Wireless networks must re-solve hard optimization problems, such as how much power each transmitter uses or which beam it points, every few milliseconds, before the radio channel changes and the old answer goes stale. The leading fast approach is deep unfolding: take a classical iterative solver, cut it to a handful of iterations, and turn each iteration into a trainable neural-network layer. Current unfolded solvers spend the same effort on every user at every layer. A user at the edge of a cell, tangled in interference, gets the same refinement as an isolated user whose answer settled after one step. Meanwhile, the real constraint is a hard deadline. What matters is the slowest case, not the average one.
This note proposes to borrow one idea from sparse "mixture-of-experts" language models. In expert-choice routing, each sub-network picks a fixed number of inputs to process, rather than each input picking a sub-network (Mixture-of-Experts with Expert Choice Routing). We apply this inside the unfolded solver. At every layer, a learned scorer chooses exactly $c$ of the $N$ users to receive the expensive update, and the rest receive a cheap update or are carried forward unchanged. Because $c$ is fixed, the compute per layer and the latency in the worst case are constant by construction. The saved budget can then buy more layers within the same deadline. Put in two sentences: unfolded solvers should spend their scarce per-layer compute on the users that still need it, and expert-choice selection is the one adaptive-compute mechanism that does this without making latency unpredictable.
Core Claim
Under a fixed wall-clock deadline, an unfolded MU-MIMO beamforming solver (multi-user, multi-antenna: one base station serving several users at once) can beat a uniform unfolded solver of equal latency. It does so by trading width for depth: each layer gives the costly per-user update, which involves inverting a matrix, to only the $c$ users a learned scorer selects, and the layer count rises accordingly. The scorer must satisfy three properties that prior expert-choice work never needed and unfolded-solver work never had to face. It must respect relabeling of users, it must give an output that changes continuously with the channel, and it must be compatible with the existing stability analysis.
The selection pool here is the set of users within a single problem instance, all of which are available at once. For this reason the main weakness of expert choice in language models, namely that future tokens are unavailable during word-by-word generation (Mixture-of-Experts with Expert Choice Routing), does not arise. Its other known weakness does carry over: selection quality degrades when the candidate pool is small. We test that weakness directly at small $N$.
Relation to Prior Work
Three graph-unfolding papers supply the base method and motivate the gap. UWMMSE embeds graph neural networks in a truncated WMMSE solver (the weighted minimum mean-square-error algorithm). It matches classical WMMSE in 4 layers instead of 100 iterations (Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation). It extends to MU-MIMO beamforming with about 3,300 shared parameters (Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks). USCA does the same for successive concave approximation applied to energy efficiency (Graph-based Algorithm Unfolding for Energy-aware Power Allocation in Wireless Networks). All three update every node at every layer. We keep their layer structure and graph parameterization but make the per-layer update sparse across users.
The non-learned accelerated solver A-MMMSE (An Accelerated Mixed Weighted-Unweighted MMSE Approach for MU-MIMO Beamforming) constrains the proposal. It replaces cubic-cost matrix inversions with GPU-parallel block coordinate gradient steps and matches WMMSE quality, so any learned gain must be measured against it at equal hardware.
Expert-choice routing supplies the selection mechanism (Mixture-of-Experts with Expert Choice Routing). In that setting, a minority of "hard" tokens attract several experts while most get one or two. That pattern is the analogue we hope to see for cell-edge users. A vision study finds expert choice and soft mixtures both beat token choice (Routers in Vision Mixture of Experts: An Empirical Study). Mixture of Sparse Attention shows that expert-choice selection can sparsify an operator other than a feed-forward block (Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing). That is the closest structural precedent. It, however, targets language-model attention and has no deadline or physical constraint.
Finally, the stability analysis of unfolded WMMSE (Stability Analysis of Unfolded WMMSE for Power Allocation) bounds output changes using the Lipschitz smoothness of every layer. A hard top-$c$ selection would break that argument, which shapes our gate design.
Research Program
Base architecture. The base is the MU-MIMO UWMMSE layer. Its expensive part is the per-user receive-filter and beamformer update, which inverts an $M \times M$ matrix for $M$ antennas. The cheap part is the graph aggregation and weight update.
Scorer. At layer $\ell$, a small shared graph network maps each user's state to a score $s_i^\ell$. Its inputs include the user's current WMMSE residual, its interference in and out, and its last change. The top $c = \lceil \rho N \rceil$ users receive the expensive update, where $\rho$ is a fixed fraction of users. Defining $c$ as a fraction keeps the method meaningful when a model trained on 20 users is deployed on 100.
Continuous gate. To keep the output continuous in the channel, the update is applied as
$$x_i^{\ell+1} = x_i^{\ell} + g_i^{\ell}\,\big(U(x^{\ell})_i - x_i^{\ell}\big), \qquad g_i^{\ell} = \sigma\!\big(\beta\,(s_i^{\ell} - s_{(c+1)}^{\ell})\big)\cdot \mathbb{1}[i \in \text{top-}c].$$
Here $x_i^\ell$ is user $i$'s beamformer, $U$ is the full unfolded update, $s_{(c+1)}^\ell$ is the $(c{+}1)$-th largest score, $\sigma$ is a shifted sigmoid equal to zero at zero, and $\beta$ is a sharpness constant. A user entering or leaving the selected set at the boundary therefore receives a near-zero update, so small channel changes cause small output changes. We will extend the per-layer bound of the stability analysis to this gate. Selection depends on scores alone, so relabeling users relabels the selected set identically; the selection respects relabeling except for ties, which we break randomly.
Training. Training uses the same unsupervised sum-rate loss as UWMMSE, with gradients through $g$.
Variants. A two-expert variant gives one cheap and one expensive operator, each with a fixed capacity. This tests whether specialization adds anything beyond selection alone.
Iso-latency schedules. The main object of study is the latency–quality frontier. For a deadline $D$, we compare a uniform network of $L$ layers with selected networks of roughly $L/\rho$ layers that fit the same measured GPU time.
Baselines
The naive reference points are classical WMMSE run to convergence, which is the quality ceiling, and WMMSE truncated to the iterations that fit the deadline. On the SISO Rayleigh ad-hoc benchmark, Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation reports a WMMSE sum-rate of 82.94 at about 16 ms per sample. On the MU-MIMO benchmark, WMMSE takes 1,305 ms per sample (Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks).
The strongest learned competitor is uniform UWMMSE itself. It reports a sum-rate of 83.21 in 4 layers at about 2 ms per sample on the SISO benchmark (Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation) and 54 ms per sample on MU-MIMO (Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks). No MU-MIMO sum-rate value is reported in the file, so we will reproduce it. We will retrain it with the same data, the same number of training steps (10,000, as in the reference), and the same hyperparameter search budget. We will also give it extra layers until it reaches each deadline.
The strongest non-learned competitor is A-MMMSE. It reportedly matches WMMSE's weighted sum-rate and runs 3x faster than GPU-accelerated reduced-WMMSE at $M=2048$ (An Accelerated Mixed Weighted-Unweighted MMSE Approach for MU-MIMO Beamforming); no absolute latency at our problem sizes is reported in the file. We run it on the same GPU, truncated to each deadline.
The ablations isolate the selection ingredient. One updates a random set of $c$ users. One uses greedy residual-ranked selection, a classical rule that picks the users whose current error is largest, with no learning. One uses token-choice routing with a tuned auxiliary balancing loss. One uses per-user early halting, where each user stops once its learned confidence crosses a threshold. One uses a hard top-$c$ gate without the continuous gate.
Beating the baselines means a higher sum-rate than iso-latency uniform UWMMSE at two or more deadlines, a tail (99th-percentile) latency no worse than the uniform model, and a margin larger than the seed-to-seed spread over at least five seeds.
Why not just…?
Why not just pick the users with the largest residual, with no learning? This is a classical greedy coordinate-selection rule, and it is a real threat. Our scorer takes the residual as an input, so it can at worst copy this rule. The question is whether learning adds anything; for example, a scorer might anticipate which users' updates will relieve interference on their neighbours. The greedy-residual ablation settles this. If it matches the learned scorer, we will report that the gain comes from fixed-budget sparse updates plus extra depth, not from learned routing, and recommend the simpler rule.
Why not just let each user halt early when it has converged? Per-user halting allocates compute adaptively. However, the number of active users per layer then varies with the instance, so worst-case latency is set by the hardest channel draw. A coherence-time deadline punishes exactly that case. We will measure the 99th-percentile and maximum latency of halting against fixed-$c$ selection. If halting's tail latency stays within the deadline at equal mean quality, this threat stands and we will say so.
Why not use token-choice routing with a balancing loss, or a soft mixture? The expert-choice paper documents that auxiliary balancing losses are brittle and drop tokens when experts overflow (Mixture-of-Experts with Expert Choice Routing). Dropping a user's update in a beamforming solver is a silent quality loss. Soft mixtures performed well in vision (Routers in Vision Mixture of Experts: An Empirical Study), but they blend all inputs together, which here would mix different users' beamformers and their power constraints. They also do not reduce the number of matrix inversions. We include token choice as a direct ablation.
Why not just make the uniform unfolded network deeper, or use A-MMMSE? Both are covered by the iso-latency comparison. If uniform depth or A-MMMSE reaches the same quality within the same deadline, the proposal fails on its own terms.
Proposed Experiments
The primary testbed is the MU-MIMO interference-network beamforming benchmark of Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks with Rayleigh fading. We train with 20 users and test at 20, 50 and 100. We sweep the antenna count $M$, because the payoff should grow as the per-user inversion cost grows.
The secondary testbed is the SISO Rayleigh ad-hoc benchmark: path-loss exponent 2.2, 20 transceiver pairs, 6,400 test channel states. Here per-node updates are cheap, so we expect little latency benefit. This makes it an honest negative control, and it is also where we check the extended stability bound under standard-normal channel perturbations, mirroring the original analysis. A third testbed transfers the method to USCA on the energy-efficiency benchmark (train on 8 users, test on 100) to see whether it generalizes beyond WMMSE.
We report sum-rate, or energy efficiency for the third testbed, against measured GPU time per sample, with mean, 99th-percentile and maximum latency. We also report the distribution of how many layers select each user, correlated with that user's interference level. The decisive checks are four. First, the latency–quality frontier must dominate both uniform UWMMSE and A-MMMSE at matched hardware. Second, the learned scorer must beat greedy-residual selection. Third, gains must hold at $N=10$, where the selection pool is smallest. Fourth, the continuous gate must keep perturbation sensitivity within the extended bound while the hard gate does not.
References
- Mixture-of-Experts with Expert Choice Routing
- Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation
- Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks
- Graph-Based Algorithm Unfolding for Energy-Aware Power Allocation in Wireless Networks
- An Accelerated Mixed Weighted-Unweighted MMSE Approach for MU-MIMO Beamforming
- Routers in Vision Mixture of Experts: An Empirical Study
- Mixture of Sparse Attention: Content-Based Learnable Sparse Attention via Expert-Choice Routing
- Stability Analysis of Unfolded WMMSE for Power Allocation