Relaying a short packet over a long distance through several relays that each decode it and re-transmit a clean copy (decode-and-forward relaying) has a tidy textbook answer: there is an optimal number of relays, and the packet length and transmit power at each hop follow a closed-form rule. That answer comes from Low-Latency Short-Packet Transmission over a Large Spatial Scale, and it assumes each hop's signal strength is fixed by distance alone. Real links fade, get shadowed, and change character as nodes speed up. Under decode-and-forward, one hop that fails to decode kills the whole packet. So when fading is realistic, a fixed allocation built for the average channel will either miss its reliability target or spend energy it did not need to. This proposal asks how much of the known latency-energy optimum survives realistic channels. It then builds a controller that sizes each hop's packet length and power from a statistically calibrated, distribution-free lower bound on that hop's next-slot signal quality, so the end-to-end error target holds under non-stationary fading at the lowest energy cost.
In two sentences: today's multi-hop short-packet allocations are optimized for a channel that does not fade, and we measure how badly that breaks and where the optimal relay count moves. We then replace the nominal channel in the same convex allocation problem with an online-calibrated worst-plausible channel per hop, with each hop's error budget set by end-to-end reliability and energy prices.
Core Claim
The claim has two parts. The first is an empirical claim that can be tested. The strict optimum in relay count and the inverse blocklength-to-power rule reported by Low-Latency Short-Packet Transmission over a Large Spatial Scale are partly artifacts of an idealized path-loss model. Once per-hop fading variance and non-stationarity are introduced, the energy-optimal relay count and the split of error budget across hops change in ways the nominal model cannot predict.
The second is a method claim. We propose feeding the finite-blocklength allocation a per-hop conformal lower bound on signal-to-noise ratio (SNR). Conformal inference is a wrapper that turns any predictor into prediction sets with guaranteed error frequency, without assuming a channel distribution. The miscoverage level of each hop's bound is adapted online. The result should meet an end-to-end packet error target at lower energy per delivered packet, including decoding energy, than any static fading margin tuned in advance. The advantage should appear specifically when the channel regime changes.
The mechanism is structural. Under decode-and-forward, end-to-end error over $N$ hops is
$$ \varepsilon_{\text{e2e}} = 1-\prod_{i=1}^{N}(1-\varepsilon_i) \le \sum_{i=1}^{N}\varepsilon_i , $$
where $\varepsilon_i$ is the decoding error of hop $i$. The total reliability target $\varepsilon^\star$ is therefore a budget to be divided among hops. The division should shift toward hops whose channels are currently volatile. The nominal allocation divides it by geometry alone.
Research Program
We first build a simulated $N$-hop decode-and-forward chain with explicit latency and energy accounting. The latency budget is $T \ge \sum_i n_i + \sum_i d_i$, where $n_i$ is the blocklength of hop $i$ in channel uses and $d_i$ is its decoding delay. The energy budget counts transmit energy $p_i n_i$ plus a per-hop decoding and circuit cost. The decoding cost follows Optimizing Energy Harvesting Decode-and-Forward Relays With Decoding Energy Costs and Energy Storage, because the research file notes that circuit and decoding costs dominate in the short-packet regime. We add a circuit-activation threshold of the kind modeled in Outage Performance of Wireless Powered Decode-and-Forward Relaying Networks in Rician Fading as an optional stress condition. The per-hop error $\varepsilon_i(n_i,p_i,\gamma_i)$ uses the finite-blocklength normal approximation, as in the source paper.
The first study sweeps channel realism while holding everything else fixed. The ladder runs from deterministic path loss (the source paper's setting), to static Rayleigh and Rician fading, to 3GPP tapped-delay-line and clustered-delay-line link-level models, to velocity-transition schedules (3, then 120, then 500 km/h) taken from Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading. Where available, the ladder ends with scenario-level spatially varying models such as QuaDRiGa or ray-tracing, which Wireless Channel Modeling for Machine Learning -- A Critical View on Standardized Channel Models argues are needed to escape the "Gaussian trap." That paper shows link-level models collapse into Gaussian channels on which linear methods are optimal. It does not ask what this does to relaying.
For each rung we re-solve the allocation and record three things: the optimal relay count, the energy per delivered packet at target reliability, and the realized error of the nominal allocation deployed unchanged. We state no expected direction for the relay-count shift in advance. Fading variance rewards shorter hops, while Low-Latency Short-Packet Transmission over a Large Spatial Scale shows decoding delay and resource dilution penalize extra hops. Which effect wins is the empirical question.
The second study is the controller. At each frame, each relay forms a one-step-ahead SNR prediction from pilots, using linear prediction as the default in line with the Gaussian-trap finding. It wraps the prediction in adaptive conformal inference to obtain a lower bound $\underline{\gamma}_i$ that is violated with target frequency $\alpha_i$. The allocator then solves the convex program of the source paper with $\underline{\gamma}_i$ in place of the nominal SNR. The per-hop levels are updated online:
$$ \alpha_{i,t+1} = \alpha_{i,t} + \eta_i\,(\alpha_{i,t}^{\text{target}} - \mathrm{err}_{i,t}), \qquad \sum_i \alpha_{i,t}^{\text{target}} = \varepsilon^\star , $$
where $\mathrm{err}_{i,t}\in\{0,1\}$ records whether hop $i$ failed to decode at frame $t$. The step size $\eta_i$ is set from estimated Doppler spread using the closed-form rule of the conformal link-adaptation paper. The split of targets $\alpha^{\text{target}}_i$ across hops is re-chosen periodically so as to minimize the Lagrangian energy cost of meeting $\varepsilon^\star$. The coupling across hops is new here. It enters through the budget constraint, which the union bound above makes valid.
The conformal guarantee is marginal over time. The research file flags conditional coverage as open, so we report guarantees end-to-end in the marginal sense only, and we separately measure per-regime coverage.
Baselines
The naive reference points are direct single-hop transmission at the same energy budget and amplify-and-forward relaying at the same relay count. Amplify-and-forward has no decoding delay but forwards noise. Björnson et al. and Optimal energy-harvesting design for AF and DF two-way relay beamforming in 6G predict it loses. Including it checks whether that prediction still holds once decoding delay and energy are charged honestly under a short deadline.
The strongest existing method is the nominal finite-blocklength allocation of Low-Latency Short-Packet Transmission over a Large Spatial Scale, re-implemented exactly and given the same pilot-based SNR estimates as our controller. The research file reports no numeric error or energy value for this method (no reported result found on Lacuna for that allocation's end-to-end error or energy on a fading benchmark), so we will produce its numbers ourselves on every channel rung.
The strongest reliability reference for the adaptive component is Conformal-AMC. Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading reports a realized error of 10.3% (±0.012) against a 10% block-error target across the 3/120/500 km/h transitions. Its singleton rate is 82.4%, versus 75.3% for 3GPP outer-loop link adaptation. We run it per hop with equal budgets $\varepsilon^\star/N$, which isolates our cross-hop budget coupling.
Three further competitors are the nominal allocation with a static fading margin grid-searched offline, per-hop outer-loop link adaptation, and the allocation re-solved with ergodic or outage-based fading statistics. All methods share the same channel traces, pilot overhead, deadline, energy model and random seeds, and each gets the same hyperparameter search budget on a held-out trace set.
Beating the field means two things together. First, the realized end-to-end error must stay within the target's confidence band on every regime, including transitions. Second, energy per delivered packet must be strictly lower than the best baseline that also meets the target, at matched deadline-violation rate.
The key ablations are the controller with nominal SNR in place of the conformal bound, the controller with equal rather than adapted per-hop budgets, and the controller with a fixed step size in place of the Doppler-adaptive one.
Why not just…?
Why not just add a fixed fading margin to the nominal allocation? This is the real threat. A margin tuned for a known regime may be nearly as efficient as our controller. The case against it is that a margin tuned for pedestrian speeds is either wasteful or unreliable after a transition to vehicular speeds. Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading documents this failure mode for fixed-step outer-loop adaptation on a single link. We will test it directly. If the best offline-tuned margin matches our energy at target reliability across the transition schedule, we will say so. The contribution then reduces to the first study's finding that the nominal optimum moves, plus a recommended margin per channel regime.
Why not just run standard outer-loop link adaptation independently on each hop? Per-hop adaptation keeps each hop near its own target, but the targets never move between hops. A volatile hop then consumes reliability that a stable hop could have given up more cheaply. Its slow fixed step also falls behind during speed changes; the conformal paper reports a 75.3% singleton rate for outer-loop adaptation versus 82.4% for Conformal-AMC. The equal-budget per-hop conformal baseline and the outer-loop baseline together show whether the coupling, rather than the conformal wrapper alone, produces any gain.
Why not just re-solve the convex program with average fading statistics? Ergodic or outage-based statistics fix Rayleigh or Rician fading for single hops, as in Outage Performance of Wireless Powered Decode-and-Forward Relaying Networks in Rician Fading. They assume the statistics are stationary and known. Under the transition schedule they are neither. On the static fading rungs, this baseline should be close to our controller, and it will serve as a sanity check that the conformal layer does not cost energy when the channel does not drift.
Why not use a deep network to predict the channel instead of linear prediction? On link-level models, Wireless Channel Modeling for Machine Learning -- A Critical View on Standardized Channel Models finds linear estimators optimal. It reports an nMSE of 0.00915 for the linear minimum-mean-squared-error estimator versus 0.01204 for a diffusion model at 10 dB on CDL-A. A learned predictor would also add relay compute energy. We therefore use a learned predictor only on the scenario-level rung, and only as a secondary comparison with its inference energy charged.
Proposed Experiments
The first decisive check is whether the nominal allocation, deployed unchanged, misses $\varepsilon^\star$ on fading rungs, and by how much, as a function of relay count. If the misses are negligible even on the transition schedule, the problem is not worth a new controller, and we report that.
The second is whether the energy-optimal relay count differs between the path-loss rung and the scenario-level rung. We report the energy-versus-relay-count curves for each rung at fixed deadline and reliability.
The third is the head-to-head. We compare realized end-to-end error, energy per delivered packet (transmit plus decoding plus circuit), and deadline-violation rate across all baselines. The comparison uses 2 to 8 hops, targets $\varepsilon^\star$ of $10^{-1}$ and $10^{-3}$, and the 3/120/500 km/h schedule, with confidence intervals over seeds and per-regime breakdowns that expose any conditional-coverage failure.
The fourth, run as a stress test, adds the circuit-activation threshold and asks whether the controller avoids the outage cliff that Outage Performance of Wireless Powered Decode-and-Forward Relaying Networks in Rician Fading identifies. It also checks whether the throughput-versus-energy-efficiency divergence reported by Throughput Analysis of Buffer-Aided Decode-and-Forward Wireless Relaying with RF Energy Harvesting grows or shrinks under calibrated allocation.
The proposal succeeds if the coupled conformal controller meets the target in every regime and beats the best tuned static margin on energy after transitions. It is refuted, and reframed as a pure modeling study, if the tuned margin ties it.
References
- Low-Latency Short-Packet Transmission over a Large Spatial Scale
- Optimizing Energy Harvesting Decode-and-Forward Relays With Decoding Energy Costs and Energy Storage
- Outage Performance of Wireless Powered Decode-and-Forward Relaying Networks in Rician Fading
- Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading
- Wireless Channel Modeling for Machine Learning -- A Critical View on Standardized Channel Models
- A New Look at Dual-Hop Relaying: Performance Limits with Hardware Impairments
- Optimal energy-harvesting design for AF and DF two-way relay beamforming in 6G
- Throughput Analysis of Buffer-Aided Decode-and-Forward Wireless Relaying with RF Energy Harvesting