Abstract
For geometric progressions with common ratios greater than 4, the speed of convergence to the uniform distribution is determined for almost all initial values.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
A sequence \(\{x_k\}\) of real numbers is said to be uniformly distributed mod 1 if
for all \(0\le a< b< 1\), where \(\langle x\rangle \) denotes the fractional part \(x - [x]\) of a real number x. Since the convergence is uniform in a and b, the following discrepancy is used to measure its speed.
For an arithmetic progression \(\{n_k\}\), the order of convergence of \(D_N\{n_k x\}\) was studied by Khintchin [9] and Kesten [8]. For uniformly distributed i.i.d. \(\{U_k\}\), Chung-Smirnov theorem asserts the law of the iterated logarithm
By various studies on lacunary series, it is known that a sequence \(\{n_kx \}\) behaves like uniformly distributed i.i.d. when \(\{n_k\}\) diverges rapidly. Actually Philipp [10] followed Takahashi’s method [11] and proved the result below by assuming the Hadamard gap condition \(n_{k+1}/ n_k \ge q > 1\).
Dhompongsa [3] proved that the limsup equals to \(\tfrac{1}{2}\) when \(\{n_k\}\) satisfies very strong gap condition \( \lim _{k\rightarrow \infty }({\log (n_{k+1}/ n_k)})/{ \log \log k} = \infty \). Beside of these results, any concrete value of limsup for exponentially growing sequence was not determined before the recent results below on divergent geometric progressions \(\{\theta ^k x\}\).
Theorem 1
[4–7] For any \(\theta \notin [-1,1]\), there exists a constant \(\varSigma _\theta \ge 1/2\) such that
If \(\theta ^j \notin \mathbf{Q}\) for any \(j \in \mathbf{N}\), then \(\varSigma _\theta = \frac{1}{2}\), and otherwise \(\varSigma _\theta > \frac{1}{2}\). When \(\theta \) satisfies \(\theta ^j \in \mathbf{Q}\) for some \(j \in \mathbf{N}\), we take p, q, and r as below.
If p and q are odd, then
If \(q=1\), then
according as p is odd, \(|p|\ge 4\) is even, \(p=2\), or \(p=-2\). If \(p=\pm 5\) and \(q=2\), then
For related works, see [1, 2]. In this paper, we prove the next result and determine \(\varSigma _\theta \) when \(\theta \) is large.
Theorem 2
Suppose that \(\theta \) satisfies (1). If p is odd, q is even and \(|p/q|\ge 9/4\), or if p is even, q is odd and \(|p/q|\ge 4\), then
where \(v(x)=\langle x \rangle (1-\langle x\rangle )\) and \(I=\min \{ n\in \mathbf{N}\mid q^n = \pm 1 \mod |p|-q\}\).
By (2), we can calculate the concrete values of \(\varSigma _\theta \) in the following way.
When \(q=1\) and \(|p|\ge 4\) is even, we have \(I=1\), and (2) gives the values stated in Theorem 1. It also gives the value \(\varSigma _{\pm 5/2}\) in Theorem 1.
We here emphasize that the values of \(\varSigma _\theta \) are determined for large \(\theta \), say, that satisfying \(|\theta |\ge 4\). On the other hand, we can find some smaller p / q for which (2) fails to hold. \(\theta =\pm 2\) is one such example.
Before closing this section, we note that the conditions \(|p/q|\ge 9/4\) and \(|p/q|\ge 4\) are superfluous, and much weaker conditions are sufficient. Set
When p is odd and q is even, if the conditions \(A_1(|p|, q) >0\), \(A_2(|p|, q)>0\), and
are satisfied, then we have (2). The condition \(|p/q|> 9/4\) implies these conditions. Although \(A_1(9, 4)< 0\) and \(A_2(9, 4)>0\), we can still prove (2) for \(p/q=\pm 9/4\). We here note that \(|p/q|\ge (2+\sqrt{6})/2 + 1/q\) also implies \(A_1(|p|, q) >0\) and \(A_2(|p|, q)>0\).
When p is odd and q is even, if (3), \(A_3(|p|, q)>0\) and \(A_4(|p|, q)>0\) are satisfied, then (2) holds. The condition \(|p/q|\ge 4\) implies these conditions.
2 Preliminary
We prepare some results. Proofs can be found in [4, 6, 7]. For a, b, \(a'\), \(b'\in [0,1)\), put
and
Here we list general properties that we use.
When \(\theta \) satisfies (1), \(\varSigma _\theta \) does not depend on r and is given by
Put \(b_i:= \frac{i}{p-q}\), \(c_1:= \frac{q-2}{2q}\), \(c_2 := \frac{p-3}{2p}\), \(c_3 := \frac{p-1}{2p}\), \(c_4 := b_{(p-q-1)/2} = \frac{(p-q-1)/2}{p-q}\), \(c_5 := \frac{q-1}{2q}\), \(c_6 := \frac{p-2}{2p}\), and \(c_7 := b_{(p-q-3)/2} = \frac{(p-q-3)/2}{p-q}\). When p and q are positive and satisfying \(p/q\ge 2\), one can verify \( c_1<c_2 < c_4 < c_3 \) and \(c_5 \le c_6 \le c_4\). By \(p=q\) mod \(p-q\), we have
Thanks to \(v(-x)= v(x)\), we have \(v(q^{m+nI} c_4) = v(\pm q^{m} c_4)=v(q^{m} c_4)\). Hence \(\sigma _{p/q}(0,c_4)\) equals to the left hand side of (2).
In the next two sections, we prove \( \varSigma _{p/q} = \sigma _{p/q}(0,c_4) \) in the case p is positive and p / q is large. By assuming this, we here prove \(\varSigma _{-p/q}= \varSigma _{p/q}\). In [6] we proved \(\varSigma _{-p/q} \le \varSigma _{p/q}\). If we find \(0\le \widehat{b}< \widetilde{b}\le 1\) with \(\sigma _{-p/q}(\widehat{b}, \widetilde{b}) = \sigma _{p/q}(0, c_4)=\varSigma _{p/q}\), we have the equality.
We put \( \widehat{b} = (1-c_4)p/({p+q})\) and \(\widetilde{b} = \widehat{b} + c_4\). It holds that \( \left\langle (-p)^{n} \widehat{b} \,\right\rangle =\left\langle q^{n} \widehat{b} \,\right\rangle \) and \(\left\langle (-p)^{n} \widetilde{b} \,\right\rangle =\left\langle q^{n} \widetilde{b} \, \right\rangle \) if n is even, and that \( \left\langle (-p)^{n} \widehat{b} \,\right\rangle =\left\langle q^{n} \widetilde{b} \,\right\rangle \) and \(\left\langle (-p)^{n} \widetilde{b} \,\right\rangle =\left\langle q^{n} \widehat{b} \,\right\rangle \) if n is odd. Actually, we have \((p+q)\widehat{b} =p-p c_4 = -p c_4 = -qc_4\) mod 1, which implies \(-p \widehat{b} = q\widetilde{b}\), \(-p\widetilde{b} = q\widehat{b}\) mod 1, and \((-p)^2 \widehat{b} = -pq\widetilde{b} = q^2 \widehat{b}\) mod 1, and so on. By noting (5), (7), and \( \left\langle \langle q^{n} \widetilde{b} \,\rangle - \langle q^{n} \widehat{b} \,\rangle \right\rangle = \langle q^{n} c_4\rangle \), we have
for odd n, and
for even n, to have \(\sigma _{-p/q}(\widehat{b}, \widetilde{b}) = \sigma _{p/q}(0, c_4)\).
We prepare some inequalities to prove \(\varSigma _{p/q} = \sigma _{p/q}(0,c_4)\). We denote \(\sigma _{p/q}(0,a)\) simply by \(\sigma (a)\).
Put
\(S_t(a):=2v(t^2a)/{(pq)^2}\), and \(Z(a) := V(a, a) + 2 v(qa)/{pq}\). Thanks to (6) and (10), we have
and
where
By \(\gcd (q, p-q)=1\), \(\langle p^k c_4\rangle = \langle q^k c_4\rangle \in \{b_1, \dots , b_{p-q-1}\}\) and \(v(b_i)\ge v(b_1)\), we have
Note that the evaluation of U strongly depends on the parity of p, since \(\langle qc_4\rangle = qc_4 - (q-2)/2\) or \(\langle qc_4\rangle = qc_4 - (q-1)/2\) according as q is even or odd.
We denote the derivative \(\frac{d}{da}f\) of f by Df, the right derivative by \(D^+f\), and the left derivative by \(D^-f\).
3 Odd p and even q
Assume that q is even, and that p is odd and positive. We divide [0, 1 / 2) into subintervals \([ 0, c_1)\), \([ c_1, c_2)\), \([c_2, c_3)\), and \([c_3, 1/2)\), and prove \(\sigma ^2(a) \le \sigma ^2(c_4)\) on each.
3.1 \([c_2, c_3)\) part
We assume (3), \(A_1(p,q)>0\), and \(a\in [c_2, c_3)\). We have \(\langle pa\rangle = pa- (p-3)/2\) and \(\langle qa\rangle = qa - (q-2)/2\). Since \(\langle pa\rangle <\langle qa\rangle \) holds on \([c_2, c_4)\), and \(\langle pa\rangle \ge \langle qa\rangle \) on \([c_4, c_3)\), we can evaluate \(X_1\) as
and verify that \(D X_1(a)\) decreases on \([c_2, c_3)\). We also have \(|D v(q^n a)| \le q^n \) a.e. a, and therefore \( |D Y_1| \le 2/{p(p-1)} \) a.e. By combining these, we have
and
Hence \( \sigma _{}^2(a) \le W_1(a) \le W_1(c_4) = \sigma _{}^2(c_4) \) for \(a\in [c_2, c_3)\). Here we note that functions appearing here are bounded and absolutely continuous, and the exceptional set of null measure for \(D W_1(a) > 0\) or \(D W_1(a) < 0\) does not harm the argument to show that \(c_4\) is the maximal point.
Here we prove \(A_1(p,q)>0\) by assuming \(p/q> 9/4\). Since \(2p^2 -(4q+2)p + (q^2-2q)\) is increasing in \(p\ge (9/4)q\), \(A_1(p,q)\) also. Because of \(A_1((9/4)q, q) = 3q^2(3q-28)/32>0\) if \(q\ge 10\), we see that \(A_1(p,q)>0\) if \(q\ge 10\) and \(p/q> 9/4\). Thanks to \(A_1(5, 2)> 0\), \(A_1(11, 4)>0\), \(A_1(17,6)>0\) and \(A_1(19,8)>0\), we see that \(A_1(p,q)>0\) if \(p/q> 9/4\).
We prove the case \(p/q=9/4\). Due to the above proof, we see that \(\sigma _{}^2(a) \le \sigma _{}^2(c_4)\) for \(a\in [c_2, c_4]\). Since we have \(T_1= \frac{1}{2520}\) and \(X_1\) is decreasing in \(a\ge c_4\), we have \(\sigma ^2(a) \le W_3(a) \le X_1(\frac{34}{81})+\frac{1}{2520} < \frac{222}{875}= \sigma ^2(c_4)\) on \([\frac{34}{81}, c_3)\). On \([c_4, \frac{33}{81})\), we have \(\langle 16 a\rangle \le \langle 81 a\rangle \) and see \(D X_2\) is decreasing and \(|D Y_2(a)| \le \frac{1}{324}\). Hence we have \(D W_2(a) \le D X_2(c_4) + \frac{1}{324}< 0\) and \(\sigma ^2(c_4) = W_2(c_4) \ge W_2(a)\ge \sigma ^2(a)\) on \([c_4, \frac{33}{81})\). Note that \(Y_2\le \frac{1}{90,720}\). On \([\frac{33}{81}, \frac{27}{65})\), by \(\langle 16 a\rangle \ge \langle 81 a\rangle \) we see that \(X_2\) equals to a quadratic function having axis at \(\frac{881}{2160}\). Hence \( \sigma ^2(a) \le X_2(\frac{881}{2160}) + \frac{1}{90,720}< \sigma ^2(c_4) \). On \([\frac{27}{65}, \frac{34}{81})\), by \(\langle 16 a\rangle < \langle 81 a\rangle \), \(X_2\) is decreasing. Hence \(\sigma ^2(a) \le X_2(\frac{27}{65})+ \frac{1}{90,720}< \sigma ^2(c_4)\) for \(a\in [\frac{27}{65}, \frac{34}{81})\).
3.2 \([c_3, 1/2)\) part
Let \(a\in [c_3, 1/2)\). We have \(\langle pa\rangle = pa-(p-1)/2\) and \(\langle qa\rangle = qa-(q-2)/2\). We see \(\langle qa\rangle - \langle pa\rangle = -(p-q)a + (p-q+1)/2 > 1/2\). Since \(W_3(a)\) maximizes at \(a_1:= (3p-1)/6p\), we have \( W_3(a) \le W_3(a_1) \). By
we have \( \sigma _{}^2(a) < U_{} \) for \(a\in [c_3, 1/2)\).
We derive \(A_2(p,q)>0\) from \(p/q\ge 9/4\). Put \(\widehat{A}_2(p,q)=(8 q^2 - 6)p^2 +(- 16q^3+4q)p -q^4\). We see
We see \( \widehat{A}_2(p,q)\) is increasing in p if \(p\ge 2q\). By \(\widehat{A}_2((9/4)q, q)= q^2(28q^2-171)/8>0\) (\(q\ge 4\)), and by \(\widehat{A}_2(5,2)>0\), the proof is over.
3.3 \([0, c_1)\) part
We consider on \([0, c_1)\). We assume \(q\ge 4\), since \(c_1=0\) otherwise. Since V(a, a) increases on \([0, c_1)\) and \(W_4(a)-V(a,a)\) has period 1 / q, the first equality of
holds. On \([c_1-1/q, c_1)\), we have \(\langle q a\rangle = qa-(q-4)/2\), and we see Z equals to a quadratic function having axis at \(a_2:=(2q+p-6)/(4q+2p)\). By \(a_2-c_1 = (p-q)/q(2q+p)>0\), we see that it is increasing on \([c_1-1/q, c_1)\). Since \(S_q(a)\) has period \(1/q^2\), we verified the second equality of (12). On \([c_1-1/q^2, c_1)\), we have \(\langle q^2 a\rangle = q^2 a-q(q-2)/2+1\). We see that \(W_4(a)\) equals to a quadratic function with axis \( a_3:=({2 q^2 + (2 p - 4) q + p^2 - 6 p - 2})/({4 q^2 + 4 p q + 2 p^2}) \). By
we have the third equality. The rest is proved by
by noting
3.4 \([c_1, c_2)\) part
Considering on \([c_1, c_2)\), we have \(\langle qa\rangle = qa-(q-2)/2\). Since Z equals to a quadratic function with axis \(a_4:=(2q+p-2)/(4q+2p)\), and since \(a_4 - c_2 = (6q+p)/2p(2q+p)>0\), it is increasing on \([ c_1, c_2)\).
First, we consider the case when \(2q\le p\le \frac{3}{2} q^2 + q\) holds. Since \(S_p \) has period \(1/p^2\), and since
we have the first equality of
On \([c_2-1/p^2, c_2)\), we have \(\langle p^2 a\rangle = p^2 a -p(p-3)/2+1\), and we see that \(W_5(a)\) equals to a quadratic function with axis \(a_5:=(2q^3 + ( p - 2) q^2 + 2 p^3 - 6 p^2 - 2p)/(4 q^3 + 2 p q^2 + 4 p^3)\), hence we have the third inequality. We have the rest by
Actually, when \(q\ge 4\), we see
When \(q=2\), and \(p=5\) or 7, one can verify the above inequality.
Next, we consider the case \(p> \frac{3}{2} q^2 + q\). We have already seen that Z and \(W_6\) are increasing on \([c_1, c_2)\). Note that \(p/q > \frac{3}{2} q + 1 \ge 4\) and \((c_2 -1/p) - c_1 = (p/q-5/2)/p> 0\). We can verify
by \(8p^3 -(8q+8)p^2 \ge 8p^3 -12qp^2>0\) and
since \(g'(t)>0\) for \(t\ge 2\) and \(g(2)>0\). Hence we see
We consider on \([ c_2-1/p, c_2)\). We have \(\langle pa \rangle = pa-(p-5)/2\), \(\langle qa \rangle = qa-(q-2)/2\), and can see that \(\langle pa \rangle < \langle qa \rangle \) on \([ c_2-1/p, c_7)\) and \(\langle pa \rangle \ge \langle qa \rangle \) on \([ c_7, c_2)\). By recalling the bound \(|D Y_1| \le 2/p(p-1)\), we can verify that \(D X_1\) is decreasing on \([ c_2-1/p, c_2)\),
on \([ c_2-1/p, c_7)\), and
on \([ c_7, c_2)\). Actually, the first inequality is clear and the second is proved by
by \(\frac{p}{q} \ge 7\) if \(q\ge 4\). In case \(q=2\), we have
for \(p\ge 9\). Hence we see the second inequality of
Put \( \widetilde{U}:= Z(c_4) + S_q(c_4) < \sigma ^2(c_4) \). Because of \(0< q^2/2(p-q)< 3q^2/2(p-q)< 1\),
we see \(\langle q^2 c_4\rangle = 1-q^2/2(p-q)\) and \(\langle q^2 c_7\rangle = 1-3q^2/2(p-q)\). Hence we can calculate \(S_q(c_4)\) and \(S_q(c_7)\), and can verify the rest by
4 Even p and odd q
Assume that q is odd and p is even and positive. We may assume \(q\ge 3\).
4.1 \([ c_6, 1/2)\) part
We consider on \([ c_6, 1/2)\). We have \(\langle pa\rangle = pa-(p-2)/2\), and by \(c_5\le c_6\), \(\langle qa\rangle = qa - (q-1)/2\). Since \(\langle pa\rangle < \langle qa\rangle \) holds on \([ c_6, c_4)\), and \(\langle pa\rangle \ge \langle qa\rangle \) on \([ c_4, 1/2)\), we see that \(D X_1\) is decreasing on \([ c_6, 1/2)\). Recalling \(|D Y_1(a) | \le 2/p(p-1) \), we have
on \([c_6, c_4)\), and
on \([c_4, 1/2)\). Hence \( W_1(a) \le W_1(c_4)= \sigma ^2(c_4) \) for \(a\in [c_6, 1/2)\).
We derive \(A_3(p,q)>0\) from \(p/q\ge 4\). Because of \(p\ge 4q\) and p and q are relatively prime, we see \(p> 4q\) or \(p\ge 4q+2\). Hence we have \(A_3(p,q)= p^2(p-4q-1)+2pq+2q^2> 0\).
4.2 \([0, c_5)\) part
We consider on \([0, c_5)\). The condition (3) implies \(p\ge 2q+2\). Since V(a, a) is increasing on \([0, c_5)\) and \(W_4(a) - V(a,a)\) has period 1 / q, we have the first equality of
On \([c_5-1/q, c_5)\), we have \(\langle qa\rangle = qa -(q-3)/2\) and we see that Z equals to a quadratic function having axis at \(a_6:=({p+ 2q-4})/({2p+4q}) \). Because of
it is increasing on \([c_5-1/q, c_5)\). Since \(S_q\) has period \(1/q^2\), we have the second equality of (14). On \([c_5-1/q^2, c_5)\), we have \(\langle q^2a\rangle = q^2a -q(q-1)/2+1\) and we see that \(W_4(a) \) equals to a quadratic function having axis at \( a_7 := ({2 q^2 + (2 p - 2) q + p^2 - 4 p - 2})/({4 q^2 + 4 p q + 2 p^2}) \). Because of
\(W_4(a)\) is increasing on \([c_5-1/q^2, c_5)\) and the third equality of (14) is proved. The rest is by
4.3 \([c_5, c_6)\) part
On \([c_5, c_6)\), we have \(\langle qa \rangle = qa-(q-1)/2\) and we see that Z equals to a quadratic function having axis at 1 / 2. Hence it is increasing on \([c_5, c_6)\). Since \(S_p\) has period \(1/p^2\), we have the first equality of
On \([(c_6-1/p^2)\vee c_5, c_6)\), by \(\langle p^2a \rangle = p^2 a - p(p-2)/2+1\) we see that \(W_5(a)\) is a quadratic function having axis at \( a_8 := ({2 q^3 + p q^2 + 2 p^3 - 4 p^2 - 2 p}) /(4 q^3 + 2 p q^3+ 4 p^3)\). Hence we have the middle of (15). Rest is by
Finally, we prove \(A_4(p,q)>0\) by assuming
We decompose as \(3A_4(p,q)=h_1 + \cdots + h_7\). Here
since the last cubic function in p / q is increasing for \(p/q\ge 2\) and equals to 24 at \(p/q=2\),
References
Aistleitner, C.: On the law of the iterated logarithm for the discrepancy of lacunary sequences. Trans. Am. Math. Soc. 362, 5967–5982 (2010)
Aistleitner, C., Berkes, I., Tichy, R.: On the law of the iterated logarithm for permuted lacunary sequences. Proc. Steklov Inst. Math. 276, 3–20 (2012)
Dhompongsa, S.: Almost sure invariance principles for the empirical process of lacunary sequences. Acta Math. Hung. 49, 83–102 (1987)
Fukuyama, K.: The law of the iterated logarithm for discrepancies of \(\{\theta ^n x\}\). Acta Math. Hung. 118, 155–170 (2008)
Fukuyama, K.: A central limit theorem and a metric discrepancy result for sequence with bounded gaps, Dependence in probability, analysis and number theory. In: Berkes, I., Bradley, R., Dehling, H., Peligrad, M., Tichy, R. (eds.) A Volume in Memory of Walter Philipp, pp. 233–246. Kendrick press, Heber (2010)
Fukuyama, K.: Metric discrepancy results for alternating geometric progressions. Monatsh. Math. 171, 33–63 (2013)
Fukuyama, K.: A metric discrepancy result for the sequence of powers of minus two. Indag. Math. (NS) 25, 487–504 (2014)
Kesten, H.: The discrepancy of random sequences \(\{kx\}\). Acta Arith. 10, 183–213 (1964/1965)
Khintchine, A.: Einige Sätze über Kettenbrüche, mit anwendungen auf die theorie der diophantischen approximationen. Math. Ann. 92, 115–125 (1924)
Philipp, W.: Limit theorems for lacunary series and uniform distribution mod 1. Acta Arith. 26, 241–251 (1975)
Takahashi, S.: An asymptotic property of a gap sequence. Proc. Jpn. Acad. 38, 101–104 (1962)
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by J. Schoißengeier.
Dedicated to Professor Norio Kôno on his 77th birthday.
K. Fukuyama was supported by KAKENHI 24340017 and 24340020.
Rights and permissions
About this article
Cite this article
Fukuyama, K., Yamashita, M. Metric discrepancy results for geometric progressions with large ratios. Monatsh Math 180, 731–742 (2016). https://doi.org/10.1007/s00605-015-0791-y
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00605-015-0791-y