bf 72}:517--528, 1986. \bibitem[My]{My}G.~Myerson. \newblock A sampler of recent developments in the distribution of sequences. \newblock {\em Lecture Notes in pure and applied Mathematics}, {\bf 147}:163--190, 1993. \bibitem[Ox]{Ox}J.~C. Oxtoby. \newblock {\em Measure and Category}. \newblock Graduate texts in Mathematics. Springer, 1980. \bibitem[R-W]{R-W}J.~M. Rosenblatt and M.~Wierdl. \newblock Pointwise ergodic theorems via harmonic analysis. \newblock In {\em Ergodic theory and its connections with harmonic analysis}, volume 205 of {\em London Math. Soc. Lecture Notes}, pages 3--151, Cambridge, 1995. Cambridge University Press. \newblock Proceedings of conference in Alexandria, Egypt, May 24--28, 1993. \bibitem[Si]{Si}K.~Sigmund. \newblock On minimal centers of attraction and generic points. \newblock {\em J. Reine Angew. Math.}, {\bf 295}:72--79, 1977. \bibitem[Sa]{Sa}T.~\v{S}al\'at. \newblock A remark on normal numbers. \newblock {\em Rev. Roumaine Math. Pures Appl.}, {\bf 11}:53--56, 1966. \bibitem[W1]{W1}R.~Winkler. \newblock Strong {W}eyl property. \newblock {\em Lecture Notes (Number-Theoretic Analysis)}, {\bf 1452}:208--220, 1990. \bibitem[W2]{W2}R.~Winkler. \newblock On the distribution behavior of sequences. \newblock {\em Mathematische Nachrichten}, {\bf 186}:303--312, 1997. \end{thebibliography} \end{document} % LocalWords: Theo Cor Def Mal pt diam cf