Layered Multicast Beamforming for the Channel That Arrives, Not the One Measured

When a base station sends the same live video to a group of phones, everyone has to decode the same signal. The group's rate is therefore set by its weakest member. Scalable video coding is the usual answer. It splits the stream into a small base layer that everyone must receive and optional enhancement layers that only better-placed users need to decode. With multiple antennas, the transmitter can then shape a separate beam for each layer. The difficulty is that designing those beams is a hard non-convex problem, and the solvers people actually use take from tens of milliseconds to seconds. Channels in moving vehicles change on millisecond timescales. So the beams are always computed for a channel that no longer exists. Part of that staleness is caused by the solver itself: the more iterations it runs, the older its answer is when it goes on air.

This note proposes to treat that delay as part of the design problem. We would train a fast, fixed-depth solver built by "unrolling" a classical iterative algorithm into a small neural network with learned step sizes. Its loss would be scored on the channel at the moment of transmission, which includes the time the solver spends computing. The key point is that layered video makes this trade-off asymmetric. Losing the base layer freezes a user's video, so its beam should be robust to channel drift. Losing an enhancement layer only lowers resolution, so those beams can aim aggressively at a predicted channel. We claim that a solver trained this way learns that asymmetry, and that the number of solver iterations has an optimal value that falls as user speed rises.

Core Claim

We make three connected claims. The first is that, for layered video multicast in mobile channels, the right objective is expected video quality evaluated on the channel after both the pilot-to-data delay and the solver's own compute time. The quality of the solution on the measured snapshot is the wrong target. The second is that optimizing this objective produces qualitatively different beams per layer: conservative beams for the base layer and sharper, prediction-driven beams for the enhancement layers. The third is that a solver's depth $K$ trades optimization quality against channel aging. The resulting optimal depth $K^\*$ can be measured, and it moves with Doppler, meaning with how fast the channel changes because of motion.

Concretely, consider superposition layered multicast with layer precoders $w_1,\dots,w_L$. Users decode layers in order, cancelling each decoded layer before attempting the next. User $u$ displays up to layer $\ell_u$, the largest $\ell$ for which its signal-to-interference-plus-noise ratio clears the threshold $\gamma_j$ for every layer $j \le \ell$. We optimize

$$ \max_{\theta}\; \mathbb{E}\Big[\textstyle\sum_{u} Q(\ell_u)\ \big|\ H\big(t+\tau_p+\tau_c(K)\big),\ W_\theta\big(\hat H(t)\big)\Big], $$

where the symbols are defined as follows. $\theta$ are the learned parameters of a depth-$K$ unrolled solver $W_\theta$. $\hat H(t)$ is the channel estimated from pilots. $H(\cdot)$ is the true channel at transmission time. $\tau_p$ is the fixed pilot-to-data delay. $\tau_c(K)$ is the measured wall-clock compute time of $K$ unrolled steps. $Q(\ell)$ is the rate-quality (PSNR) of the video decoded through layer $\ell$. The hard "decoded or not" indicator is replaced during training by a sigmoid surrogate.

Relation to Prior Work

Two papers join multicast with layered or loss-tolerant video. Chakareski jointly chooses client clusters, source rates and error-correction parity for H.264/SVC over IPTV using nested coordinate descent. ul Zuhra et al. schedule loss-tolerant multicast with virtual token queues and bipartite matching. Both motivate this work, because they show that the video structure is worth several dB. Both also treat the radio link as fixed, static bit pipes and do not design antenna beams. We keep their video-quality objective and move it down into the precoder.

A second group designs multicast beams but ignores video and time. Zang et al. solve robust max-min multigroup multicast with majorization-minimization plus bisection over a spherical error ball. Zhang et al. schedule groups and run projected subgradient beamforming, with their fastest variant still taking 0.357 s at 128 antennas. The NP-completeness result constrains the design: exact solutions are hopeless beyond a handful of antennas, so we target good stationary points quickly. We reuse the majorization-minimization / projected-gradient iteration as the step to be unrolled. The spherical error ball is exactly what we replace with channel trajectories drawn from a simulator.

The unrolling technique itself comes from Shlezinger et al., UWMMSE and USCA. The last two use graph neural networks so that one trained model generalizes across network sizes. All three target unicast sum-rate or energy efficiency on a single static snapshot, and none consider max-min shared beams or layer dependencies.

On channel aging, the cell-free rate-splitting study shows that a common message decodable by all users absorbs interference from outdated channel estimates. That is structurally a one-layer multicast. It also treats the aging delay as exogenous and ignores the cost of its own bisection solver. MambaCSP predicts future channels in linear time but does not design beams.

The closest cross-direction system is Deep-GRAIL. It learns rate-splitting VR multicast policies with a reinforcement learner guided by imitation of a classical solver. It assumes perfect static channel knowledge and tiled rather than layered content. Its ablation, which found that imitation of the classical solver mattered most, constrains us: we do not start from a model-free learner.

Research Program

We would build a layered multicast simulator on the QuaDRiGa 3GPP urban-macro non-line-of-sight channel model, sweeping speeds from 10 to 100 km/h. We would use 16 to 64 transmit antennas and multicast groups of 5 to 30 users. Channels are sampled as continuous trajectories rather than independent blocks.

Video quality per layer comes from H.264/SVC rate-PSNR operating points on CIF Foreman, as in Chakareski, with two to four layers. The Arizona State University traces in the file are single-layer H.264/AVC. We would use them only for the monolithic comparison.

The solver unrolls $K$ projected-gradient steps of a smoothed max-min layered objective. A graph network over users outputs per-layer step sizes and per-user weights, and a closed-form projection keeps the power budget exact at every layer. The input is either the current estimate $\hat H(t)$ or that estimate plus a MambaCSP-style prediction. The solver is trained with the equation above. $\tau_c(K)$ is measured on the target hardware and rounded up to channel-sample resolution, so deeper networks see more-aged channels during training as well as at test time.

The analysis would answer three questions. Does $K^\*$ exist, and does it shrink with speed? Do the learned base-layer beams become wider (flatter gain across users) than the enhancement beams as speed grows? Does the learned solver generalize across group sizes without retraining?

Baselines

The naive reference is single-layer multicast at the rate the worst user can decode. Its beam is computed by the robust majorization-minimization solver of Zang et al. on the stale estimate and run to convergence. Its real solve time is charged as channel aging. This is what a careful deployment does today. We would count the proposal as beating it if the proposal delivers higher mean PSNR across users together with a lower frozen-video (base-layer outage) rate at every speed at or above 30 km/h.

The strongest existing method is the same majorization-minimization solver extended to layered superposition, with one max-min subproblem per layer solved in sequence. It is fed the MambaCSP-style predicted channel for the actual transmit time and given a robustness margin tuned by grid search on validation trajectories. We run it both to convergence and truncated at its best iteration count, with that count also chosen on validation data. Both variants run on the same hardware and trajectories as the proposal, and both get the same number of validation evaluations for tuning. There is no reported result found on Lacuna for layered multicast beamforming under channel aging, so this number must be produced here. The closest reported figures come from different settings. Chakareski reports 39.52 dB expected Y-PSNR at 1200 kbps for Foreman with packet-level losses and no beamforming. Zhang et al. report 0.357 s for their fastest multicast scheduler-plus-beamformer at 128 antennas. To count as beating the tuned, truncated, prediction-fed baseline, the proposal must exceed it by at least 0.5 dB mean PSNR at matched base-layer outage, or cut outage by a fixed relative margin at matched PSNR, across three random seeds.

The ablation that isolates the key ingredient is the identical unrolled architecture trained on the snapshot objective, where $H(t)$ is scored instead of the aged channel. A second ablation keeps the aged-channel loss but ties all layers to one shared robustness level, which removes the asymmetry. If the snapshot-trained network matches the proposal, the time-aware objective contributes nothing. If the tied-robustness network matches, the layered-asymmetry claim fails.

Why not just…?

Why not just predict the channel and run a fast classical solver on the prediction? This is the most serious threat, which is why it is the strongest baseline rather than a straw man. The evidence against it is partial. The rate-splitting study shows that conventional precoding saturates under any residual channel error. MambaCSP reports only relative gains, about 9 to 12% lower normalized squared error than language-model predictors, so residual error remains. Neither fact shows that a well-tuned robust margin cannot close the gap. The truncated, margin-tuned comparison settles it. If that baseline comes within 0.3 dB of the proposal, we would report the time-aware, layer-asymmetric objective as the contribution and drop claims about unrolling.

Why not just stop a classical solver early instead of learning an unrolled one? Early stopping implements the same depth-for-freshness trade-off without any learning. That is why the baseline includes the truncated solver at its tuned best depth. Prior unrolling results suggest learned steps reach good solutions in far fewer iterations. UWMMSE matched 100 WMMSE iterations (82.94) with 4 layers (83.21), and USCA ran 58× faster than SCA while shrinking its suboptimality gap by 65%. Those results are for unicast, however, and may not transfer. The curve of PSNR against depth at each speed, for both methods, will show whether learning shifts the frontier or merely matches it.

Why not just send single-layer video and drop or queue the weak users, as in ul Zuhra et al.? Their scheduler meets per-user loss budgets, but every served user gets the same quality, so strong users are wasted. Chakareski measured an 8.5 dB gain from layered clustering over a uniform midpoint scheme. We include a monolithic arm that combines worst-user dropping with virtual-queue scheduling on the Arizona State traces. That arm reports what fraction of users are dropped and what PSNR the rest receive, so the comparison is direct rather than assumed.

Why not just use rate-splitting's common message instead of video layers? A single common message is a one-layer multicast. As the file's tension analysis notes, it recreates the weakest-user bottleneck at the common-stream level when user signal quality differs widely. Our $L=1$ configuration is exactly this case. If adding layers gives no gain over $L=1$ at high speed, the layered hypothesis is refuted.

Decisive Checks

The central check is the depth-aging curve. For each speed, we plot mean PSNR and base-layer outage against unrolled depth $K$, using real measured compute times. The proposal predicts an interior optimum $K^\*$ that decreases monotonically from 10 to 100 km/h. The snapshot-trained ablation should show no such optimum, or a worse one. A flat curve would mean compute-induced aging is negligible at these solver speeds, and the premise fails.

The second check inspects the learned beams. For each layer, we measure how evenly gain is spread across group members, and how quickly delivered signal quality decays as the transmit delay grows. The claim holds only if the base-layer beams become measurably more uniform than the enhancement beams as speed increases.

The third check tests generalization. We train on 10-user groups and test on 5 and 30 users, following the equivariance protocol of UWMMSE. The proposal must stay ahead of the tuned truncated baseline without retraining.