人类移动中的幂律是一种混合假象：来自疫情自然实验的证据
The power law in human mobility is a mixture artifact: evidence from a pandemic natural experiment
浏览论文内容
中文总结 AI 辅助
该研究利用COVID-19封锁自然实验，分析440万用户21亿次位移，发现人类移动的幂律是不同空间容器移动分布混合产生的假象，而非个体移动特征。
中文摘要 AI 辅助
近二十年来，人类移动被视为无标度的，位移分布呈现重尾特征，看似是截断幂律，其成因被归因于个体的莱维飞行。另一种对立观点认为，每个空间容器内的移动服从对数正态分布，而总体幂律是不同大小容器混合产生的假象。由于两种观点都能拟合相同的总体数据，相关争论难以定论。我们将COVID-19封锁作为自然实验，该实验会消除长途旅行但保留本地移动。我们分析了来自440万手机用户在三个不同时期的21亿次位移。总体指数在封锁期间从1.66上升至1.74。对封锁前的旅行人群进行重采样以匹配封锁人群，可单独重现该变化，因此需从个体层面判断问题。对数正态分类是稳定吸引子（81%的分类结果保持不变），而幂律分类具有脆弱性（仅32%的分类结果保持不变），且发生分类切换的用户是旅行范围收缩最严重的用户。匹配样本量后，我们发现单个用户的尾部分布拒绝幂律，而同等大小的混合样本通过幂律检验，说明重尾行为源于总体行为而非个体。尾部检验在全样本量下未发现幂律阈值，且明显的幂律在超过1万个数据点后消失。水平混合可重建总体数据（R²最高达0.98），回转半径在封锁期间收缩失败且恶化（变异系数CV从0.62变为0.76），且幂律陡化集中在旅行范围广的用户中（P=0.0003）。人类旅行的幂律是聚合的特征，而非个体移动的特征。
英文摘要
For nearly two decades, human mobility has been read as scale free. Displacement distributions follow heavy tails that look like truncated power laws, traced to individual Lévy flights. A rival account holds that movement within each spatial container is lognormal, and the aggregate power law is an artifact of mixing containers of different sizes. The two fit the same aggregate data, so the debate has been hard to settle. We use the COVID-19 lockdowns as a natural experiment that removes long-distance travel and leaves local travel intact. We analyze 2.1 billion displacements from 4.4 million mobile-phone users across three distinct periods. The aggregate exponent increases under lockdown, from 1.66 to 1.74. Resampling the pre-lockdown traveling population to match the lockdown population reproduces that shift on its own, so we must look at the individual level to decides the question. Lognormal classification is a stable attractor (81\% retained) while the power-law classification is fragile (32\% retained), and the users who switch are the ones whose travel range collapsed the most. Matching the sample size we find that single users' tails reject the power law and pooled mixtures of equal size pass, so the heavy tail behavior originates in the aggregate behavior and not on the individual. Tail tests find no power-law threshold at full sample size, and the apparent power law disappears above ten thousand points. A level mixture rebuilds the aggregate ($R^2$ up to 0.98), the radius-of-gyration collapse fails and worsens under lockdown (CV $= 0.62$ to $0.76$), and the steepening concentrates in wide-ranging users ($P = 0.0003$). The power law of human travel is a feature of aggregation, not of individual movement.
