Cities are often described as statistically self-similar. When researchers divide a metropolis into grid cells and track how many phones are present in each cell over time, the fluctuations follow power laws, meaning the same statistical pattern repeats across time scales. The exponent of these power laws also appears to change smoothly from the dense core to the sparse periphery. This proposal points out a confound that has not been examined. Every such measurement mixes two kinds of noise. One is the real fluctuation of people. The other is counting noise, which comes from observing only a small and uneven fraction of those people. Counting noise has a known, uninteresting signature: it is uncorrelated in time and shrinks like one over the square root of the count. Peripheral cells contain far fewer observations, so counting noise grows there. The measured exponent should then drift toward the value pure noise would give, independent of any change in urban dynamics. Part of the celebrated core-to-periphery gradient may therefore be produced by how densely the data were sampled.
We propose to separate these two components using controlled thinning of the raw observations. The separation also turns the usual grid-size dilemma into a calculation. Cells that are too fine are dominated by counting noise, and cells that are too coarse wash out detail. The scaling law itself tells us where the crossover between the two regimes sits, and that crossover depends on local sampling density. In this proposal, data density is the mechanism, power-law scaling is the object being measured, and the choice of spatial discretization is the practical decision that follows.
Core Claim
Let a grid cell contain a true population $S$, observed through a sampling rate $p$ (observations per person per time step). Let $c \ge 1$ be a dispersion factor capturing the fact that one device produces many correlated pings. Let $A$ be the amplitude and $\beta$ the exponent of the intrinsic fluctuation law. The relative variance we observe is approximately
$$v(S) = A\,S^{-2\beta} + \frac{c}{pS}.$$
The first term is the intrinsic anomalous scaling of coupled systems. Universal Features in the Growth Dynamics of Complex Organizations found $\beta \approx 0.15$ for firms and countries, where independent subunits would give 0.5. The second term is ordinary counting variance. Because $\beta < 1/2$, intrinsic fluctuations dominate only above a crossover size
$$S^{*} = \left(\frac{c}{pA}\right)^{1/(1-2\beta)}.$$
Below $S^{*}$, a cell mostly measures the sensor. The same mixture acts on the temporal exponent: counting noise is white and pulls the detrended fluctuation exponent $\alpha$ toward 0.5. We claim three things. First, naive per-cell exponents are biased toward noise values in proportion to how far a cell falls below $S^{*}$. Second, $\beta$, $A$ and $c$ can be estimated from the data by device-level thinning. Third, a cell-size rule of "keep each cell above a multiple of $S^{*}$" is a principled replacement for fixed minimum-count thresholds, and it adapts to local density.
Relation to Prior Work
The claim is motivated by The spatiotemporal scaling laws of urban population dynamics. That study applied temporally weighted detrended fluctuation analysis to gridded mobile pings in six metropolitan regions. It reported that $\alpha$ decays logarithmically with distance from the center, with average fits of $R^2 = 0.994$ for temporal scaling and $0.921$ for spatial scaling. It used one fixed regular grid and did not test whether ping density, which also falls with distance, explains part of the decay. We do not dispute its fits. We ask whether the fitted gradient survives once counting noise is removed.
Universal fluctuations in growth dynamics of economic systems and the Amaral et al. work above supply the method. Data collapse, which rescales fluctuations by size-conditional standard deviation, is what we apply after correction. Those studies worked with fully observed aggregates and never faced a sampling term. Multi-spatial Scale Event Detection from Geo-tagged Tweet Streams via Power-law Verification runs power-law tests inside quadtree cells and reports that 24 of 30 event series pass. It assumes cells hold enough tweets for the test to have power, and our crossover analysis predicts where that assumption fails.
On the density side, the closest system is AdaSTEM, from Adaptive Spatio-Temporal Exploratory Models. It splits a quadtree until cells reach a minimum observation count, which sizes cells inversely to data density. Our rule differs because its threshold is not a single count. It depends on the measured $\beta$, on the overdispersion $c$, and on what one wants to estimate. Wrong Model, Right Uncertainty constrains our goals: under density shift, honest intervals widen sharply, and naive model-trusting intervals can fall to 0% coverage. We therefore report intervals on corrected exponents rather than only point estimates. Hypothesis Testing for Class-Conditional Noise Using Local Maximum Likelihood names the "empty neighborhood problem", which is the extreme case of our regime below $S^{*}$.
On the discretization side, A Unified Model for Spatio-Temporal Prediction Queries with Arbitrary Modifiable Areal Units (One4All-ST) and UCTB treat the dependence of results on unit size and shape as a forecasting nuisance and handle it with hierarchical grids. Neither checks whether scaling properties are preserved across levels. We use them as the downstream testbed, not as competitors on the scientific question.
Research Program
The work has three stages. The first stage is a simulator. It generates a population field with a prescribed radial profile of $\alpha$, which can be flat or decaying, and a prescribed $\beta$ under aggregation. It then observes the field through a device-level sampling process with a spatially varying rate $p(r)$ and clustered pings per device. We use it to verify that the naive pipeline of the urban scaling study produces a spurious radial decay when the true $\alpha$ is flat but $p$ falls with distance. We also use it to check that our corrected estimator recovers the truth.
The second stage is the estimator. For each cell we subsample devices, not pings, at fractions $q \in \{1, 1/2, 1/4, 1/8, \dots\}$. Subsampling devices preserves the within-device clustering that makes the counting noise non-Poisson. We then fit how the observed fluctuation function varies with $q$. Extrapolating to $q \to \infty$, the hypothetical infinite-sampling limit, gives the intrinsic fluctuation function, from which we compute a corrected $\alpha$. The slope with respect to $q$ identifies $c/p$. Repeating the analysis over nested grid aggregations gives $\beta$ and $A$ via data collapse. Uncertainty comes from a spatial block bootstrap.
The third stage is the partition rule. Starting from a fine grid, we merge cells in a quadtree until each cell's estimated population exceeds $\kappa S^{*}$, with a single tuned constant $\kappa$. We randomize offsets of the partition, as AdaSTEM does, to soften boundary effects.
Baselines
The "do nothing" reference is the fixed fine regular grid with naive per-cell detrended fluctuation analysis, reproducing the pipeline of The spatiotemporal scaling laws of urban population dynamics. Its reported fits are temporal $R^2 = 0.994$ and spatial $R^2 = 0.921$. We do not count goodness of fit as a win, because a confounded gradient can fit well. For the scientific question, beating this baseline means the following on the simulator: it reports a radial decay when the true $\alpha$ is flat, and our estimator does not. On real data, it means the decay attributable to density shrinks by a margin larger than the bootstrap interval.
The strongest existing competitor for cell sizing is the AdaSTEM-style minimum-count quadtree. We give it the same data and the same randomized-ensemble smoothing. We tune its count threshold over the same grid of values and the same validation split we use to tune $\kappa$, so each method gets an equal tuning budget. No reported result was found on Lacuna for AdaSTEM's predictive accuracy as a number; the file only says it outperformed fixed-scale models. Our rule beats it if it gives lower bias in recovered exponents on the simulator and lower held-out forecasting error downstream. The gain must hold across cities with different sampling rates, because a single tuned count should transfer poorly when $p$ changes.
For downstream forecasting we feed each partition into One4All-ST on NYC Taxi. The reported reference there is a 10.6% RMSE improvement over multi-model baselines on specific NYC tasks, using about 20% of their parameters. For the NYC bus data we use the UCTB pipeline, whose reported result is an RMSE of 80.63 and MAPE of 0.194 after region generation. We count a win as lower RMSE than the same model trained on its default partition, with identical training budget.
Two ablations isolate the key ingredient. The first replaces device-level thinning with a plain variance-minus-mean Poisson correction. The second uses $S^{*}$ computed with $\beta$ fixed at 0.5. If either matches the full method, the measured overdispersion or the anomalous exponent is not doing work.
Why not just…?
Why not just subtract the Poisson variance from each cell, as in standard noise-corrected fluctuation scaling? This is the most serious alternative, and it is included as an ablation. Phones ping in bursts, so a cell's count variance exceeds its mean by the factor $c$, and a Poisson correction would under-correct in exactly the sparse cells that matter. Device-level thinning measures $c$ directly. If the simulator shows $c \approx 1$ on real ping structure, then the Poisson correction suffices. In that case we would report the simpler method and keep the substantive contribution, which is the test of whether the radial gradient is an artifact.
Why not just use AdaSTEM's tuned minimum-count threshold? Within one dataset, a well-tuned count threshold may perform as well as our rule. That would be fine, since $S^{*}$ is itself a count. Our rule is expected to win on transfer. Its threshold scales as $(c/pA)^{1/(1-2\beta)}$, so moving from Milan's telecom records to sparse Melbourne tweets should change it in a predictable way. A single count cannot follow that change. We test this by tuning on one city and evaluating on another. If transfer shows no advantage, the partition rule is not a contribution, and we will say so.
Why not just bootstrap confidence intervals on each cell's exponent? Bootstrapping exposes variance but not bias. Counting noise shifts $\alpha$ systematically toward 0.5, and resampling the same noisy observations reproduces that shift with tight intervals. The simulator with a flat true $\alpha$ checks this directly: naive bootstrap intervals should exclude the truth in sparse cells.
Proposed Experiments
The decisive experiment is a thinning test on the Telecom Italia Milan records, 57.4 million records, and on NYC Taxi trajectories, 36 million records. We thin core cells down to the observation density of peripheral cells and recompute $\alpha$. If core $\alpha$ falls toward peripheral values, density explains the gradient. If it stays put, the gradient is genuine urban structure, and the published law is strengthened by having survived a real test. Either outcome is publishable.
The second experiment addresses how results depend on unit size. We aggregate Milan and Beijing taxi grids at four or five nested resolutions and test whether corrected $\beta$ and $\alpha$ are invariant across resolution while naive ones drift. This directly answers the open question of whether published exponents depend on the grid.
The third experiment carries the analysis to sparse data. For the Melbourne tweets (about 920 thousand) and New York tweets (9.5 million), we compute $S^{*}$ per quadtree cell and ask whether the power-law event tests that fail fall disproportionately in cells below $S^{*}$. That would suggest some apparent non-events are cases where the test lacked power.
The fourth experiment is the downstream forecasting comparison on NYC Taxi and NYC bus described under Baselines. We report RMSE and MAPE with the same model and budget across partitions. Tencent and SafeGraph data are used only if access allows; the core claims rest on the Milan, NYC and Beijing data.
References
- Universal Features in the Growth Dynamics of Complex Organizations
- The spatiotemporal scaling laws of urban population dynamics
- Universal fluctuations in growth dynamics of economic systems
- Multi-spatial Scale Event Detection from Geo-tagged Tweet Streams via Power-law Verification
- Adaptive Spatio-Temporal Exploratory Models: Hemisphere-wide species distributions from massively crowdsourced eBird data
- Wrong Model, Right Uncertainty: Spatial Associations for Discrete Data with Misspecification
- Hypothesis Testing for Class-Conditional Noise Using Local Maximum Likelihood
- A Unified Model for Spatio-Temporal Prediction Queries with Arbitrary Modifiable Areal Units
- UCTB: An Urban Computing Tool Box for Spatiotemporal Crowd Flow Prediction