Researchers who analyze large satellite constellations often treat the satellites as points scattered uniformly at random over a sphere. This modelling style is called stochastic geometry: the network is described as a random point pattern, and averages over that pattern give closed formulas. The formulas predict routing latency and the probability that a multi-hop path is available, and they reach those predictions in thousands of evaluations instead of the hundred million a brute-force simulation can need. The trouble is that these formulas have only been checked against Monte Carlo runs of the same idealized model. Systems papers that drive real orbits and real population maps through packet-level simulators tell a different story. Shortest-path routing loses most of its packets over crowded regions, and the losses come from congestion over cities, which a purely geometric model never sees.
This note proposes two things. The first is a trace-driven audit that measures exactly where the uniform-sphere predictions fail. The second is a repair that adds one object to the analytical model: the cluster tree of user demand. A cluster tree is the nested family of connected "high-demand islands" that appears when you threshold a density map at every level. Routing breaks when a path must cross a saturated region. At a given load, which regions are saturated, how large they are, and whether neighbouring islands have merged into a barrier are all questions about this tree. Our claim is that a stochastic geometry model which removes links inside the tree's components, using a probability set by link capacity, predicts trace-level availability and drop rates much better than either the uniform model or a model corrected only for where satellites concentrate. The model stays cheap enough to evaluate inside design sweeps.
Core Claim
The analytical LEO routing line models satellites as a homogeneous binomial point process on a sphere, meaning a fixed number of points placed independently and uniformly. Stochastic Geometry-Based Low Latency Routing in Massive LEO Satellite Networks uses this to bound nearest-neighbour relay latency. Ultra Reliable Low Latency Routing in LEO Satellite Constellations adds signal-quality coverage and a "zigzag" distance penalty. Reliability Analysis of Multi-hop Routing in Multi-tier LEO Satellite Networks extends the model to stacked shells. Satellite-Terrestrial Routing or Inter-Satellite Routing? compares ground-relayed paths with space-mesh paths. These papers motivate our work, because they show the model is tractable. They also leave the gap we target: each one ignores queuing and traffic, and each validates only against its own idealization. SkyLink and SatFlow supply the evidence that this gap matters. They use real orbital elements (Two-Line Element files propagated with PyEphem) and real demand: the Gridded Population of the World (GPWv4) map in SkyLink and YouTube traces in SatFlow. In those settings congestion dominates the outcome, but neither paper has an analytical model.
We claim that most of the error between uniform-sphere predictions and trace-driven outcomes comes from demand hotspots, not from satellite non-uniformity. The audit will test this rather than assume it. We further claim that the hotspot effect enters the model as a position-dependent thinning of the links. Thinning means each link is kept or dropped at random, and the drop probability depends on location. That probability is a function of the demand cluster tree. Let $D$ be the demand density on Earth's surface, let $C$ be per-link capacity, and let $\rho(x)$ be the local satellite density at location $x$. A link over $x$ is treated as saturated when the demand it must carry, roughly $D(x)/\rho(x)$, exceeds $C$. Saturated links are therefore exactly those lying inside the superlevel set $\{x : D(x)/\rho(x) \ge C\}$. As $C$ or the offered load changes, this set moves through the levels of the tree for the ratio $D/\rho$. A path's availability and its detour penalty then depend on the tree's components along the great-circle route: how many it crosses, how wide they are, and the level at which they merge. A single hotspot mask or a fixed grid cannot supply that information.
Research Program
The first component is the audit. We build a packet-level reference simulator in the style SkyLink describes. It propagates TLE orbits for Starlink-like, Kuiper-like and OneWeb-like shells, generates user traffic from GPWv4, and gives each link a finite buffer. We then compute the published analytical predictions for availability, latency and interruption probability, using the formulas from the four stochastic geometry papers at matching constellation parameters, and compare them with the simulator one source–destination pair at a time. Following CHB, which showed that aggregate clustering scores hide failures concentrated in structurally hard cases, we will not report a single mean error. We stratify errors by route type: paths over oceans, paths that cross one persistent hotspot, paths that cross merged hotspot chains, and high-latitude paths near the inclination band. CHB worked only on static image and tabular data. We carry its "decompose by structure" logic into a dynamic network setting, where the structure comes from the demand tree rather than from a homology fingerprint of embeddings.
The second component is the model. We replace the uniform point process with a two-part intensity. The first part is the satellite density, corrected for latitude using the classical result for circular orbits at fixed inclination: density climbs sharply near the latitude equal to the inclination. This is the "shell effect" that Wang et al. 2022 flag as an open issue. The second part is the tree-governed thinning described above. We estimate the cluster tree of $D/\rho$ with a consistent level-set estimator. Rates of convergence for the cluster tree supplies the robust single-linkage method and its rates. Choosing the prominence cut-off is a known weak point, and we automate it with the bootstrap confidence rule from AuToMATo. That paper reports 65% confidence as the best setting on 2D benchmarks; we treat that value as a starting point to recheck, not as a fact about our data.
The model is used through a small set of quantities read off the tree. For each route we record the number of components it crosses above the capacity level, the total length of the route inside them, and the merge level of adjacent components. These enter the existing great-circle availability and zigzag formulas as region-dependent link-survival probabilities. Within each region the calculation keeps the single- and double-integral form that makes stochastic geometry cheap.
The third component handles motion. Earth rotates under the shells and $\rho$ drifts with the orbits, so the ratio $D/\rho$ changes over time while the demand map itself does not. Re-estimating the tree from scratch at every snapshot would be wasteful, and Direction 1's open problem about spatial indexes under continuous updates applies here. We will test whether the tree's topology, meaning which components exist and when they merge, stays constant over windows of minutes while only the levels shift. If it does, the analytical model needs a structural update only when a merge or split event occurs. This would be a first, restricted answer to the open problem of tracking superlevel set trees under node mobility.
Baselines
The naive reference point is the homogeneous spherical binomial model exactly as published in Wang et al. 2022 and 2023, evaluated at the same constellation parameters. The ground truth for every baseline is our trace-driven simulator, so all models are compared on the same orbits, the same demand map and the same buffer sizes. On its own idealized Monte Carlo, Wang et al. 2023 report a deviation of 1.70% to 4.53% between their formulas and simulation. Wang et al. 2022 report nearest-neighbour routing efficiency of 99.44% for Starlink and 97.91% for Kuiper. We expect these numbers to degrade substantially under traces; we have no trace-level number for them, and none is reported on Lacuna.
The strongest analytical competitor is the satellite-only inhomogeneous model: the same formulas with the latitude-corrected intensity but no demand thinning. It isolates whether orbital clustering alone explains the error. It receives the same per-region integration and the same evaluation budget.
The isolating ablation is a single-threshold hotspot mask. It thins links inside one superlevel set of $D/\rho$ at the capacity level, chosen to minimize validation error, but uses no hierarchy and no merge structure. A second flat competitor clusters demand with $k$-means, as KMDUC does in Cluster-Based Multi-Agent Task Scheduling for SAGIN. It assigns each cluster a thinning rate, and we tune $k$ on the same validation split.
Prediction is the primary target. For readers who care about routing decisions, we also add a secondary comparison against SkyLink's decentralized bandit link manager, reimplemented in our simulator with its published design. SkyLink reports a 29% reduction in the weighted sum of average delay and drop rate compared with bent-pipe, and 95% lower drop rates than $k$-shortest paths. We do not expect a tree-informed nearest-neighbour relay to beat SkyLink outright. For that secondary result, beating it means matching its drop rate within a few points while requiring no online learning.
On the primary question, the tree model beats the baselines if it cuts per-pair absolute error in availability and drop rate by at least half relative to the satellite-only model on hotspot-crossing routes, with no loss on ocean routes. It must also clearly beat the single-threshold mask on routes that cross merged chains.
Why not just…?
Why not just correct the satellite density for inclination and stop? This is the most serious threat. The shell effect is real, and the classical latitude formula is cheap. If the audit shows that the satellite-only model already closes most of the trace-level gap, the demand tree is unnecessary and we will report that. Our expectation that it will not comes from SkyLink and SatFlow, where the large failures (drop rates 95–99% worse under shortest-path routing) concentrate where population concentrates, not where orbits concentrate. That is evidence, not proof. The stratified audit settles it directly by comparing error on hotspot-crossing routes with error on high-latitude ocean routes.
Why not just bin demand on a grid or use one hotspot threshold? On a raster, a single-threshold mask is a slice of the tree at one level. The question is whether the rest of the tree matters. It should matter when load changes, because islands that were separate merge into barriers, and a fixed mask then gives the wrong detour cost. The single-threshold ablation, evaluated across a sweep of offered loads with its threshold tuned at one load, tests exactly this. If it matches the full tree at every load, we drop the hierarchy claim.
Why not just simulate, or learn a policy like SkyLink? Simulation is the ground truth here, so we do not dispute its accuracy. The cost is the issue. The stochastic geometry line reports cutting evaluation from $10^8$ Monte Carlo iterations to $10^3$–$10^4$ (Satellite-Terrestrial Routing or Inter-Satellite Routing?), and that saving matters for sweeps over altitude, satellite count and capacity. A learned policy such as SkyLink manages links well, but it answers "what should this satellite do now," not "what availability will this design deliver." We will report wall-clock time per design point for the simulator and for our model, so readers can judge the trade.
Why not use deterministic time-expanded contact graphs? Orbits are predictable, but demand bursts and buffer overflow are not. A contact graph also gives no closed-form sensitivity to design parameters. We include deterministic snapshot shortest-path routing in the simulator as a reference policy, but we do not treat it as a prediction model.
Proposed Experiments
We will run three constellation regimes: a Starlink-like dense shell, a Kuiper-like medium shell, and a OneWeb-like sparse shell. Wang et al. 2022 report frequent structural interruptions for OneWeb, so it tests the model where the uniform approximation is already weak. Each regime uses TLE trajectories over a 24-hour window, GPWv4 demand, and offered loads swept from light to saturating. We hold out alternate hours for evaluation and tune every baseline's free parameters on the remaining hours with the same budget. The metrics are per-pair absolute error in path availability, end-to-end latency and drop rate, each reported within the route strata defined above, together with model evaluation time.
Three outcomes decide the proposal. First, the audit must show a large, structured gap between the published formulas and the traces; if the gap is small and uniform, the analytical line is vindicated and the repair is moot. Second, the tree model must beat the satellite-only model on hotspot-crossing routes. Third, it must beat the tuned single-threshold mask under load changes, which is the test of whether the hierarchy, and not just a hotspot indicator, carries the predictive signal. A further check examines whether tree topology stays stable over minute-scale windows. That determines whether the model can be updated only at merge and split events or must be rebuilt at every snapshot.
References
- Stochastic Geometry-Based Low Latency Routing in Massive LEO Satellite Networks
- Ultra Reliable Low Latency Routing in LEO Satellite Constellations: A Stochastic Geometry Approach
- Reliability Analysis of Multi-Hop Routing in Multi-Tier LEO Satellite Networks
- Satellite-Terrestrial Routing or Inter-Satellite Routing? A Stochastic Geometry Perspective
- SkyLink: Scalable and Resilient Link Management in LEO Satellite Networks
- SatFlow: Scalable Network Planning for LEO Mega-Constellations
- CHB: A Diagnostic Toolkit for Hardness-Aware Clustering Evaluation
- Rates of convergence for the cluster tree
- AuToMATo: An Out-Of-The-Box Persistence-Based Clustering Algorithm
- Cluster-Based Multi-Agent Task Scheduling for Space-Air-Ground Integrated Networks