Abstract Generated abstract
This paper gives a sufficient criterion for a sequence of real numbers in the unit interval to be completely uniformly distributed. The criterion assumes a uniform upper bound, proportional to volume, for the limiting frequency with which consecutive blocks of any dimension fall in an arbitrary parallelepiped of the corresponding unit cube. The proof uses binomial measure estimates for deviations in the number of hits in intervals and higher-dimensional parallelepipeds, then applies a grouping argument to control bad blocks and pass to shifted sequences. It concludes that the empirical frequencies of all finite consecutive blocks in arbitrary parallelepipeds converge to their volumes, which is the asserted complete uniform distribution.
Full Text
MATHEMATICS
A. G. POSTNIKOV
A CRITERION FOR A COMPLETELY UNIFORMLY DISTRIBUTED SEQUENCE
(Presented by Academician I. M. Vinogradov on 17 I 1958)
In the present paper a criterion is given for a sequence of real numbers \(\alpha\), taken from the interval \([0,1]\):
\(\alpha=\alpha_1,\alpha_2,\alpha_3,\ldots\), to be completely uniformly distributed.* Our arguments will be analogous to the arguments of § 4 of paper \((^2)\).
By \(\Delta_s=(\delta_1,\delta_2,\ldots,\delta_s)\) we shall denote a parallelepiped lying in the \(s\)-dimensional unit cube, defined by the condition that the \(i\)-th coordinate of its points belongs to the interval \(\delta_i\); by \(|\Delta_s|\) we shall denote the volume of the parallelepiped \(\Delta_s\).
Theorem. Let the sequence \(\alpha\) be such that there exists a constant \(c\) such that, for any \(s \geqslant 1\) and any \(\Delta_s\),
\[ \lim_{P\to\infty}\frac{N_P(\Delta_s)}{P}<c|\Delta_s|. \]
Then the sequence \(\alpha\) is completely uniformly distributed.
Lemma 1. Let \(r\geqslant 1\) be an integer. Let \(\delta\) be some interval in \([0,1]\), and let \(|\delta|\) be its length. Consider the unit cube in \(l\)-dimensional space. The Lebesgue measure of those points of this cube for which the number of coordinates falling in the interval \(\delta\) (denote this number by \(A_\delta^l\)) satisfies the inequality
\[ \left|A_\delta^l-l|\delta|\right|\geqslant \frac{l}{r}, \]
does not exceed \(r^4/4l^2\).
Proof. There are \(C_l^k\) ways by which one can distribute \(k\) marks “fell in \(\delta\)” among \(l\) positions. Let \(\delta\) be the interval \([a,b]\). The volume of the region defined by the condition that \(k\) specified coordinates satisfy the inequalities \(a\leqslant \xi\leqslant b\), and the remaining \(l-k\) the inequalities either \(0\leqslant \xi\leqslant a\), or \(b<\xi\leqslant 1\), is equal to \((b-a)^k(1-(b-a))^{l-k}=|\delta|^k(1-|\delta|)^{l-k}\). Therefore the volume of the region of points \(k\) of whose coordinates have fallen in \(\delta\), and the remaining ones outside \(\delta\), is \(C_l^k|\delta|^k(1-|\delta|)^{l-k}\). Hence the measure sought in the lemma is equal to
\[ L=\sum_{|k-l|\delta||\geqslant l/r} C_l^k|\delta|^k(1-|\delta|)^{l-k}. \]
Carrying out the usual estimate of this expression, we obtain the lemma.
* For the definition of a completely uniformly distributed sequence, see paper \((^1)\).
Let \(s \geqslant 1\) and \(l \geqslant 1\) be natural numbers. Consider the unit cube in \(ls\)-dimensional space. Let \(\Delta_s=(\delta_1,\ldots,\delta_s)\) be a fixed parallelepiped in the unit cube of \(s\)-dimensional space; \(|\Delta_s|=|\delta_1|\cdots|\delta_s|\) is the volume of this parallelepiped. Let some point of the cube have coordinates
\[
(a_1,a_2,\ldots,a_s,a_{s+1},\ldots,a_{2s},\ldots,a_{(l-1)s+1},\ldots,a_{ls}).
\]
Group the coordinates \(s\) at a time, \((B_1\ldots B_l)\), where
\[
B_k=(a_{(k-1)s+1}\ldots a_{ks}),\qquad k=1,2,\ldots,l,
\]
points of the unit cube of \(s\)-dimensional space. Denote by \(A_{\Delta_s}^{(l)}\) the number of these points that fall into \(\Delta_s\).
Lemma 2. Let \(r \geqslant 1\) be an integer. The Lebesgue measure of the points of the unit cube of \(ls\)-dimensional space for which
\[
\bigl|A_{\Delta_s}^{(l)}-l|\Delta_s|\bigr| \geqslant \frac{l}{r}
\]
does not exceed \(r^4/4l^2\).
Proof. We compute the Lebesgue measure of the set of those points of the \(ls\)-dimensional cube for which, in the representation \((B_1,\ldots,B_l)\), at certain \(x\) places there is a hit in \(\Delta_s\), and at the remaining places there is not. Let \(E_s\) be the unit cube of \(s\)-dimensional space. Obviously, this measure is equal to
\[
\underbrace{\int_{\Delta_s}\cdots\int_{\Delta_s}}_{x\ \text{times}}
\underbrace{\int_{E_s/\Delta_s}\cdots\int_{E_s/\Delta_s}}_{l-x\ \text{times}}
dx_1\cdots dx_{ls}
=
|\Delta_s|^x(1-|\Delta_s|)^{l-x}
\]
(each integral is \(s\)-fold).
Since the \(x\) hits can be arranged among the \(l\) places in \(C_l^x\) ways, the measure sought in the lemma is equal to
\[
L=\sum_{\substack{x=0\\ |x-l|\Delta_s||\geqslant l/r}}^{l}
C_l^x|\Delta_s|^x(1-|\Delta_s|)^{l-x}.
\]
Carrying out the usual estimate, we obtain that \(L\leqslant r^4/4l^2\).
Proof of the theorem. Let the sequence \(\alpha=\alpha_1,\alpha_2,\ldots,\alpha_s,\ldots\) satisfy the condition of the theorem. Let \(\Delta_s=(\delta_1,\ldots,\delta_s)\) be a parallelepiped in the unit cube of \(s\)-dimensional space. Arrange the numbers of the sequence into groups of \(s\):
\[
P_1^{(s)},P_2^{(s)},\ldots,
\]
where
\[
P_k^{(s)}=(\alpha_{(k-1)s+1},\ldots,\alpha_{ks}),
\]
i.e. consider the sequence of points of the \(s\)-dimensional unit cube. From \(X\) terms of the sequence we obtain \([X/s]\) points. Denote by \(A_{\Delta_s}^{[X/s]}\) the number of times the points \(P_j^{(s)}\) \((j=1,\ldots,[X/s])\) fall into \(\Delta_s\). Choose arbitrarily \(l\geqslant 1\) and arrange the points \(P^{(s)}\) into groups of \(l\):
\[
\underbrace{P_1^{(s)}P_2^{(s)}\cdots P_l^{(s)}}_{l}\
\underbrace{P_{l+1}^{(s)}P_{l+2}^{(s)}\cdots P_{2l}^{(s)}}_{l}\ \cdots
\]
Take a natural \(r\geqslant 1\) and call an \(l\)-term group of points “good” if the number of points falling into \(\Delta_s\) is equal to \(l(|\Delta_s|+\theta/r)\), \(|\theta|\leqslant 1\), and “bad” otherwise. Denote by \(M([X/s])\) the number of times bad groups occur in the sequence of groups formed from the first \(l\left[\frac{[X/s]}{l}\right]\) numbers of the sequence \(\alpha\).
Then the number of good groups will be
\[
\left[\frac{X}{sl}\right]-M\left(\left[\frac{X}{s}\right]\right)
=
\frac{X}{sl}-M\left(\left[\frac{X}{s}\right]\right)+O(1)
\]
with an absolute constant in \(O\). A good group contributes \(l(|\Delta_s|+\theta/r)\) points falling into \(\Delta_s\). Therefore, among the \([X/s]\) points \(P^{(s)}\), the number falling into \(\Delta_s\) is
\[ A_{\Delta_s}^{[X/s]} = l\left(|\Delta_s|+\frac{\theta}{r}\right) \left(\frac{X}{sl}-M\left(\left[\frac{X}{s}\right]\right)\right) + l\theta_1 M\left(\left[\frac{X}{s}\right]\right)+O(l). \]
The term \(O(l)\) arises from the fact that, possibly, there is an incomplete group of points. Hence
\[ A_{\Delta_s}^{[X/s]} = \frac{X}{s}\left(|\Delta_s|+\frac{\theta}{r}\right) + O\left(M\left(\left[\frac{X}{s}\right]\right)l\right)+O(l). \]
Let \(\mathfrak M\) denote the set of those points in the \(ls\)-dimensional unit cube for which
\[
\left|A_{\Delta_s}^{(l)}-l|\Delta_s|\right|\geq l/r.
\]
Each bad combination represents a point of \(\mathfrak M\). Therefore
\[
M\left(\left[\frac{X}{s}\right]\right)\leq N_{s[X/s]}(\mathfrak M).
\]
But for \(X\geq X_0\), according to the hypothesis of the theorem,
\[
N_{s[X/s]}(\mathfrak M)\leq 2C\,\operatorname{mes}\mathfrak M\,X,
\]
\[
A_{\Delta_s}^{[X/s]}
=
\frac{X}{s}\left(|\Delta_s|+\frac{\theta}{r}\right)
+
O\left(X\frac{r^4}{l}\right)+O(l)
\]
(by Lemma 2).
Hence, letting \(X\) tend to infinity, we obtain
\[ \overline{\lim}_{X\to\infty} \left| \frac{A_{\Delta_s}^{[X/s]}}{X/s} - |\Delta_s| \right| \leq \frac{1}{r} + O\left(\frac{sr^4}{l}\right). \]
Now letting the parameter \(l\) tend to infinity, we obtain
\[ \overline{\lim}_{X\to\infty} \left| \frac{A_{\Delta_s}^{[X/s]}}{X/s} - |\Delta_s| \right| \leq \frac{1}{r}. \]
Consider the \(s\) sequences \(T^j\alpha,\ j=0,1,2,\ldots,s-1\) (among which \(T^0\alpha\) is our sequence \(\alpha\)),
\[ T^j\alpha=\alpha_{j+1},\alpha_{j+2},\alpha_{j+3},\ldots \]
Each of these sequences satisfies the condition of the criterion. Denote the quantity \(A_{\Delta_s}^{[X/s]}\), constructed for the sequence \(T^j\alpha\), by \(A_{\Delta_s}^{[X/s]}(T^j\alpha)\). We have
\[ \lim_{X\to\infty} \frac{A_{\Delta_s}^{[X/s]}(T^j\alpha)}{X/s} = |\Delta_s|, \qquad j=0,1,\ldots,s-1. \]
But, obviously,
\[ N_X(\Delta_s)=\sum_{j=0}^{s-1} A_{\Delta_s}^{[X/s]}(T^j\alpha)+O(s). \]
Hence
\[ \lim_{X\to\infty} \frac{N_X(\Delta_s)}{X} = s\frac{|\Delta_s|}{s} = |\Delta_s|. \]
Since \(s\geq 1\) and \(\Delta_s\) is an arbitrary parallelepiped, the theorem is proved.
Mathematical Institute named after V. A. Steklov
Academy of Sciences of the USSR
Received
12 I 1958
References
- A. G. Postnikov, Izv. AN SSSR, Ser. Mat., 22, No. 3 (1958).
- A. G. Postnikov, I. I. Pyatetskii, Izv. AN SSSR, Ser. Mat., 21, No. 4, 501 (1957).