Ride-hailing platforms that try to equalize driver income over a week face a familiar trap. If they send whichever driver has earned least, pickups become long and costly. If they send the nearest driver, income piles up among drivers who happen to sit near demand. We propose a middle route taken from the physics of liquids. The idea is to specify how idle drivers should be arranged relative to one another, then use liquid-state integral equation theory to compute the soft pairwise "repulsion" that produces that arrangement. Integral equation theory is the set of equations that link the forces between particles to how the particles end up spread out.
Each idle driver repositions by following this repulsion, using only the positions of nearby drivers. Dispatch then gives each request to the most under-paid driver within a short pickup radius. The spatial arrangement is what keeps "most under-paid nearby" from becoming "far away." The whole policy is closed-form: it needs no training and no city-scale optimization. It can also be recomputed every hour as fleet density changes.
Core Claim
Equal-income dispatch fails in practice mainly because of coverage fluctuations. Some requests have no idle driver nearby, and others have many. The long pickups caused by worst-off-first dispatch (Sühr et al. report passenger waits rising by up to 284%) are the symptom.
The number of idle drivers inside a pickup radius has a variance fixed by the fleet's pair correlation $g(r)$. This is the probability, relative to random placement, of finding another idle driver at distance $r$. The long-range part of that variance is controlled by the structure factor $S(k)$, the Fourier-space version of the same information. The Ornstein–Zernike equation gives $S(k) = 1/(1-\rho \hat c(k))$, where $\rho$ is idle-driver density and $\hat c$ is the Fourier transform of the "direct correlation," the part of the correlation not passed on through third drivers.
We therefore choose a target $g(r)$ with two properties. It should have a correlation hole at the scale of the pickup radius, and $S(k\to 0)$ should be small, meaning every neighbourhood has a steady, small number of idle drivers. We invert this target through the hypernetted-chain closure to obtain a pair potential, $$\beta V(r) = h(r) - c(r) - \ln(1+h(r)), \qquad h = g-1,$$ where $\beta V$ is the dimensionless repulsion drivers feel from one another. We then add debt-aware local dispatch.
The claim is that this policy reaches the fairness–wait trade-off of an Integer Linear Program (ILP), the exact but slow optimization used by Sühr et al., at a small fraction of its per-decision compute. It should also clearly beat hand-tuned repulsions of the same functional family.
Why expect the inversion to be accurate here? Penetrable, soft-core fluids become more mean-field-like as density rises. That result is from Hansen and co-workers on soft mixtures (Finken et al.; Hansen et al.). Bounded potentials inverted this way have also been shown to reproduce structure while making simulations much cheaper (Finken et al.). A dense urban fleet with bounded, overlapping repulsions is exactly that regime. These papers supply the method; we move it from polymer coils to vehicles.
How This Differs From What Exists
The closest work imports physical pair potentials into optimization. Pair-potential energy functions in multi-objective optimization spreads candidate solutions along a static trade-off curve using pre-chosen Coulomb or Gaussian repulsions. It tunes their single strength parameter with Newton iteration or a neural network. We instead derive the whole shape of $V(r)$ from a target structure through Ornstein–Zernike inversion. We also apply it to moving agents with a fairness state. That paper is our natural ablation.
The amortized-fairness literature motivates the problem but offers no spatial mechanism. Sühr et al. solve repeated ILPs and note that these are intractable at city scale. MOMAQL learns multi-agent Q-functions on a zone graph with a hand-tuned fairness weight. LoTFair gives primal-dual guarantees on accumulated unfairness, but only under convexity. BankFair and Equity of Attention amortize exposure in rankings, where space plays no role. None of these shapes where idle agents wait.
Johnson et al. and Representability problems for coarse-grained water potentials show that a potential which matches $g(r)$ is unique and cannot also match other properties. This paper constrains the proposal. Our potential controls coverage statistics only. It cannot, by construction, also be tuned to hit a throughput target, so the fairness–efficiency trade-off must be set through dispatch rules and the choice of target $g$, not the potential.
Restricted cubic splines enter in a secondary, constraining role. These are piecewise cubics that are forced to be straight lines beyond the outermost breakpoints. Clinical studies use them to find tipping points such as the TyG-index threshold near 9.0 in Tao et al. and the J-shaped stress-hyperglycemia curve in the MIMIC-IV/CIN-II study.
Amortized fairness assumes a driver can be "paid back later." If the hazard of leaving the platform accelerates past some earnings deficit, borrowing beyond that point breaks the contract permanently. We fit such a hazard curve and use its inflection as a debt ceiling. We also use the same spline basis, in distance $r$, to regularize the numerically inverted potential. For the potential we add one extra condition, zero slope beyond the outer breakpoint, so the potential vanishes at long range.
Research Program
The work has four parts.
Build the simulator. We build an event-driven dispatch simulator from trip records. Idle drivers move with small overdamped steps down the gradient of the summed pair potential, plus noise. Requests go to the highest-debt driver within radius $R$, falling back to the nearest driver otherwise. Debt is the running gap between a driver's net earnings and the fleet mean, which is the accumulator used by Sühr et al.
Choose the target structure. Within zones of roughly uniform demand (a local-density approximation), we specify target $g(r)$ as a smooth correlation hole of width near $R$. Its depth is chosen so that the predicted count variance, obtained from $S(k)$, gives one to three eligible drivers per request. We solve the two-dimensional Ornstein–Zernike equation with Hankel transforms, invert through the hypernetted-chain closure, and refine with a few iterative corrections in the spline basis. This is repeated every hour as $\rho$ changes, because inverted potentials do not transfer across densities.
Diagnose the physics. Two questions matter here. First, does the non-equilibrium fleet, in which particles vanish on dispatch and reappear at drop-offs, actually reproduce the target $g(r)$? Second, what happens with separate debt classes? If we use a two-species mixture, following Finken et al., we must avoid cross-repulsions that drive demixing. Demixing would cluster high-debt drivers together. It would be flagged by the closure failing to converge near a spinodal, the density at which a mixture becomes unstable and separates.
Fit the churn curve. Where driver identities allow, we fit a Cox model with a restricted cubic spline in cumulative earnings deficit. The outcome is multi-day absence, used as a proxy for exit. The estimated inflection becomes the debt ceiling.
Baselines
The naive reference is Nearest Driver First. On the Asian-city data of Sühr et al., it leaves the bottom 50% of drivers with 28% of income. Every method must beat it on the Generalized Entropy Index (GEI, the income-inequality measure used by Sühr et al.) without raising wait times materially.
The opposite extreme is Worst-Off Driver First, the reference for the long-pickup failure. We report its wait inflation and its share of trips where costs exceed the fare (negative-utility trips).
The strongest existing method on this question is the fairness-weighted ILP of Sühr et al. They report that ILP1.0 reduced GEI by 13.1% relative to Nearest Driver First and improved passenger utility by about 4%. We rerun it in our simulator with the same matching windows, the same cost model, and the same driver cohort. We sweep its fairness weight so that it traces a full trade-off curve. We record per-window solve time on identical hardware. Beating it means our curve lies on or above the ILP's at equal wait time, and at least an order of magnitude cheaper per decision. Matching it at far lower cost would already be a useful result.
On NYC Manhattan data we compare against MOMAQL Long-term Fairness in Ride-Hailing Platform. It reports total utility 95,823.79 and fairness variance 85,194.48 using a three-week train / one-week test split. We reuse that split and report both metrics. We give MOMAQL's weight $\omega$ the same tuning budget, in simulator hours, as our target-$g$ search.
The key ablation keeps the dispatch rule and the potential's family fixed but removes the inversion. We use a Gaussian repulsion whose single strength parameter is tuned by grid search per hour, in the style of the multi-objective pair-potential paper. A second ablation removes repositioning entirely, leaving debt-aware local dispatch alone; this isolates the contribution of spatial structure.
Why not just…?
Why not just tune a simple Gaussian repulsion instead of inverting a target structure? This is the most serious threat. The multi-objective pair-potential paper shows that the best repulsion strength shifts with problem geometry, which suggests a one-parameter family may not track changing fleet densities. That is a hint, not proof. Our key ablation settles it directly. If a per-hour tuned Gaussian matches the inverted potential's trade-off curve, we will report that the value lies in "any bounded repulsion plus debt-aware local dispatch." The inversion would then be only a cheap tuning-free way to set it.
Why not just add debt weights to nearest-driver matching, as in primal-dual methods like LoTFair? Debt weights decide who among nearby drivers gets the request. They do nothing about whether anyone is nearby, and that is where Worst-Off Driver First loses. The no-repositioning ablation tests this. If debt-aware dispatch alone reaches the ILP curve, the spatial component is unnecessary.
Why not just learn a repositioning policy with reinforcement learning? MOMAQL does this with good results. Its ablation shows that fairness variance nearly doubles, to about 153k, without demand forecasting, which indicates sensitivity to look-ahead and tuning. Our approach needs no training and recomputes hourly in closed form. We compare on MOMAQL's own split. We expect to trade some raw utility for robustness and cost, and we will say so if the gap is large.
Why not just use a fixed debt cap instead of a spline-estimated churn threshold? A grid-searched cap may well perform equally well in simulation, because the simulator's churn model is itself an assumption. The spline's advantage is that the cap is estimated from observed behaviour rather than tuned against the simulator. We report both. We treat the churn component as a sensitivity analysis if exits cannot be identified in the data. No reported result on driver churn was found on Lacuna for either taxi dataset.
Proposed Experiments
We use the two ride-hailing datasets in the file. The Asian-city platform data covers 1,462 drivers and more than 230,000 requests over 29 days; access must be confirmed with the authors of Sühr et al. The NYC Manhattan March 2016 data follows the MOMAQL protocol. The metrics are GEI of net driver income, mean and 95th-percentile passenger wait, share of negative-utility trips, weekly fairness variance, per-decision compute, and simulated exits under the fitted hazard.
Three checks decide the proposal.
The first is structural. The simulated fleet's measured $g(r)$ and neighbourhood-count variance should match the targets within a few percent at typical daytime densities. If the dispatch churn of drivers disappearing and reappearing breaks the equilibrium mapping, the physics component fails regardless of the fairness numbers.
The second is the trade-off curve against the ILP and against the tuned-Gaussian ablation, at matched wait times.
The third is transferability across time of day. We compare an hourly re-inverted potential with one inverted once at peak density. The integral-equation literature predicts that the fixed potential degrades off-peak. Observing that degradation, together with its removal by cheap re-inversion, is the concrete payoff that neither learned nor hand-tuned alternatives provide for free.
References
- Two-Sided Fairness for Repeated Matchings in Two-Sided Markets: A Case Study of a Ride-Hailing Platform
- Phase Separation of Penetrable Core Mixtures
- Can Polymer Coils Be Modeled as “Soft Colloids”?
- Coarse-graining polymers as soft colloids
- On the utilization of pair-potential energy functions in multi-objective optimization
- Long-Term Fairness in Ride-Hailing Platform
- Long-Term Fairness for Real-Time Decision Making: A Constrained Online Optimization Approach
- Guaranteeing Accuracy and Fairness under Fluctuating User Traffic: A Bankruptcy-Inspired Re-ranking Approach
- Equity of Attention: {A}mortizing Individual Fairness in Rankings
- Representability problems for coarse-grained water potentials
- Prognostic value of triglyceride-glucose index in patients with chronic coronary syndrome undergoing percutaneous coronary intervention
- Impact of stress hyperglycemia ratio on mortality in patients with critical acute myocardial infarction: insight from american MIMIC-IV and the chinese CIN-II study