Abstract
We prove a conjecture of Ambrus, Ball and Erdélyi that equally spaced points maximize the minimum of discrete potentials on the unit circle whenever the potential is of the form
where \(f:[0,\pi ]\rightarrow [0,\infty ]\) is non-increasing and convex and \(d(z,w)\) denotes the geodesic distance between z and w on the circle.
Similar content being viewed by others
1 Introduction and Main Results
Let \(\mathbb{S }^1:=\{z=x+{\text{ i }}y \mid x,y \in \mathbb{R },\,\,x^2+y^2 =1\}\) denote the unit circle in the complex plane \(\mathbb{C }\). For \(z,w\in \mathbb{S }^1\), we denote by \(d(z,w)\) the geodesic (shortest arclength) distance between z and w. Let \(f:[0,\pi ]\rightarrow [0,\infty ]\) be non-increasing and convex on \((0,\pi ]\) with \(f(0)=\lim _{\theta \rightarrow 0^+}f(\theta )\). It then follows that f is a continuous extended real-valued function on \([0,\pi ]\).
For a list of n points (not necessarily distinct) \(\omega _n=(z_1 ,\ldots , z_n)\in (\mathbb{S }^1)^n\), we consider the f-potential of \(\omega _n\),
and the f-polarization of \(\omega _n\),
In this note, we are chiefly concerned with the n-point f-polarization of \(\mathbb{S }^1\) (also called the nth f-Chebyshev constant of \(\mathbb{S }^1\)),
which has been the subject of several recent papers (e.g., [1, 2, 5, 6]).
In the case (relating to Euclidean distance) when
we abbreviate the notation for the above quantities by writing
The main result of this note is the following theorem conjectured by Ambrus et al. [2]. Its proof is given in the next section.
Theorem 1
Let \(f:[0,\pi ]\rightarrow [0,\infty ]\) be non-increasing and convex on \((0,\pi ]\) with \(f(0)=\lim _{\theta \rightarrow 0^+}f(\theta )\). If \(\omega _n\) is any configuration of n distinct equally spaced points on \(\mathbb{S }^1\), then \(M^f(\omega _n;\mathbb{S }^1)=M^f_n(\mathbb{S }^1)\). Moreover, if the convexity condition is replaced by strict convexity, then such configurations are the only ones that achieve this equality.
Applying this theorem to the case of \(f_s\) given in (4) we immediately obtain the following.
Corollary 2
Let \(s>0\) and \(\omega _n^*:=\{\mathrm{{e}}^{\mathrm{{i}}2\pi k/n}\,: k=1,2,\ldots ,n\} \). If \((z_1,\ldots , z_n)\in (\mathbb{S }^1)^n\), then
with equality if and only if \((z_1,\ldots , z_n)\) consists of distinct equally spaced points.
The following representation of \(M^s(\omega _n^*;\mathbb{S }^1)\) in terms of Riesz s-energy was observed in [2]:
where
Thus, applying the asymptotic formulas for \(\mathcal{E }_s(\mathbb{S }^1;n)\) given in [3], we obtain the dominant term of \( M^s_n(\mathbb{S }^1)\) as \(n\rightarrow \infty \):
where \(\zeta (s)\) denotes the classical Riemann zeta function and \(a_n \sim b_n\) means that \(\lim _{n \rightarrow \infty }{a_n/b_n} = 1\). These asymptotics, but for \(M^s (\omega _n^*;\mathbb{S }^1)\), were stated in [2]Footnote 1.
For s an even integer, say \(s=2m\), the precise value of \(M^{2m}_n(\mathbb{S }^1)=M^{2m}(\omega _n^*;\mathbb{S }^1)\) can be expressed in finite terms, as can be seen from formula (1.20) in [3].
Corollary 3
We have
where \(\alpha _j(s)\) is defined via the power series for \(\mathrm{sinc\,} z = (\sin \pi z)/(\pi z)\):
In particular,
The case \(s=2\) of the above corollary was first proved in [1, 2] and the case \(s=4\) was first proved in [5]. We remark that an alternative formula for \(\alpha _j(s)\) is
where \(B_j^{(\alpha )}(x)\) denotes the generalized Bernoulli polynomial. Asymptotic formulas for \(M^f_n(\mathbb{S }^1)\) for certain other functions f can be obtained from the asymptotic formulas given in [4].
As other consequences of Theorem 1, we immediately deduce that equally spaced points are optimal for the following problems:
and
with the solution to (8) being well-known. Furthermore, various generalizations of the polarization problem for Riesz potentials for configurations on \(\mathbb{S }^1\) are worthy of consideration, such as minimizing the potential on circles concentric with \(\mathbb{S }^1\).
2 Proof of Theorem 1
For distinct points \(z_1,z_2\in \mathbb{S }^1\), we let \(\widehat{z_1z_2}\) denote the closed subarc of \(\mathbb{S }^1\) from \(z_1\) to \(z_2\) traversed in the counterclockwise direction. We further let \(\gamma (\widehat{z_1z_2})\) denote the length of \(\widehat{z_1z_2}\) (thus, \(\gamma (\widehat{z_1z_2})\) equals either \(d(z_1,z_2)\) or \(2\pi -d(z_1,z_2)\)). Observe that the points \(z_1\) and \(z_2\) partition \(\mathbb{S }^1\) into two subarcs: \(\widehat{z_1z_2}\) and \(\widehat{z_2z_1}\). The following lemma (see proof of Lemma 1 in [2]) is a simple consequence of the convexity and monotonicity of the function f and is used to show that any n-point configuration \(\omega _n\subset \mathbb{S }^1\) such that \(M^f(\omega _n; \mathbb{S }^1)=M_n^f(\mathbb{S }^1)\) must have the property that any local minimum of \(U^f(\omega _n; \cdot )\) is a global minimum of this function (Fig. 1).
For \(\phi \in \mathbb{R }\) and \(z\in \mathbb{S }^1\), we let \(\rho _\phi (z):=e^{i\phi }z\) denote the counterclockwise rotation of z by the angle \(\phi \).
Lemma 4
([2]) Let \(z_1, z_2\in \mathbb{S }^1\) and \(0<\varepsilon <\gamma (\widehat{z_2z_1})/2\). Then with f as in Theorem 1,
for z in the subarc \(\widehat{\rho _\varepsilon (z_2) \rho _{-\varepsilon }(z_1)},\) while the reverse inequality holds for z in the subarc \(\widehat{z_1 z_2}\). If f is strictly convex on \((0,\pi ]\), then these inequalities are strict. If \(z_1=z_2\), then we set \(\widehat{z_1 z_2}=\{z_1\}\) and \(\widehat{z_2 z_1}=\mathbb{S }^1\).
We now assume that \(\omega _n=(z_1 ,\ldots , z_n)\) is ordered in a counterclockwise manner and also that the indexing is extended periodically so that \(z_{k+n}=z_k\) for \(k\in \mathbb{Z }\). For \(1\le k\le n\) and \(\Delta \in \mathbb{R }\), we define \(\tau _{k,\Delta }:(\mathbb{S }^1)^n\rightarrow (\mathbb{S }^1)^n\) by
If \(z_{k-1}\ne z_k\) and \(z_{k+1}\ne z_{k+2}\), then \(\tau _{k,\Delta }(\omega _n)\) retains the ordering of \(\omega _n\) for \(\Delta \) positive and sufficiently small. Given \(\varvec{\Delta }:=(\Delta _1,\ldots , \Delta _n)^T\in \mathbb{R }^n\), let \(\tau _{\varvec{\Delta }}\!:=\!\tau _{n,\Delta _n}\circ \cdots \circ \tau _{2,\Delta _2}\circ \tau _{1,\Delta _1}\) and \(\omega _n^{\prime }\!:=\!\tau _{\varvec{\Delta }} (\omega _n).\) Letting \(\alpha _k\!:=\!\gamma (\widehat{z_k z_{k+1}})\) and \(\alpha _k^{\prime }:=\gamma (\widehat{z_k^{\prime } z_{k+1}^{\prime }})\) for \(k=1,\ldots , n\), we obtain the system of n linear equations:
which is satisfied as long as \(\sum _{k=1}^n\alpha _k^{\prime }=2\pi \) or, equivalently, if \(\omega _n^{\prime }\) is ordered counterclockwise. Let
Then (10) holds if
in which case, the configurations
are all ordered counterclockwise. If the components of \(\varvec{\Delta }\) are nonnegative, then we may replace the ‘(1/4)’ in (11) with ‘(1/2)’.
Lemma 5
Suppose \(\omega _n=(z_1 ,\ldots , z_n)\) and \(\omega _n^{\prime }=(z^{\prime }_1 ,\ldots , z^{\prime }_n)\) are n-point configurations on \(\mathbb{S }^1\) ordered in a counterclockwise manner. Then there is a unique \(\varvec{\Delta }^*=(\Delta _1^*,\ldots , \Delta _n^*)\in \mathbb{R }^n\) so that
-
(a)
\(\Delta _k^*\ge 0, \quad k=1,\ldots , n\),
-
(b)
\(\Delta _{j }^*=0\) for some \(j \in \{1,\ldots , n\}\), and
-
(c)
\(\tau _{\varvec{\Delta ^*}}(\omega _n)\) is a rotation of \(\omega _n^{\prime }\).
Proof
The system (10) can be expressed in the form
where
It is elementary to verify that \(\ker A=(\text{ range }A)^\perp = \text{ span } (\varvec{1})\), where \(\varvec{1}=(1,1,\ldots ,1)^T\). Since \(\varvec{\beta }^T\varvec{1}=\sum _{k=1}^n (\alpha _k^{\prime }-\alpha _k)=0\), the linear system (13) always has a solution \(\varvec{\Delta }\). Let \(j\in \{1, \ldots , n\}\) satisfy \(\Delta _{j}= \min _{1\le k\le n} \Delta _k \). Then subtracting \(\Delta _{j} \varvec{1}\) from \(\varvec{\Delta }\), we obtain the desired \(\varvec{\Delta }^*\). Since \(\ker A=\text{ span }\,\varvec{1}\), there is at most one solution of (13) satisfying properties (a) and (b), showing that \(\varvec{\Delta ^*}\) is unique.
Part (c) holds as a direct result of the fact that both \(\omega _n\) and \(\omega _n^{\prime }\) are ordered counterclockwise.
Lemma 6
Let \(\Omega _n=(z_1,\ldots , z_n)\) be a configuration of \(n\) distinct points on \(\mathbb{S }^1\) ordered counterclockwise, and with f as in Theorem 1, suppose \(\varvec{\Delta }=(\Delta _1,\ldots , \Delta _n)\in \mathbb{R }^n\) is such that
-
(a)
\(0\le \Delta _k \le (1/2) \mathrm{sep}(\Omega _n)\) for \(k=1,\ldots , n\), and
-
(b)
there is some \(j\in \{1,\ldots , n\}\) for which \(\Delta _j=0\).
Let \(\Omega _n^{\prime }:=\tau _{\varvec{\Delta }}(\Omega _n)=(z^{\prime }_1,\ldots , z^{\prime }_n)\). Then \(\widehat{z^{\prime }_j z^{\prime }_{j+1}}\subset \widehat{z_jz_{j+1}}\) and
If f is strictly convex on \((0,\pi ]\) and \(\Delta _k>0\) for at least one k, then the inequality (14) is strict.
We remark that \(\Delta _k=0\) for all \(k=1,\ldots , n\) is equivalent to saying that the points are equally spaced.
Proof
Recalling (12), it follows from condition (a) that \((z^{(\ell )}_1,\ldots , z^{(\ell )}_n):= \omega _{n,\varvec{\Delta }}^{(\ell )}\) are counterclockwise ordered. Since \(\Delta _j=0\) and \(\Delta _k\ge 0\) for \(k=1,\ldots ,n\), the points \(z^{(\ell )}_j\) and \(z^{(\ell )}_{j+1}\) are moved at most once as \(\ell \) varies from 1 to n and move toward each other, while remaining in the complement of all other subarcs \(\widehat{z^{(\ell )}_k z^{(\ell )}_{k+1}}\), i.e.,
for \(k\in \{1,\ldots ,n\}\setminus \{j\}\) and \(\ell \in \{1,\ldots , n\}\). Lemma 4 implies that, for \(\ell =1,\ldots , n\), we have \(U^f(\omega ^{(\ell -1)}_n;z) \le U^f(\omega ^{(\ell )}_n;z)\) for \(z\in \widehat{z^{(\ell )}_j z^{(\ell )}_{j+1}}\) (where \(\omega ^{(0)}_n:=\omega _n\)) and the inequality is strict if \(\Delta _\ell >0\). Hence, (14) holds and the inequality is strict if f is strictly convex and \(\Delta _k>0\) for some \(k=1,\ldots , n\). \(\square \)
We now proceed with the proof of Theorem 1. Let \(\omega _n=(z_1,\ldots , z_n)\) be a non-equally spaced configuration of n (not necessarily distinct) points on \(\mathbb{S }^1\) ordered counterclockwise. By Lemma 5, there is some equally spaced configuration \(\omega _n^{\prime }\) (i.e., \(\alpha ^{\prime }_k=2\pi /n\) for \(k=1,\ldots , n\)) and some \(\varvec{\Delta ^*}=(\Delta _1^*,\ldots , \Delta _n^*)\) such that (a) \(\omega _n^{\prime }=\tau _{\varvec{\Delta ^*}}(\omega _n)\), (b) \(\Delta ^*_k\ge 0\) for \(k=1,\ldots ,n\), and (c) \(\Delta ^*_j=0\) for some \(j\in \{1,\ldots ,n\}\). Then (10) holds with \(\alpha _k:=\gamma (\widehat{z_k,z_{k+1}})\) and \(\alpha ^{\prime }_k:=2\pi /n\). Since \(\omega _n\) is not equally spaced, we have \(\Delta ^*_k>0\) for at least one value of k.
For \(0\le t\le 1\), let \(\omega _n^t:=\tau _{(t\varvec{\Delta ^*})}(\omega _n)=(z_1^t,\ldots , z_n^t)\) and, for \(k=1,\ldots , n\), let \(\alpha _k^t:=\gamma (\widehat{z_k^t z_{k+1}^t})\). Recalling (10), observe that
for \(0\le t\le 1\) and \(k=1,\ldots , n\), and so \(\text{ sep }(\omega _n^t)\ge t(2\pi /n)\). Now let \(0<t<s<\min (1,t(1+\pi /(nD)))\), where \(D:=\max \{\Delta _k : 1\le k\le n\}\). Then Lemma 6 (with \(\Omega _n=\omega _n^t, \varvec{\Delta }=(s-t)\varvec{\Delta ^*}\), and \(\Omega _n^{\prime }=\tau _{\varvec{\Delta }}(\Omega _n)=\omega _n^s\)) implies that \(\widehat{z_j^s z_{j+1}^s}\subseteq \widehat{z_j^t z_{j+1}^t}\) and that
where the inequality is sharp if f is strictly convex.
Consider the function
Observe that
for \(0<t<s<\min (1,t(1+\pi /(nD)))\). It is then easy to verify that h is non-decreasing on \((0,1)\). Since \(\omega _n^t\) depends continuously on t, the function h is continuous on \([0,1]\) and thus h is non-decreasing on \([0,1]\).
We then obtain the desired inequality
where the last equality is a consequence of the fact that \(\omega _n^{\prime }\) is an equally spaced configuration and so the minimum of \(U^f(\omega _n^{\prime };z)\) over \(\mathbb{S }^1\) is the same as the minimum over \(\widehat{z_j^{\prime }z_{j+1}^{\prime }}\). If f is strictly convex, then \(h(0)<h(1)\) showing that any optimal f-polarization configuration must be equally spaced. This completes the proof of Theorem 1. \(\square \)
Notes
We remark that there is a factor of \(2/(2\pi )^p\) missing in the asymptotics given in [2] for the case \(p:=s>1\).
References
Ambrus, G.: Analytic and probabilistic problems in discrete geometry. Ph.D. Thesis, University College London (2009)
Ambrus, G., Ball, K., Erdélyi T.: Chebyshev constants for the unit circle. Bull. Lond. Math. Soc. 45(2), 236–248 (2013)
Brauchart, J.S., Hardin, D.P., Saff, E.B.: The Riesz energy of the \(N\)th roots of unity: an asymptotic expansion for large \(N\). Bull. Lond. Math. Soc. 41(4), 621–633 (2009)
Brauchart, J.S., Hardin, D.P., Saff, E.B.: Discrete energy asymptotics on a Riemannian circle. Unif. Distrib. Theory 6, 77–108 (2011)
Erdélyi, T., Saff, E.B.: Riesz polarization inequalities in higher dimensions, J. Approx. Theory 171, 128–147 (2013)
Nikolov, N., Rafailov, R.: On the sum of powered distances to certain sets of points on the circle. Pac. J. Math. 253(1), 157–168 (2011)
Acknowledgments
We thank the referees for their helpful suggestions to improve the manuscript. This research was supported, in part, by the U.S. National Science Foundation under Grants DMS-0808093 and DMS-1109266.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Hardin, D.P., Kendall, A.P. & Saff, E.B. Polarization Optimality of Equally Spaced Points on the Circle for Discrete Potentials. Discrete Comput Geom 50, 236–243 (2013). https://doi.org/10.1007/s00454-013-9502-4
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00454-013-9502-4