Deciding which radio heads to switch off, which users to admit, and how to steer their multicast beams is a joint problem: each discrete choice changes the continuous optimum, and the reverse is also true. A Cloud Radio Access Network (Cloud-RAN) is a cellular design in which many simple antenna sites, called remote radio heads, are run from one central processing pool. The standard way to solve the joint problem in such a network wraps two approximations together. The beam design is "lifted" into positive semidefinite matrices whose rank-one constraint is then dropped. This step is called semidefinite relaxation (SDR), and its output is a matrix that has to be turned back into a single beam. Heavy-tailed penalties then push whole radio heads to zero. Once SDR stops being exact, a random-sampling heuristic recovers the beams, and each outer iteration solves a large semidefinite program (Smoothed Lp-Minimization for Green Cloud-RAN with User Admission Control).
Our observation is that switching off a radio head and dropping a user are rank-reducing events. If the beam matrix for multicast group $k$ is factored as $W_k = V_k V_k^H$, then turning radio head $n$ off zeroes a block of rows of $V_k$, and dropping a whole group sends $V_k$ to zero. The resource-allocation outer loop therefore drives the beam variables to the rank-deficient boundary of the bounded-rank set. Recent geometry work shows this boundary is exactly where first-order methods can converge to points that look stationary but are not (Finding stationary points on bounded-rank matrices: a geometric hurdle and a smooth remedy). We propose to optimize the factored beams directly with that paper's second-order remedy, so that no semidefinite program is ever solved. When a final certificate is needed, we use an extreme-point rank-reduction step. We then test whether the boundary pathology actually occurs in green Cloud-RAN and whether handling it buys power savings.
Core Claim
In multicast green Cloud-RAN with joint radio-head selection and admission control, a factored, rank-pursuing second-order method is predicted to have two advantages over SDR plus randomization. It should reach lower total network power at equal admitted users, and it should run without interior-point solvers. We also predict that the advantage concentrates on instances where the outer sparsity loop shuts down radio heads or users, since that is where the iterates land on rank-deficient points. The core problem, for a fixed admitted set, is
$$ \min_{\{V_k \in \mathbb{C}^{N\times r}\}} \; \sum_k \frac{\|V_k\|_F^2}{\eta} + \sum_{n} P^c_n\, \mathbb{1}\Big[\textstyle\sum_k \|V_k^{(n)}\|_F > 0\Big] \quad \text{s.t.}\quad \frac{\operatorname{tr}(H_i V_k V_k^H)}{\sum_{j\neq k}\operatorname{tr}(H_i V_j V_j^H)+\sigma^2} \ge \gamma_i \;\; \forall i \in \mathcal{G}_k . $$
Here $N$ is the total number of transmit antennas and $r$ is the working rank. $V_k^{(n)}$ is the block of rows belonging to radio head $n$, $P^c_n$ is its circuit and fronthaul power, $\eta$ is amplifier efficiency, $H_i = h_i h_i^H$ is user $i$'s channel outer product, $\mathcal{G}_k$ is the set of users in multicast group $k$, $\gamma_i$ is the target signal-to-interference-plus-noise ratio, and $\sigma^2$ is the noise power. With $r = N$ this is the SDR. With $r = 1$ it is the original beamforming problem. The indicator term is handled by reweighted group penalties, as in Group Sparse Beamforming for Green Cloud-RAN.
Research Program
The algorithm has three layers. The outer layer reuses the iterative reweighting of group norms and the admission penalties from the smoothed-$\ell_p$ paper and the group-sparse beamforming paper. It is taken unchanged so that any gain can be attributed to the inner solver. The inner layer replaces each semidefinite program with optimization over the factors $V_k$. We use the lifted trust-region method with factor rebalancing from the bounded-rank stationarity paper, with the SINR constraints handled by an augmented Lagrangian. We start at a moderate rank $r$, where the factored landscape is better behaved, and decrease it on a schedule. This direction is the reverse of the rank-increasing strategy of Low-Rank Matrix Completion for Topological Interference Management by Riemannian Pursuit. That paper is the closest precedent: it uses fixed-rank manifold optimization in a wireless problem where convex proxies collapse. However, it optimizes a topology-only completion with no power, fronthaul, or on/off variables.
The third layer is certification. For a fixed active set, the inner problem has a linear objective and linear inequalities in $W_k$ over a positive semidefinite domain, which is exactly the low-rank spectral setting of On the partial convexification for low-rank spectral optimization: rank bounds and algorithms. We use that paper's extreme-point rank-reduction step to bound the rank of the returned solution without solving an SDP. We use it only as a post-processor and a diagnostic. Its column-generation solver is a natural competitor for the inner layer, so we run it as a baseline as well. The analysis question is whether the rank bound, which grows with the number of constraints in each group, predicts when rank one is attainable. That would tell practitioners when randomization is unavoidable.
The diagnostic contribution is to instrument every inner solve. Each time the outer loop zeroes a radio-head block or a group, we log the rank of the iterate and the first-order stationarity measure. We also log whether a second-order step finds further descent. This gives a direct count of how often the boundary trap described in the stationarity paper occurs in a real engineering problem rather than in constructed examples.
Baselines
The naive reference is decoupled allocation: all radio heads on, users associated by signal strength, then minimum-power multicast beamforming by SDR with randomization. The cross-layer paper (Cross-Layer Resource Allocation With Elastic Service Scaling in Cloud Radio Access Network) and the group-sparse paper both report that such decoupled schemes use much more power. Group sparse beamforming reports "up to 40% reduction in total power consumption under low QoS" and about 20% at high load relative to non-joint schemes. Our method must at least reproduce gaps of this kind against the naive reference, or our simulator differs from theirs.
The strongest existing method is the smoothed-$\ell_p$ algorithm with SDR inner solves and Gaussian randomization, from the smoothed-$\ell_p$ paper. It addresses exactly this joint problem (multicast, radio-head selection, admission control), and it is the method whose rank relaxation we replace. We found no reported numerical result on Lacuna for its power on the standard setups, so we will reimplement it on identical channel draws. It will get the same outer reweighting schedule, the same $p$ and smoothing parameters chosen by an equal grid search, and a randomization budget matched in wall-clock time to our inner solver. Beating it means lower mean total power at an equal or larger number of admitted users across QoS targets, or equal power at clearly lower runtime.
A second strong competitor for the inner solve is column generation for the partial convexification. The low-rank spectral optimization paper reports solving MIMO instances in "1–2 seconds" where MOSEK took over 180 seconds or failed within an hour. We run it inside the same outer loop. On small instances, branch-and-bound with an imitation-learned pruning policy (Scalable Network Adaptation for Cloud-Rans: An Imitation Learning Approach) gives a near-optimal reference; that paper reports over 20x speedup with less than 0.6% objective loss versus full branch-and-bound. We use plain branch-and-bound on the smallest networks to measure optimality gaps.
Two ablations isolate the key ingredient. The first keeps our factorization and rank schedule but replaces the second-order trust region with first-order projected gradient. The second fixes $r = 1$ from the start, which is direct beam optimization with no lifting.
Why not just…?
Why not just keep SDR and randomize? For unicast this may be fine. In multicast with several users per group, however, the relaxation is not tight; that is why the smoothed-$\ell_p$ paper needs randomization at all. Its cost also grows with the size of the semidefinite programs, which that paper identifies as the bottleneck. Our main comparison measures both effects separately: power gap at matched time, and time at matched power. The rank-reduction diagnostic shows on which instances the relaxation was loose.
Why not just optimize the beam vectors directly ($r = 1$) with a first-order method? This is the real threat, because it is the simplest non-SDP option. If the boundary trap is rare in practice, plain gradient methods on beam vectors could match everything we build. The $r = 1$ and first-order ablations settle this. If they match our method in power, runtime, and failure rate, we will report that the stationarity pathology is irrelevant for green Cloud-RAN. That would be a useful negative result, and the method would reduce to a simpler recommendation. We expect the gap, if any, to sit specifically on instances with many shutdown events, and the logging lets us check that prediction rather than average it away.
Why not just learn the discrete decisions and leave the beams to SDR? Imitation-learned pruning speeds up the search over on/off patterns, but every node it evaluates still needs a continuous solve. That paper also notes that the learned policy is bounded by the cost of its offline oracle. These approaches are complementary: our inner solver could replace the per-node solve. We will measure per-node solve time in both cases to quantify this, without claiming to beat learned search at discrete selection.
Proposed Experiments
We use the synthetic 3GPP-style Cloud-RAN setups from the file: 10 radio heads with 15 users and 6 radio heads with 8 users, with Rayleigh fading and path loss as in the group-sparse and imitation-learning papers. Users are grouped into multicast groups of two to five. We sweep the SINR target from low to overload, so that admission control becomes active. We also scale to denser networks to stress runtime. Metrics are total network power (transmit plus circuit and fronthaul), number of admitted users, wall-clock time, the fraction of instances that end exactly rank one without randomization, the optimality gap to branch-and-bound on the 6-by-8 instances, and the logged count of shutdown events at which first-order stationarity vanished but a second-order step still found descent.
The decisive check is a stratified comparison. We split instances by the number of radio-head or group shutdowns along the outer loop. The hypothesis is supported if the advantage of the second-order factored method over both SDR and the first-order ablation grows with this count. The hypothesis is refuted if the gap is flat or absent. A secondary check is whether the extreme-point rank bound from the partial convexification paper, evaluated per group, predicts which instances still need randomization under the SDR baseline.
References
- Smoothed Lp-Minimization for Green Cloud-RAN With User Admission Control
- Finding stationary points on bounded-rank matrices: a geometric hurdle and a smooth remedy
- Group Sparse Beamforming for Green Cloud-RAN
- Low-Rank Matrix Completion for Topological Interference Management by Riemannian Pursuit
- On the partial convexification for low-rank spectral optimization: rank bounds and algorithms
- Cross-Layer Resource Allocation With Elastic Service Scaling in Cloud Radio Access Network
- Scalable Network Adaptation for Cloud-Rans: An Imitation Learning Approach