Ride-hailing, courier dispatch and similar platforms face a "quick or cheap" choice for every request. Matching a rider to a driver immediately keeps waits short, but the driver may be far away. Waiting lets more drivers appear nearby, which makes the eventual match cheaper, but the waiting itself is a cost. The theory of this trade-off almost always assumes a seamless space in which every location sees the same flow of new arrivals. Real cities have edges: coastlines, rivers, closed roads, the municipal limit. A request at the edge of the service area sees roughly half the arrival stream an interior request sees, and a request in a corner sees about a quarter, because the rest of its neighborhood lies outside the region. In sensor-network connectivity this kind of truncation is known to wreck predictions. This note asks whether it also changes how long a matching platform should wait, and for whom.
The claim is deliberately two-sided, and the first half is a prediction that the obvious idea may be unnecessary. A short scaling argument, given below, says that the standard distance-triggered "break-even" rule already waits about the right amount at the border when requests do not compete for the same partners. The argument also says the rule should fail in two-sided markets where border requests compete with interior requests for a shared, scarce driver pool. There, interior requests trigger first and take the drivers, and border requests are starved. We propose to test both halves. If the second holds, we propose a location-dependent stopping rule, which we call a "patience clock," that fixes the starvation without giving up total cost. Road closures create new internal borders as the day goes on, and we maintain the truncation map under those closures using decremental planar connectivity.
What the directions contribute
The boundary literature motivates the question. Results on Finite Wireless Sensor Networks splits a finite domain into interior, edge and corner zones and shows that edge and corner nodes drive disconnection. At n=100, asymptotic formulas that ignore borders misestimate disconnectivity by up to ten orders of magnitude. Toroidal wrap-around also distorts spectral epidemic thresholds (Spectral analysis of virus spreading). Cube Versus Torus Models constrains how far these effects reach: border effects vanish as density grows, because the border fraction shrinks. Any matching gap caused by borders should therefore shrink as the market thickens, and our experiments check for exactly that signature.
The matching literature supplies the objective and the competitors. Total cost is the sum of match distances plus the sum of waiting times (Min-Cost Bipartite Perfect Matching with Delays). Online Matching: Haste makes Waste! gives a randomized tree-embedding algorithm, "Stilt-walker," with ratio $O(\log^2 n + \log \Delta)$, where $\Delta$ is the ratio of largest to smallest distance. Online Matching with Delays and Stochastic Arrival Times analyzes the break-even rule, which matches two pending requests once their accumulated delay equals the distance between them. Quick or Cheap? Breaking Points in Dynamic Markets shows that waiting and matching cost can both be near-optimal only if match cost falls faster than $1/x^2$ in pool size $x$. In two dimensions, nearest-neighbor distance falls only as $x^{-1/2}$, so spatial markets sit firmly in the no-free-lunch regime. That paper uses one clearing rule for all agents. We ask whether the rule should depend on location.
The empirical bridge is Understanding Inequalities in Ride-Hailing Services Through Simulations. In a 20×20 grid calibrated to NYC taxi data, nearest-driver matching starves drivers on the outskirts, and the Gini coefficient of driver income rises to 0.2–0.45 under oversupply. A "Poorest" rule repairs this without losing revenue. That study has no waiting-cost model and no theory, but it shows that border starvation is real and that a simple fairness rule is a serious competitor.
Asynchronous neural maturation serves as a lens here, not as a method. Brain development stabilizes primary local circuits early and keeps long-range association circuits plastic far longer (Mapping the microstructure of human cerebral cortex; Structural and Maturational Covariance). Staged maturation in networks also changes what gets learned (Gradual sensory maturation). The transferable idea is that commitment speed need not be uniform across a spatially organized system. Different regions can run on different clocks, set by how much each gains from waiting. We borrow nothing more than this. The biology does not tell us what the clock rates should be; the scaling argument does.
Core Claim
Let a request at location $z$ face partner arrivals of rate $\lambda$ per unit area. Let $\phi(z) \in (0,1]$ be the fraction of its interaction disk that lies inside the service region: 1 in the interior, about 1/2 at an edge, about 1/4 at a corner, or measured on the road network. After waiting time $t$, its nearest available partner is at distance roughly $c/\sqrt{\phi(z)\lambda t}$, where the constant $c$ comes from area scaling of nearest neighbors. For a single request choosing when to stop, the cost is
$$ J(t) = t + \frac{c}{\sqrt{\phi(z)\lambda t}}, \qquad t^*(z) = \left(\tfrac{c}{2}\right)^{2/3} (\phi(z)\lambda)^{-1/3}, $$
so border requests should wait longer by a factor $\phi^{-1/3}$: about 1.26 at edges and 1.59 at corners. Break-even fires when $t$ equals the current distance. That gives $t = c^{2/3}(\phi\lambda)^{-1/3}$, which has the same $\phi^{-1/3}$ scaling and differs only by a constant. So in one-sided settings, break-even already adapts to borders, and a border-calibrated clock should gain little.
The gain should appear when requests compete for shared supply. A border request's best driver usually lies toward the interior, where other requests can claim it. Break-even ignores this competition hazard, so interior requests with short distances fire first. Our patience clock is an index rule. A request fires when the expected rate at which waiting improves its match falls below the sum of the delay rate and the competition hazard $h(z)$ times the cost of losing its current best partner. Here $h(z)$ is the rate at which competing requests near that partner would claim it. Both $h(z)$ and the improvement rate are estimated from the truncation-corrected local arrival intensities $\phi(z)\lambda$. What is new is the prediction that borders matter for delayed matching only through competition, and a stopping rule that accounts for that competition. A location-dependent waiting rule alone is not the contribution.
Research Program
The work has four parts. The first is the analysis above, extended to the bipartite setting. We will derive how $h(z)$ depends on $\phi$ and on the supply/demand ratio, and state the regime, undersupply near borders, in which break-even's border starvation should grow. The second part is the estimator. $\phi(z)$ and the local intensities are computed on the road graph as the reachable road length within the match radius, relative to an interior reference, smoothed over a rolling time window. The third part is maintenance under road closures. Closures are edge deletions in a planar road graph. When a deletion splits a component, every request near the split gets a smaller $\phi$. We detect splits with decremental planar connectivity (Optimal Decremental Connectivity in Planar Graphs), which offers $O(1)$ connectivity queries and $O(n)$ total time over all deletions, and update $\phi$ only inside affected regions. We expect this to affect latency, not decision quality, and we will report it that way. The fourth part is evaluation across synthetic and real markets, laid out below.
Baselines
The naive reference is greedy nearest-driver dispatch with zero waiting. It minimizes delay, carries the $\Omega(n)$ deterministic lower bound of Online Matching: Haste makes Waste!, and produced the border starvation in Understanding Inequalities.
The second baseline is a single global batch interval, and separately a single global patience multiplier on break-even. Each is tuned by grid search on held-out days, with the same tuning budget we give our own clock's two constants. This is the strongest "no spatial structure" competitor.
The third baseline, and the strongest competitor on this specific question, is the break-even rule of Online Matching with Delays and Stochastic Arrival Times. It is the strongest because our own analysis predicts it already handles borders in one-sided markets. We found no reported result on Lacuna for break-even total cost on a real spatial dataset, so we will establish its numbers ourselves on identical request streams.
The fourth baseline is Stilt-walker (Online Matching: Haste makes Waste!), run on a tree embedding of the same road metric.
The fifth baseline is Bokányi and Hannák's Poorest rule. Its reported reference point is a Gini coefficient of 0.2–0.45 for nearest matching under oversupply, which Poorest reduces without revenue loss. No exact Poorest Gini value is given in the file.
On Chengdu, the strongest full system is the Model-Based Modular Approach of Online Relocating and Matching of Ride-Hailing Services, which reports a 10.79% higher request completion rate than batch matching and 3.09% more completed requests than RL methods. It also relocates idle drivers, which our clock does not. The fair comparison is therefore our clock plugged in as its matching layer, against its own matching layer, with the relocation layer held fixed.
Three ablations isolate the key ingredient. The first sets $\phi\equiv 1$, which is the torus-calibrated clock. The second keeps $\phi$ but sets $h\equiv 0$, which should roughly reproduce break-even if the analysis is right. The third keeps $h$ but drops $\phi$. Beating break-even means lowering the ratio of border-zone to interior-zone total cost by at least half, while keeping platform-wide total cost (match distance plus delay) within 2% of the best tuned baseline, or below it.
Why not just…?
Why not just use break-even, since it already scales with distance? This is the main threat, and our own derivation partly endorses it. The one-sided experiment below is designed to confirm that break-even matches the patience clock to within a few percent. If break-even also matches it in the two-sided undersupplied regime, the honest conclusion is that borders do not require special clocks for delayed matching. We would publish that as a negative result that contrasts sharply with the connectivity literature.
Why not just tune one global waiting parameter? A global parameter can trade total cost against total delay, but the border/interior gap comes from a spatial difference in arrival intensity and competition, and one scalar cannot remove it. The tuned-global baseline, given the same search budget, measures exactly how much of the gain a single knob recovers.
Why not just apply a fairness rule like Poorest? Poorest targets driver income inequality and was shown to work in Understanding Inequalities. It does not model rider waiting cost, so it may move border starvation from drivers to riders. We report both rider-side border delay and driver Gini for every method. If Poorest wins on both, the clock adds nothing beyond it, and we will say so.
Why not just ignore borders, as torus models do? On a torus, all location-dependent rules coincide by construction. The torus control experiment tests whether any of our gains are artifacts. The scaling of the gap with market density tests whether they behave as genuine border effects should, given the asymptotic washout in Cube Versus Torus Models.
Why not just recompute connectivity from scratch after each closure? At city scale that may be fast enough. We run the from-scratch recomputation as an ablation and claim the decremental structure only for its measured latency advantage under high closure rates.
Proposed Experiments
Synthetic markets place Poisson riders and drivers on the unit square, a square with internal cut lines, and the torus. Sizes range from about 100 agents, the finite regime of Results on Finite Wireless Sensor Networks, up to tens of thousands. We vary the supply/demand ratio across undersupply and oversupply. Real-data experiments replay the NYC TLC 2013 trips on the 20×20 grid of Bokányi and Hannák, and on the road graph. On the Chengdu Gaode data used by Gao et al., subject to access, we add historical closures as edge deletions. Metrics are total cost relative to the hindsight-optimal min-cost matching on the same stream (an empirical competitive ratio), the border/interior cost ratio by zone, mean and tail rider wait by zone, driver income Gini, completion rate, and per-event decision latency.
Three checks decide the outcome. First, in one-sided synthetic markets, the patience clock and break-even should agree to within a few percent. A large gap there would falsify the scaling argument. Second, on the torus, all methods with tuned constants should coincide. On bounded domains, the clock's advantage over break-even should appear only in undersupply and should shrink roughly in proportion to the fraction of requests within one match radius of a border, which falls as $n^{-1/2}$ for uniform arrivals. Third, under mid-day road closures that create internal borders in NYC and Chengdu, the clock with up-to-date $\phi$ should cut the border/interior cost ratio by at least half relative to both break-even and its own $\phi\equiv1$ ablation, with no more than a 2% increase in total cost. If the first two checks pass and the third fails, borders matter for connectivity but not for delayed matching.
References
- Results on finite wireless sensor networks: Connectivity and coverage
- Spectral analysis of virus spreading in random geometric networks
- Cube Versus Torus Models and the Euclidean Minimum Spanning Tree Constant
- Min-Cost Bipartite Perfect Matching with Delays
- Online matching: haste makes waste!
- Online Matching with Delays and Stochastic Arrival Times
- Quick or Cheap? Breaking Points in Dynamic Markets
- Understanding Inequalities in Ride-Hailing Services Through Simulations
- Mapping the microstructure of human cerebral cortex in vivo with diffusion MRI
- Structural and Maturational Covariance in Early Childhood Brain Development
- Gradual sensory maturation promotes abstract representation learning
- Optimal Decremental Connectivity in Planar Graphs
- Online relocating and matching of ride-hailing services: A model-based modular approach