Convergent points for random power series on the unit circle

Marcus Michelen, Mehtaab Sawhney · 2025

Read Convergent points for random power series on the unit circle by Marcus Michelen, Mehtaab Sawhney (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-02729.txt.

Introduction

Marcus Michelen and Mehtaab Sawhney wrote a 2025 probability paper. It settles a question Erdős posed in 1961. The topic is random power series on the unit circle. Random signs multiply fixed coefficients. The issue is where the series converges on the circle.

Classic results show divergence almost everywhere when coefficients decay slowly. A stronger size condition gives divergence everywhere. Erdős asked if slower decay guarantees some convergence point. The authors prove a strong yes. When coefficients shrink faster than the threshold rate, the convergence set has Hausdorff dimension one. That is the largest possible size for a subset of the circle. So convergence points form a rich fractal set, not a few strays.

Catalog record

CREATORMarcus Michelen, Mehtaab Sawhney
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE56 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-02729

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Convergent points for random power series on the unit circle Marcus Michelen, Mehtaab Sawhney Abstract . Consider a random power series of the form P ⁡ ( z ) = ∑ n ≥ 1 ε n ​ a n ​ z n P(z)=\sum_{n\geq 1}\varepsilon_{n}a_{n}z^{n} where a n ∈ ℂ a_{n}\in\mathbb{C} are deterministic and ε n \varepsilon_{n} are chosen independently and uniformly at random from { ± 1 } \{\pm 1\} . Kolmogorov’s three-series theorem states that if ∑ n | a n | 2 = ∞ \sum_{n}|a_{n}|^{2}=\infty then P ⁡ ( z ) P(z) almost-surely diverges at almost every z z with | z | = 1 |z|=1 . Dvoretzky and Erdős proved in 1959 that if | a n | = Ω ⁡ ( 1 / n ) |a_{n}|=\Omega(1/\sqrt{n}) then in fact P P almost surely diverges at every | z | = 1 |z|=1 . Erdős then asked in 1961 if this is sharp, meaning that if | a n | = o ⁡ ( 1 / n ) |a_{n}|=o(1/\sqrt{n}) then there is almost surely some convergent point z z with | z | = 1 |z|=1 . We prove this in a strong sense and show that if a n = o ⁡ ( 1 / n ) a_{n}=o(1/\sqrt{n}) then in fact the set of convergent points of P P with | z | = 1 |z|=1 has Hausdorff dimension 1 1 . 1. Introduction Consider a random power series P(z)=\sum_{n\geq 1}\varepsilon_{n}a_{n}z^{n} where a_{n}\in\mathbb{C} are deterministic and \varepsilon_{n} are chosen independently and uniformly from \{\pm 1\} . A special case of Kolmogorov’s three-series theorem shows that for a fixed z with |z|=1 , P(z) converges almost-surely if and only if \sum_{n\geq 1}|a_{n}|^{2}<\infty . Fubini’s theorem then allows one to show that if \sum_{n\geq 1}|a_{n}|^{2}<\infty , then almost-surely P converges at almost-every point with |z|=1 . Seeking to better understand convergence of random series, a classic result of Salem and Zygmund [ 9 , Theorem 5.1.5] asserts that if \sum_{n\geq 1}|a_{n}|^{2}<\infty and R_{k}=\sum_{n\geq k}|a_{n}|^{2} satisfies \sum_{k\geq 1}^{\infty}\frac{R_{k}^{1/2}}{k(\log k)^{1/2}}<\infty then P is almost-surely uniformly convergent on the unit circle. Further, under mild conditions on the coefficient sequence, this result is if and only if. The purpose of this paper is to consider the case when we expect that the power series will not converge uniformly on the unit circle. Dvoretzky and Erdős [ 2 ] proved that if there exists a positive sequence c_{n} with \limsup_{n\to\infty}\frac{\sum_{j=0}^{n}c_{j}^{2}}{\log(1/c_{n})}>0 and that |a_{n}|\geq c_{n} then almost surely the series \sum_{n\geq 1}\varepsilon_{n}a_{n}z^{n} diverges everywhere on the unit circle |z|=1 . An important special case of this result is that when |a_{n}|\geq 1/\sqrt{n} , then almost surely the series \sum_{n\geq 1}\varepsilon_{n}a_{n}z^{n} diverge everywhere on the unit circle |z|=1 . This paper is concerned with the converse of the result of Dvoretzky and Erdős [ 2 ] . In particular, Erdős [ 4 , pg 254] (see also [ 1 , Problem #527] ) asked if when |a_{n}|=o(1/\sqrt{n}) is it true that almost all series \sum_{n\geq 1}\varepsilon_{n}a_{n}z^{n} have a point on |z|=1 which is convergent. (Erdős in particular writes that “Perhaps every series satisfying \cdots has this property.”) Our main result confirms the conjecture of Erdős. Theorem 1.1 . Let a_{n}\in\mathbb{C} with \lim_{n\to\infty}n^{1/2}|a_{n}|=0 and let \varepsilon_{n} be chosen independently and uniformly at random from \{\pm 1\} . Then almost surely there exists z with |z|=1 such that \sum_{n=1}^{\infty}\varepsilon_{n}a_{n}z^{n} converges. Our methods in fact deliver a substantially stronger conclusion regarding the set of points on the unit circle which converge. Theorem 1.2 . Let a_{n}\in\mathbb{C} with \lim_{n\to\infty}n^{1/2}|a_{n}|=0 and let \varepsilon_{n} be chosen independently and uniformly at random from \{\pm 1\} . Then almost surely the set of points with |z|=1 such that \sum_{n=1}^{\infty}\varepsilon_{n}a_{n}z^{n} converges has Hausdorff dimension 1 . In other words, if one considers a sequence a_{n} with \sum_{n\geq 1}|a_{n}|^{2}=\infty but \lim_{n\to\infty}n^{-1/2}|a_{n}|=0 , then the set of convergent points of P on the unit circle has Lebesgue measure 0 but Hausdorff dimension 1 . 1.1. Outline of proof of Theorem 1.1 The result of Dvoretzky and Erdős [ 2 ] —which built on earlier work of Dvoretzky [ 3 ] concerning covering the unit circle with random arcs—shows that if the coefficient sequence is a_{n}=n^{-1/2} , then the series \sum_{n=1}^{\infty}a_{n}\varepsilon_{n}z^{n} diverges everywhere on the unit circle. We now sketch a short proof of this special case, which is closely related to the argument in [ 2 ] and serves as motivation for much of our analysis. Suppose that for a fixed z , the series \sum_{n=1}^{\infty}a_{n}\varepsilon_{n}z^{n} converges. Then for every \varepsilon>0 , there exists N_{\varepsilon} such that for N\geq N_{\varepsilon} we have that the event (1.1) \mathcal{C}_{N}(z)=\left\{\inf_{N\leq j\leq N^{2}}\Big|\sum_{n=N}^{j}\varepsilon_{n}a_{n}z^{n}\Big|\leq\varepsilon\right\} holds. Note that the random variable Z=\sum_{n=N}^{N^{2}}\varepsilon_{n}a_{n}z^{n} has variance \mathbb{E}[|Z|^{2}]=\sum_{n=N}^{N^{2}}|a_{n}|^{2}\approx\log N . Thus the event in ( 1.1 ) can be compared to a Brownian motion staying inside the interval [-\varepsilon,\varepsilon] for time proportional to \log N . This happens with probability \lesssim e^{-\Omega(\varepsilon^{-2}\log N)}=N^{-\Omega(\varepsilon^{-2})} . Moreover, for z and z^{\prime} with, e.g, |z-z^{\prime}|\leq N^{-5} , the quantity in ( 1.1 ) changes by o(1) and so there are essentially only N^{O(1)} distinct points z relevant to ( 1.1 ). A union bound then shows divergence at all points on the unit circle. Recall that Theorem 1.1 states that if the coefficients decay even slightly faster than n^{-1/2} , then the series converges. Already, revisiting the event \mathcal{C}_{N}(z) tells us why one could hope for this behavior: if the coefficients a_{n} now satisfy |a_{n}|\leq\delta/\sqrt{n} , then the variance of the sum is now of order \asymp\delta^{2}\log N , implying that the probability a given z satisfies \mathcal{C}_{N}(z) is more like N^{-\Theta(\delta^{2}/\varepsilon^{2})} . A back-of-the-envelope calculation to estimate the correlations of the sum in ( 1.1 ) suggests that, \sum_{n=N}^{N^{2}}\varepsilon_{n}a_{n}z^{n} and \sum_{n=N}^{N^{2}}\varepsilon_{n}a_{n}w^{n} typically decorrelate once we have |z-w|\gg N^{-2} . Provided \delta\ll\varepsilon , one expects that in fact many points satisfy ( 1.1 ). Of course, for a point z to be a convergent point for \sum_{n}\varepsilon_{n}a_{n}z^{n} , it is not enough for ( 1.1 ) to hold for some large N . We must also have, for instance, the same holds when we look between coefficients in [N^{2},N^{4}] , and [N^{4},N^{8}] and so on. We saw \mathbb{P}(\mathcal{C}_{N}(z))\approx N^{-\Theta(\delta^{2}/\varepsilon^{2})} and hope that \mathcal{C}_{N}(z),\mathcal{C}_{N}(w) are nearly independent for |z-w|\gg N^{-2}. For any given point z that satisfying \mathcal{C}_{N}(z) , it is in fact unlikely for it to then satisfy, say, \mathcal{C}_{N^{2}}(z) . However, we now suddenly have more points to work with, as the events \mathcal{C}_{N^{2}}(z) and \mathcal{C}_{N^{2}}(w) are now approximately independent for |z-w|\gg N^{-4} , meaning we have roughly N^{4} many chances for some \mathcal{C}_{N^{2}} to hold while we only had around N^{2} chances for \mathcal{C}_{N} to hold. Further, if |\zeta-z|\ll N^{-2} , then if \mathcal{C}_{N}(z) holds it is quite likely that \mathcal{C}_{N}(\zeta) holds. This leads to a branching process sort of argument: look at points that satisfy \mathcal{C}_{N}(z) ; think of their “children” as points at distance \ll N^{-2} ; show that if we have many points satisfying \mathcal{C}_{N}(z) then we have many points satisfying \mathcal{C}_{N^{2}}(w) . This is a close approximation of our argument, with two tweaks. The first is to choose our scales to grow a bit slower than N,N^{2},N^{4},\ldots. We choose a sequence of

Read the complete text