[Paper Review] How Far Might We Walk at Random?

Steven R. Finch|arXiv (Cornell University)|Feb 13, 2018

Complex Systems and Time Series Analysis31 references3 citations

TL;DR

This paper investigates the expected maximum displacement of one-dimensional random walks under strong and weak reflection at the origin, deriving asymptotic expressions for the first and second moments of the reflected maximum $M_n$. For symmetric walks, $\mathbb{E}[M_n] \sim \sqrt{\pi n / 2}$ and $\mathbb{E}[M_n^2] \sim 2G n$, where $G$ is Catalan's constant; for asymmetric walks, $\mathbb{E}[M_n] = O(\ln n)$, with no known higher-order terms.

ABSTRACT

This elementary treatment first summarizes extreme values of a Bernoulli random walk on the one-dimensional integer lattice over a finite discrete time interval. Both the symmetric (unbiased) and asymmetric (biased) cases are discussed. Asymptotic results are given as the time interval length approaches infinity. Focus then shifts to such walks reflected at the origin -- in both strong and weak senses -- and related unsolved problems are meticulously examined.

Motivation & Objective

• To derive asymptotic expressions for the expected maximum displacement $M_n$ of a one-dimensional random walk under strong and weak reflection at the origin.
• To analyze the behavior of $\mathbb{E}[M_n]$ and $\mathbb{E}[M_n^2]$ in both symmetric and asymmetric random walk models with reflection.
• To identify open problems in higher-order asymptotic expansions and moment estimation for reflected walks, particularly in the asymmetric case.
• To extend results to lazy random walks (with zero increments) and Markov-modulated walks, exploring how structure affects the maximum displacement.

Proposed method

• Uses recurrence relations for joint probabilities of maximum and minimum values in asymmetric random walks, derived via combinatorial and generating function techniques.
• Applies computer algebra (Mathematica) for non-rigorous, experimentally-based derivations of moment asymptotics.
• Employs known results from Brownian motion and excursion theory to infer asymptotic behavior for reflected walks.
• Analyzes cycle maxima between zero crossings in reflected walks, modeling them as i.i.d. geometric random variables to estimate $\mathbb{E}[M_n]$.
• Considers generalized models including lazy walks (with $\mathbb{P}(X_i=0)=r$) and Markov-dependent increments to assess robustness of asymptotic results.
• Compares strong reflection ($S_j = |S_{j-1} + X_j|$) and weak reflection ($S_j = \max\{S_{j-1} + X_j, 0\}$), conjecturing differences in higher-order terms.

Experimental results

Research questions

• RQ1 What are the asymptotic behaviors of $\mathbb{E}[M_n]$ and $\mathbb{E}[M_n^2]$ for symmetric random walks under strong reflection?
• RQ2 How do the moments of the reflected maximum $M_n$ behave in the asymmetric case, and why is $\mathbb{E}[M_n] = O(\ln n)$ the best known bound?
• RQ3 Are there higher-order terms in the asymptotic expansions of $\mathbb{E}[M_n]$ and $\mathbb{E}[M_n^2]$ for reflected walks, and if so, what are they?
• RQ4 How do the results change under lazy random walks (with $\mathbb{P}(X_i=0) = r$), and what is the impact of deterministic timing of negative steps?
• RQ5 What is the effect of Markovian dependence in step increments on the distribution of the maximum displacement $M_n$?

Key findings

• For symmetric random walks under strong or weak reflection, $\mathbb{E}[M_n] \sim \sqrt{\pi n / 2}$ and $\mathbb{E}[M_n^2] \sim 2G n$, where $G \approx 0.9159655941$ is Catalan’s constant.
• In the asymmetric case ($p < q$), $\mathbb{E}[M_n] = O(\ln n)$, with no known higher-order terms, and $\mathbb{E}[M_n^2] = O(\ln^2 n)$, indicating bounded variance.
• For lazy random walks with $\mathbb{P}(X_i = 1) = \mathbb{P}(X_i = -1) = 1/3$, $\mathbb{E}[M_n]/\sqrt{n} \sim \sqrt{\pi/3} \approx 1.023$ and $\mathbb{E}[M_n^2]/n \sim (4/3)G \approx 1.221$.
• When negative steps are deterministic (e.g., every third step is $-1$), $\mathbb{E}[M_n]/\sqrt{n} \sim \sqrt{\pi/12} \approx 0.512$ and $\mathbb{E}[M_n^2]/n \sim (1/3)G \approx 0.305$, showing significant sensitivity to step timing.
• For persistent random walks with Markov-dependent increments, $\mathbb{E}[M_n^+] \sim \sqrt{\alpha/\beta} \left( \sqrt{2n/\pi} - \frac{1}{2}\sqrt{\alpha/\beta} \right)$, with variance scaling as $\frac{\alpha}{\beta}(1 - 2/\pi)n$.
• The cross-moment $\mathbb{E}[M_n^+ M_n^-]$ remains unknown for reflected walks, and no higher-order terms are known for the symmetric reflected case.