When a phone, aircraft, or satellite user moves faster than global routing can update, networks keep sessions alive by pointer forwarding: the anchor the user just left (an anchor is a router or satellite that stands in for the user's address) forwards packets to the next anchor, which forwards them on again. The chain avoids outages, but it grows. Each extra hop adds delay and a new queue, and those queues interact because TCP's feedback loop, acknowledgements, and shared backhaul links tie them together. When a session slows down, an operator cannot tell which anchor is responsible, or whether cutting that anchor out of the chain would help. Existing schemes compact chains using fixed thresholds on domain size or chain length, and evaluate the whole bundle end to end.
We propose treating a forwarding chain the way causal-discovery work treats a gene network. Deliberately nudge a small, chosen set of anchors, for example with a slight rate limit. Record how the steady-state delay at every other anchor shifts. From those shifts, recover the coupling between anchors well enough to predict the effect of interventions that were never tried, including bypassing a given anchor. The method comes from a recent identifiability result: for linear stochastic systems, one intervention per strongly connected component (a group of variables that all influence each other through feedback) is enough to recover the full interaction matrix. The result was proven for gene expression and has never been applied to packet queues. The network question gives it a structural advantage. Traffic enters a forwarding chain at a single upstream anchor, so the "single root" condition the theorem needs holds by construction for single-sender chains.
Core Claim
We claim that per-anchor queueing delay along a pointer-forwarded chain, linearized around its operating point, behaves closely enough like a multivariate Ornstein–Uhlenbeck process that one probe per feedback-coupled anchor group suffices. By an Ornstein–Uhlenbeck process we mean a system that is pulled back toward a mean, with the pull between variables set by an interaction matrix and random noise added. Under this claim, the probes recover the chain's interaction structure and let us rank which anchor to compact away. The ranking should be about as accurate as probing every anchor individually, and clearly better than using passive telemetry alone or fixed compaction thresholds. We write the model as
$$dX_t = -(\Lambda X_t - b)\,dt + \Sigma^{1/2} dW_t,$$
Each symbol has a concrete meaning. $X_t$ is the vector of excess queueing delays at the $k$ anchors of the chain. $\Lambda$ is the matrix of how each anchor's delay pushes on the others. $b$ is the load-driven offset. $\Sigma$ is the noise covariance. $W_t$ is standard Brownian noise.
One Intervention per Component is Enough supplies the method. It proves that one intervention per strongly connected component identifies $\Lambda$, $b$ and $\Sigma$ up to a single positive scale $c$, and it gives a two-phase estimator. The first phase finds the components from downstream shifts in the steady-state mean. The second phase solves the steady-state mean and covariance (Lyapunov) equations for each component in topological order, using $\ell_1$-regularized least squares.
The unknown scale $c$ does not hurt our main use case. The steady-state mean $\Lambda^{-1}b$ is unchanged when both $\Lambda$ and $b$ are multiplied by $c$, so predicted steady-state delay after an intervention is fully identified. Only the time it takes to settle is not. That quantity has to come from a short transient measurement, a limitation the paper itself notes.
Research Program
The work has four parts: a mapping, an estimator port, a decision rule, and stress tests of the theorem's assumptions.
The mapping defines the observed variables and the allowed probes. The variables are per-anchor excess sojourn delay, which each anchor can timestamp locally, or, in a harder variant, only delays inferred from the endpoints. The probes are small hard or soft interventions that the forwarding plane already supports: a token-bucket rate cap, an added fixed delay, or diverting a fraction of traffic past the anchor. These correspond to the "hard" and "soft" interventions in the synthetic experiments of the source paper.
We then derive when the coupling graph has cycles. TCP acknowledgements returning up the same chain, shared radio or inter-satellite backhaul, and control signaling for pointer updates on the data links all create two-way coupling. The resulting components set the number of probes needed. This is where pointer-forwarding structure and component isolation actually meet. The component count predicts the probing budget before any experiment is run.
Bypassing an anchor is not literally a node intervention, because it rewires the chain. We model it as a hard intervention that clamps the bypassed anchor's excess delay to zero and adds the link stretch it saves. We then check this approximation directly against ground-truth compaction in simulation. If the approximation fails, the method is useful for diagnosis but not for choosing what to compact, and we will report that.
The decision rule compacts the anchor with the largest predicted reduction in end-to-end steady-state delay, but only when that reduction exceeds the signaling cost of re-binding. This replaces the fixed domain-size threshold used in Dynamic Pointer Forwarding Scheme for HMIPv6 Networks.
Finally, we stress the theorem's assumptions where networks are likely to break them. Multiple senders entering at different anchors create multiple root components, which the theorem does not cover. Shared backhaul creates correlated noise, violating the diagonal-noise assumption. Congestion creates nonlinearity. We will measure how prediction error degrades with each violation rather than assume it away.
Baselines
The naive reference is fixed-threshold chain management with no probing. We use the adaptive switching between access-router-level and anchor-level pointers from Dynamic Pointer Forwarding Scheme for HMIPv6 Networks. That paper reports only cost curves tracking the lower envelope of static schemes across session-to-mobility ratio, with no single number. We add a "compact after $L$ hops" rule with $L$ tuned on held-out traces.
In the satellite setting, the strongest complete system is SKYCASTLE. It places anchors on satellites, assigns users to anchors heuristically, and tunnels packets with Segment Routing. It reports a Connection Uninterrupted Ratio above 98% against 50–79% for ground-anchor schemes, and an average RTT reduction of 21.5% (peak 47.8%). We run it unmodified as the host system and add our diagnosis and compaction on top. Beating it means lower tail RTT and fewer degraded-flow minutes at equal or better uninterrupted ratio.
For the diagnosis task itself, there is no reported result on Lacuna for anchor-level fault localization or compaction ranking in forwarding chains. The comparison is therefore internal and must be fair. The strongest competitor is exhaustive single-anchor probing: nudge every anchor once, $k$ probes, and fit the same estimator. It uses the same data per probe, the same probe magnitude, and the same regularization, tuned on the same validation chains. Beating it means matching its ranking accuracy with fewer probes, and therefore less self-inflicted disruption.
The ablation that isolates the key ingredient is the covariance-only (observational) estimator from the same paper. The paper reports that the mean-plus-covariance method significantly outperformed covariance-only on unseen Perturb-seq knockouts and was comparable to an oracle given interventional means; the file gives no DES/PDS values. We also run an oracle that measures ground-truth compaction outcomes for every anchor, which gives the ceiling.
Why not just…?
Why not just instrument every anchor and read off which one is slow? In a single operator's simulator this works for "which queue is longest." It does not say which queue causes the others to grow. A downstream anchor can look congested because upstream bursts are passed through it, so bypassing it may not help. Passive per-hop telemetry is exactly the covariance-only condition of our ablation. If it ranks compaction targets as well as the probing method, the intervention component is unnecessary and we will say so. In multi-operator or satellite settings, internal anchors are often not observable at all, and the endpoint-only variant tests that case.
Why not just ablate everything in a full factorial grid? Enumerating compaction and probing subsets grows exponentially with chain length. Constellation-scale emulation in StarryNet already takes minutes to tens of minutes per environment setup at thousands of satellites. The real limit, though, is operational: every probe on a live chain degrades a real session. Factorial testing is our offline ground truth, not a deployable method. The claim to test is that one probe per coupled group loses little accuracy relative to it.
Why not just train a black-box predictor on telemetry? A regressor from RTT and loss features to "best anchor to bypass" is a serious competitor when training traces are plentiful. We include one, trained on the same trace budget. We expect it to fail when mobility patterns change between training and test, since it learns correlations tied to one traffic regime. The structural model re-estimates from fresh probes each time. The experiments settle this by training on one flight route and testing on the other.
Proposed Experiments
The first stage is synthetic. We use chains of 3–12 anchors whose ground truth is either our linearized model or packet-level queues with TCP flows. In the linear case we check that parameter error falls as the number of probes reaches the number of components, reproducing the source paper's synthetic trend. In the packet-level case we measure how far the linear approximation holds as load rises toward saturation.
The second stage uses an HMIPv6-style simulation on the 8×8 access-router grid with five anchor tiers, following the DPF-HMIPv6 setup. We replace Random Waypoint mobility with correlated trajectories and compare compaction policies on total signaling cost and 95th-percentile delivery delay.
The decisive test runs in StarryNet with Starlink (1,584 satellites) and Kuiper (1,296) orbital data, using the United 857 and Virgin Atlantic 138 flight traces and SKYCASTLE as the host system. Multi-sender traffic probes the multi-root failure case, and cross-seam handovers test sudden topology change.
We report three primary metrics: Spearman correlation between the predicted and true ranking of anchors by bypass benefit, the number of probes used, and end-to-end tail RTT together with uninterrupted ratio under the resulting compaction policy.
The idea is supported if three things hold. First, per-group probing reaches at least 90% of exhaustive probing's ranking correlation with substantially fewer probes. Second, it clearly beats covariance-only. Third, it lowers tail RTT relative to SKYCASTLE's default behavior without reducing uninterrupted ratio.
The idea is refuted if either of two things happens. Passive telemetry matches the probing method, or the bypass-as-clamping approximation mispredicts ground-truth compaction by more than the gap between the policies.
References
- One Intervention per Component is Enough: Towards Identifiability in Linear Stochastic Dynamics from Steady State
- Dynamic Pointer Forwarding Scheme for HMIPv6 Networks
- SkyCastle: Taming LEO Mobility to Facilitate Seamless and Low-latency Satellite Internet Services
- StarryNet: Empowering Researchers to Evaluate Futuristic Integrated Space and Terrestrial Networks