Noise-Aware Angular Suppression for Broadcast Storms

When a car brakes hard on a highway, it floods a warning to nearby vehicles, and each one decides whether to repeat it. If every vehicle repeats, the shared radio channel jams: the "broadcast storm." If too many stay silent, the warning never reaches the vehicles that need it. The best simple fix uses geometry. A vehicle stays quiet if it overhears a relay that lies in roughly the same direction from the original sender. ODAM-C does this with a fixed 90° test. That test trusts GPS positions exactly. Under ordinary positioning error, the angle between two nearby vehicles is close to random, so the suppression decision is close to random too.

We propose replacing the single-relay angle test with a small statistical test that each vehicle runs locally. The vehicle asks: "Given noisy positions of every relay I have overheard, how likely is it that the direction I would cover is already covered?" It retransmits only if that probability is low. We then distill the test into a two-branch linear rule cheap enough to run on every packet. The second step uses the symbolic-distillation recipe of Knowledge-Guided Machine Learning for Stabilizing Near-Shortest Path Routing. The sterile-neutrino literature enters in a limited role, as a template for inference and reporting. It is not a source of network mechanism.

Core Claim

Angular suppression rules fail under location noise for a specific reason. The error in a measured angle grows as the measured distance shrinks. Suppose two positions each carry independent isotropic error with standard deviation $\sigma$ per axis and are separated by distance $d$. To first order, the angle between them has standard deviation

$$\sigma_\theta \approx \frac{\sqrt{2}\,\sigma}{d}.$$

This follows from the small-angle approximation: the perpendicular component of the difference of two independent errors has standard deviation $\sqrt{2}\sigma$. With metre-scale GPS error and relays a few metres apart in dense traffic, $\sigma_\theta$ is tens of degrees. ODAM-C's 90° boundary is then misjudged exactly in the dense regime where suppression matters most.

The new mechanism is a coverage decision that combines evidence from all overheard relays and weights each one by its angular uncertainty. A relay is a vehicle that has already repeated the warning. Let $\phi$ be the direction from the original sender to me. Let $\hat\theta_j$ be the measured direction to overheard relay $j$, with uncertainty $\sigma_{\theta,j}$. Let $w$ be a sector half-width. I suppress when

$$\Pr[\text{sector}(\phi, w)\ \text{covered}] = 1 - \prod_j \Big(1 - \Pr\big[\,|\theta_j - \phi| \le w \mid \hat\theta_j, \sigma_{\theta,j}\big]\Big) \ge \tau.$$

Here $\tau$ is a target confidence. The product treats relays as independent coverers, which the experiments will stress-test.

Two things are new relative to ODAM-C. First, no single nearby relay can trigger suppression on the strength of an angle it cannot actually resolve. Second, several distant relays that agree can trigger suppression even when none of them alone is decisive.

Where Each Direction Fits

The routing direction supplies the noise model and the cheap-rule method. Geographic Routing Algorithm with Location Errors shows, for unicast routing, that weighing expected progress against angle under Gaussian, uniform and exponential location error beats proximity-only greedy forwarding. Its MEDA algorithm uses principal component analysis on each neighbor's position statistics. That per-neighbor matrix work is what we want to avoid at broadcast rates. We borrow its error models but apply them to a suppression decision, not to next-hop choice.

Knowledge-Guided Machine Learning for Stabilizing Near-Shortest Path Routing distills a learned forwarding policy into guarded linear commands, two-branch "if condition then action" rules. It reports two orders of magnitude lower operational complexity than the neural network, but it assumes exact coordinates. We reuse its distillation recipe on a noise-aware decision.

A data dissemination mechanism for motorway environment in VANETs motivates the work. Its tight angular constraints suppress loops on linear motorways, but it is untested on intersections, where angles and roads stop lining up. VANETs, short for vehicular ad-hoc networks, are the car-to-car setting throughout this note.

The broadcast-storm direction supplies the problem and the competitor. Position-Based Emergency Message Dissemination Schemes in the Internet of Vehicles: A Review names GPS degradation in tunnels and urban canyons as an unsolved weakness of position-based suppression. It also notes the lack of adaptive models balancing redundancy against reachability. That gap is the one we target.

The sterile-neutrino direction contributes a statistical lesson, and we are deliberately narrow about it. The research file proposes analogies such as "non-unitary leakage into sterile buffers" and "parametric resonance of beacons." We judge these decorative: each reduces to an existing networking idea, a deferred-retransmission buffer or a load-dependent collision rate, and gains nothing from the physics notation. The transferable lesson is the one in First Constraints on Light Sterile Neutrino Oscillations from Combined Appearance and Disappearance Searches with the MicroBooNE Detector. There, an appearance signal and a disappearance signal can partly cancel when one parameter is small and a background ratio is large. Fitting each channel alone hides the effect; fitting them jointly breaks the degeneracy.

Counter-based suppression has the same structure. A high duplicate count is equally consistent with "dense and my direction is covered" and "dense but all copies came from behind me." Combining the count with angular placement, weighted by angular confidence, is the joint fit that separates the two.

We also adopt the physics habit of reporting sensitivity as a region in parameter space rather than at a few operating points. Improved eV-scale sterile-neutrino constraints from the second KATRIN measurement campaign reports its reach as a contour in mass-squared and mixing. We will report each suppression rule as the region in the plane of location-noise level versus vehicle density where it holds a delivery target.

Research Program

The first component derives the coverage probability above in closed form for Gaussian, uniform and exponential position error. Each yields a different tail for $\Pr[|\theta_j-\phi|\le w]$. We will check the first-order $\sigma/d$ approximation against Monte Carlo, focusing on small $d$, where it breaks down and where the decision matters most.

The second component builds the full rule into a VANET simulator. Each vehicle keeps the list of overheard relays it already caches under ODAM-C's scheme, adds a per-relay distance-derived angular uncertainty, and evaluates the coverage probability when its contention timer expires. The rule also needs an estimate of $\sigma$. We consider two estimators: the receiver's own GPS dilution-of-precision report, or a fixed conservative value. Comparing them tests sensitivity to misspecified noise.

The third component distills the rule into a guarded linear command over two or three features, following the symbolic-regression procedure of Chen et al. Candidate features are the nearest same-sector relay distance, the count of relays with confident in-sector angles, and the noise estimate. We expect the distilled rule to look like "suppress if $k$ confident in-sector relays, where confidence requires $d_j > c\,\sigma$." If it does, the contribution is partly a principled derivation of a margin that practitioners might otherwise hand-tune. We consider that an acceptable and checkable outcome.

Baselines

The naive reference is blind flooding: every vehicle repeats once. It gives the upper bound on reach and the lower bound on channel efficiency.

The classical temporal baseline is counter-based suppression in the style of Winstanley et al.. A vehicle cancels its scheduled repeat after overhearing a threshold number of duplicates. We tune the threshold over the same grid as our $\tau$.

Plain ODAM, which suppresses on any overheard duplicate, is included because Sun et al. report it collapsing in dense traffic.

The strongest existing method is ODAM-C, the 90° angle test plus its two-tier least-recently-used cache. In SUMO+NS-2 dense-traffic scenarios, An Optimal ODAM-Based Broadcast Algorithm for Vehicular Ad-Hoc Networks reports "nearly 100% PDR" against 50–55% for ODAM and WPBM. It also reports redundancy of 0.6 versus 0.1 for ODAM and about 0.5 ms added latency. PDR, the packet delivery rate, is the fraction of vehicles that receive the warning. All of these figures assume perfect positions. No reported result was found on Lacuna for ODAM-C under location error, so our first measurement is ODAM-C's own degradation curve.

We will also run a strengthened ODAM-C that the original paper did not evaluate. Its angle threshold is tuned per scenario, and it adds a fixed angular safety margin tuned on held-out runs. This is the honest form of the strongest competitor.

All methods share the same SUMO mobility traces, NS-2 802.11p channel and MAC settings, noise draws per seed, and an equal hyperparameter grid evaluated on validation seeds distinct from test seeds.

Two ablations isolate the key ingredients. The first keeps the multi-relay coverage product but sets $\sigma=0$, isolating the value of combining relays. The second keeps the $\sigma/d$ uncertainty but uses only the single nearest relay, isolating the value of noise awareness.

Beating ODAM-C means reaching its perfect-GPS PDR level, which we treat as 95% or higher, over a strictly larger region of the noise-density plane at equal or lower redundancy. A gain in PDR alone, bought with redundancy above ODAM-C's 0.6, does not count.

Why not just…?

Why not just tune ODAM-C's angle threshold or add a fixed margin? This is the most serious threat. A fixed margin ignores that angular error scales with $1/d$. It must be wide enough for the closest relays, which makes it over-conservative for distant ones and drives redundancy up. The strengthened-ODAM-C baseline tests this directly. If a tuned fixed margin matches our coverage region at equal redundancy, the distance dependence does not matter in practice, and we will report that.

Why not just use counter-based suppression, which needs no positions at all? It is immune to GPS error, which is a real advantage in tunnels. However, the count alone cannot distinguish "covered" from "dense but uncovered in my direction." Sun et al.'s ODAM collapse to 50–55% PDR in dense, diverging traffic is evidence of exactly this conflation. We expect counters to win only at very high noise. The contour plots will show where that crossover sits, and a hybrid that falls back to counters beyond it is a natural result.

Why not just run MEDA-style per-neighbor statistics? Geographic Routing Algorithm with Location Errors shows the value of noise modeling, but it is designed for unicast next-hop choice, and its per-neighbor matrix operations scale poorly with dense broadcast neighbor sets. Our full rule is in effect a broadcast analogue of MEDA. The distilled rule is the test of whether that modeling can be made cheap, so we will report per-packet operation counts alongside delivery.

Why not just bypass the multi-hop network with infrastructure? TC-PAA uses roadside micro base stations and neural traffic classification. It reports 12.288 mW for emergency communication against 75.776 mW for AODV, and cuts paths to 1–2 hops. It requires roadside units and camera feeds, which do not exist on rural roads or after disasters, the settings where ad-hoc dissemination is the only option. We treat it as out of scope rather than as a competitor.

Proposed Experiments

The main testbed follows ODAM-C: SUMO mobility with NS-2, at sparse, medium and dense traffic. We extend it with an urban grid with four-way intersections, a roundabout, and a two-level overpass. The overpass is where planar angles mislead, because vehicles at small angles may be on different roads.

Location error is injected under Gaussian, uniform and exponential models with standard deviations from 0 to 10 m, following the noise families in MEDA's evaluation. A secondary static test uses MEDA's synthetic 1000 m × 1000 m field with 400 uniformly placed nodes. Its 2,000 Monte Carlo runs check that the results are not artifacts of road geometry.

Metrics are PDR, redundancy as defined by Sun et al., end-to-end latency, MAC collision count, and per-packet decision cost. The headline figure is the PDR ≥ 95% region for each method in the noise-density plane, drawn per road geometry.

Three checks decide the outcome. First, ODAM-C's PDR should degrade measurably as noise approaches typical inter-vehicle spacing; if it does not, the problem we target is not real. Second, the full rule must enlarge the 95% region over strengthened ODAM-C at equal or lower redundancy; this is the test that would sink the idea if a fixed margin suffices. Third, the distilled rule must keep most of that gain at decision cost comparable to ODAM-C's angle check. The two ablations will show whether any gain comes from noise awareness, from combining multiple relays, or from both.