A low-Earth-orbit satellite serving ships, mountain villages or desert sites has to point many narrow radio beams at once from a large antenna array. It cannot afford a separate radio chain behind every antenna. So it uses hybrid beamforming: a slow, cheap network of analog phase shifters steers a few wide "super-beams," and a small digital processor behind a handful of radio chains mixes those super-beams so that users do not interfere with one another.
Every published satellite design of this kind that we found assumes two things. The satellite knows each user's channel perfectly, and that channel stays frozen for the whole frame. At orbital speeds of several km/s, neither holds. By the time the digital mixing weights are applied, they are stale, and the interference they were meant to cancel returns. The damage is not uniform across users. Users whose channels drift fastest, or who were already weak, lose the most. A beamformer that was "fair" on stale data becomes unfair on the real channel.
This note proposes a way to split the problem along the satellite channel's two timescales. The analog stage mostly depends on the geometry between satellite and user: angles and range, which are predictable from the orbit. The satellite sets it once per window so that it holds up across the whole predicted drift. The digital stage depends on fast residual phase changes that cannot be predicted. It is refreshed from cheap pilots, which are short training signals of known content. Because these pilots only probe the few-dimensional channel seen through the analog beams, they cost little. The part that is new here is that the pilot budget itself becomes a fairness lever. How often each user's digital weights get refreshed is chosen jointly with the precoder under a smoothed max-min rate objective, so the satellite spends training time where aging would otherwise starve someone.
Core Claim
In a LEO downlink with hybrid hardware, the right place to fight both channel aging and unfairness is the allocation of per-user refresh rates for the low-dimensional digital stage, under an analog stage designed to be robust over a window. Re-solving a perfect-CSI hybrid design at every slot on predicted channels is the wrong place. We claim this joint design raises the worst user's time-averaged rate substantially over (a) per-slot fair hybrid precoding fed with extrapolated channels and (b) sum-rate hybrid designs. We further claim it keeps sum-rate close to the sum-rate designs. The gain should grow with the amount of unpredictable aging and vanish when aging is fully predictable. That dependence is itself a falsifiable prediction.
Why the existing pieces do not add up to this
Two satellite papers supply the setting and the hardware model. Multi-Satellite Cooperative Networks: Joint Hybrid Beamforming and User Scheduling Design splits precoding into codebook analog alignment plus digital interference cancellation across cooperating LEO satellites. It schedules users by spectral-efficiency gain and assumes perfect, instantaneous CSI. Joint Beamforming and Illumination Pattern Design for Beam-Hopping LEO Satellite Communications jointly optimizes which ground cells are lit and the hybrid weights, using manifold optimization. It maximizes sum-rate on a per-slot static channel with perfect CSI. Both motivate this work, because their gains rest precisely on the interference cancellation that stale CSI destroys. Neither handles aging or fairness.
The fairness objective and its solver come from Max-min Rate Optimization of Low-Complexity Hybrid Multi-User Beamforming Maintaining Rate-Fairness. That paper replaces the non-smooth minimum with a smooth surrogate and derives closed-form alternating updates for sub-array hardware with quantized phase shifters, but only on static terrestrial snapshots. We reuse its surrogate and update structure and change what is being optimized: window-averaged rates under aging, plus refresh periods. The constraint on the objective comes from Razaviyayn et al., who show exact max-min beamforming is NP-hard, and Proutiere et al., who show strict rate equalization collapses capacity. Both push us toward a soft minimum rather than hard equalization.
On aging, Joint Optimization of Power Allocation and Training Duration for Uplink Multiuser MIMO Communications gives the pilot-versus-payload tradeoff for fully digital, sum-rate systems with symmetric users. We extend the same tradeoff to per-user refresh periods, hybrid hardware and a fairness objective. Rate-Splitting for CF Massive MIMO Systems With Channel Aging shows that outdated CSI creates an interference floor that saturates rate as power grows, and absorbs it with a common message, but assumes fully digital access points. We treat it as a competing remedy rather than a building block. MambaCSP forecasts fast-fading channels at linear cost but never closes the loop into beamforming. It serves as the strongest "just predict better" alternative.
Research Program
We model user $k$'s channel from an array of $N_t$ antennas as a dominant line-of-sight path plus weak scatter. Its angle, range and bulk Doppler follow the orbit and the user's surveyed position. A residual phase $\phi_k(t)$, which is unpredictable and caused by oscillator drift, position error and atmosphere, wanders as a random process whose variance grows with elapsed time.
The analog precoder $F_{RF}$ is held fixed over a window of $T$ slots. With $N_{RF}$ radio chains, each user's effective channel $g_k(t) = F_{RF}^H h_k(t)$ has only $N_{RF}$ entries, so a refresh pilot costs $O(K)$ symbols for $K$ users. Probing the full $N_t$-dimensional channel through few radio chains would cost far more. That dimension reduction is what makes frequent refreshes affordable on hybrid hardware, and it resolves the conflict the research file notes between the minimum training duration in Lu et al. and hybrid sounding.
User $k$'s digital weights are refreshed every $\tau_k$ slots. Between refreshes the satellite applies the last estimate, rotated by the predicted geometric phase. The expected rate $\bar R_k(F_{RF}, F_{BB}, \tau_k)$ is averaged over the window. Here $F_{BB}$ is the digital precoder, and the average accounts for the growing residual-phase decorrelation within each refresh interval and for the payload fraction lost to pilots. The design problem is
$$ \max_{F_{RF},\,F_{BB},\,\{\tau_k\}} \; -\frac{1}{c}\log \sum_{k=1}^{K} e^{-c\,\bar R_k(F_{RF},F_{BB},\tau_k)} \quad \text{s.t. } |[F_{RF}]_{ij}|=1,\ \ \mathrm{tr}(F_{RF}F_{BB}F_{BB}^H F_{RF}^H)\le P,\ \ \sum_k \frac{L_p}{\tau_k}\le B . $$
The objective is the log-sum-exp soft minimum with sharpness $c$, in the spirit of Zhu et al. The second constraint is the transmit power budget $P$. In the third, $L_p$ is the pilot cost per refresh and $B$ is the total training budget per slot.
We solve this by alternation. The analog stage is updated by manifold steps on the unit-modulus set against the ensemble of predicted geometries across the window, and we will test both an AOSA (array-of-subarrays) variant and a fully connected variant. The digital stage uses closed-form soft max-min updates. The refresh periods are updated by a one-dimensional search per user on the concave-in-practice window-rate curve, since Lu et al.'s quasiconcavity result suggests this curve is well behaved.
We will also derive a small analytic result for a two-user line-of-sight case. It states when a slowly aging strong user should cede refresh budget to a fast-aging weak user, which shows the fairness-aging coupling in closed form.
Baselines
The naive reference is the analog-only grid of beams with proportional-fair time scheduling. The satellite steers DFT beams along ephemeris and shares slots among users. Zhang et al. report that their multi-connection hybrid scheme exceeds this analog-plus-scheduling scheme (AU) by "over 200%" in spectral efficiency in some constellations. We rerun it on identical orbits and user drops, so that every gain is relative to the cheapest deployable system.
The strongest existing competitor is Zhu et al.'s soft max-min hybrid precoder, re-solved every slot on ephemeris-extrapolated channels refreshed at a uniform period. It is the only published method that already balances throughput and fairness under sub-array hardware and phase quantization. The file reports it reaching "~40 bps/Hz" sum-rate at 25 dBm with 72 antennas and 8 users while keeping minimum per-user rates of 1.0 to 2.5 bps/Hz, with only a 3–5% drop using 3-bit phase shifters. We will first reproduce that terrestrial setting to validate our implementation. We will then give it the same training budget $B$, the same power, antenna count, quantization and iteration cap, and tune its sharpness parameter on the same held-out passes. We will also include the sum-rate beam-hopping design of Wang et al., which the file reports converging in under 8 iterations, as the throughput ceiling.
Beating the strongest competitor means a higher window-averaged minimum and 5th-percentile user rate, with sum-rate within a few percent, at equal pilot budget. The difference must hold across passes rather than on a few favorable geometries.
Two ablations isolate the key ingredient. The first keeps our window-robust analog stage but refreshes all users uniformly. The second keeps aging-aware refresh but redesigns the analog stage per slot from predicted geometry. If uniform refresh matches the full method, the pilot-allocation idea is empty. Finally, a genie that knows the true channel every slot gives the upper bound.
Why not just…?
Why not just feed a better channel predictor into the existing fair precoder? This is the real threat. If MambaCSP-style forecasting removed most residual phase error, per-slot soft max-min on predicted channels would match our method at zero added design effort. The file reports MambaCSP achieving 9% (TDD) and 12% (FDD) lower NMSE than LLM predictors, but at up to 100 km/h on terrestrial urban channels. It reports nothing on satellite residual phase, so the question is open. We will train a MambaCSP-type predictor on the same simulated passes and plug it into the Zhu baseline. We concede the point in any regime where it closes the gap. Our claim survives only if unpredictable phase is large enough that no predictor removes it, and the sweep over residual-phase variance will show where that boundary lies.
Why not just use rate-splitting to absorb stale-CSI interference? Rate-splitting stops the interference floor from saturating rate under aging in fully digital cell-free systems, per the Rate-Splitting paper. The file reports no numerical rates for it beyond the $v = 30$ m/s setting. On hybrid satellite hardware, the common stream must share the same few radio chains and unit-modulus beams. We will implement a hybrid rate-splitting variant with the same budget. Rate-splitting and refresh allocation are complementary, so if both help we will report the combination rather than claim exclusivity.
Why not just shorten frames and re-sound everyone often? Lu et al. show that frequent training starves payload, and full-dimensional sounding through $N_{RF} \ll N_t$ chains is especially expensive. The uniform-refresh ablation is exactly this alternative made as cheap as possible, because it probes only the effective channel. If it matches the full method, then per-user allocation adds nothing.
Proposed Experiments
Satellite trajectories and user geometry come from STK, following Zhang et al. We use Walker-type constellations near the 45° inclination they found favorable, with user drops that are both sparse and dense, including the sparse case they report achieving 43.2% higher mean SE. Channels are built as line-of-sight-dominant Rician links on those trajectories, with a tunable residual-phase random walk. We will cross-check terrestrial-style fading with QuaDRiGa, which MambaCSP and Detect to Learn both use.
Array sizes span 64 to 96 elements to match Wang et al.'s reported ~33% sum-rate gain from 64 to 96. The number of radio chains varies from $K$ to $2K$, bracketing the $N_{RF} \ge 2N_s$ condition from Hybrid Digital and Analog Beamforming Design for Large-Scale Antenna Arrays, and phase-shifter resolution varies from 3 bits to continuous.
Metrics are window-averaged minimum rate, 5th-percentile rate, sum-rate, Jain's fairness index, rate trajectories over a full pass, pilot fraction, and per-window compute time against a realistic onboard budget.
Three results decide the claim. First, the gain over the strongest baseline must rise with residual-phase variance and fall to near zero when aging is fully predictable. A flat gain would mean we are measuring something other than aging. Second, the uniform-refresh ablation must lose most of the fairness gain, which shows that pilot allocation, not the window-robust analog stage alone, does the work. Third, the learned-predictor baseline must fail to close the gap at realistic satellite residual-phase levels. If it closes the gap, the honest conclusion is that prediction, not joint design, is what matters, and the note will say so.
References
- Multi-Satellite Cooperative Networks: Joint Hybrid Beamforming and User Scheduling Design
- Joint Beamforming and Illumination Pattern Design for Beam-Hopping LEO Satellite Communications
- Max-Min Rate Optimization of Low-Complexity Hybrid Multi-User Beamforming Maintaining Rate-Fairness
- Linear transceiver design for a MIMO interfering broadcast channel achieving max-min fairness
- Wireless downlink data channels: user performance and cell dimensioning
- Joint optimization of power allocation and training duration for uplink multiuser MIMO communications
- Rate-Splitting for CF Massive MIMO Systems With Channel Aging
- MambaCSP: Hybrid-Attention State Space Models for Hardware-Efficient Channel State Prediction
- Hybrid Digital and Analog Beamforming Design for Large-Scale Antenna Arrays