JCPGNN-M: GNN-Based Joint Channel & Power Allocation

• The paper introduces JCPGNN-M, a method that leverages graph neural networks for joint channel and power allocation while enabling simultaneous multi-channel occupancy.
• It reformulates resource allocation by converting discrete channel choices into continuous power vectors, facilitating efficient spectrum reuse and reduced inference time.
• The approach integrates Lagrangian-based primal-dual optimization to enforce power and QoS constraints, demonstrating superior runtime, scalability, and robustness.

JCPGNN-M is a graph neural network-based method for joint channel and power allocation in interference-limited multi-channel wireless networks. It was introduced as a scalable alternative to iterative optimization for maximizing weighted sum rate under per-user total power constraints, while allowing simultaneous multi-channel allocation for each user and efficient spectrum reuse (Chen et al., 4 Jun 2025). A later formulation extends the method to user-specific Quality-of-Service (QoS) constraints by embedding the GNN in a Lagrangian-based primal-dual optimization framework, with the stated goal of ensuring satisfaction of QoS constraints and convergence to a stationary point (Chen et al., 8 Sep 2025).

1. Origin and conceptual scope

JCPGNN-M emerged from a line of GNN-based wireless resource allocation methods in which the wireless network is represented as a graph and resource decisions are learned by message passing over interference structure. Its immediate predecessor, JCPGNN, addressed joint channel and power allocation in heterogeneous wireless networks by modeling each potential user–channel allocation as a node and using a shared message computation layer with two task-specific heads for channel and power allocation (Chen et al., 2024). In that earlier formulation, each transceiver pair could use at most one channel.

JCPGNN-M changes the problem class. Instead of restricting each user to a single channel, it enables simultaneous multi-channel allocation for each user and supports efficient spectrum reuse, with several users per channel and users on multiple channels (Chen et al., 4 Jun 2025). In the QoS-constrained formulation, the method is described as a scalable and theoretically grounded solution for constrained resource allocation in future wireless networks (Chen et al., 8 Sep 2025). This shift is structurally important: the allocation variable is no longer primarily a one-hot channel choice plus a scalar power, but a per-user vector of powers across channels.

A plausible implication is that JCPGNN-M is best understood not as a minor architectural revision of JCPGNN, but as a reformulation of the underlying resource-allocation problem so that channel access is induced by nonzero power decisions rather than imposed by a discrete single-channel assignment.

2. Optimization problem and relaxation

The multi-channel formulation considers $D$ transceiver pairs and $M$ orthogonal channels. In the unconstrained version, the objective is to maximize weighted sum rate subject to nonnegativity of powers, binary channel-assignment variables, and a per-user total power budget (Chen et al., 4 Jun 2025): $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$

Because users are allowed simultaneous multi-channel access, the formulation can be simplified by dropping the channel-allocation variables and focusing directly on power: $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$ In the QoS-constrained version, the same reduction is used, but each user must also satisfy a minimum total rate (Chen et al., 8 Sep 2025): $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$

This relaxation is central to JCPGNN-M. The papers state that users’ channel activity is implicitly determined by $p_i^m>0$, so channel selection is encoded by sparse or zero-valued power outputs rather than by an explicit discrete combinatorial variable (Chen et al., 8 Sep 2025). This suggests that JCPGNN-M converts a mixed-integer joint channel-and-power problem into a continuous structured power-allocation problem whose solutions still induce channel usage patterns.

3. Graph model and message-passing architecture

JCPGNN-M represents the network as a collection of channel-wise subgraphs. For each channel $m$, a subgraph $\mathcal{G}^m$ is constructed; nodes correspond to users on that channel, denoted $v_i^m$, and edges encode pairwise interference (Chen et al., 4 Jun 2025). In the QoS-constrained presentation, each subgraph $\mathcal{G}^m$ connects all users sharing channel $M$0, with edges weighted by $M$1; node features include direct channel gain and user-specific QoS $M$2, and edges encode interference coefficients (Chen et al., 8 Sep 2025).

The core GNN computation follows a standard message-passing pattern. At layer $M$3, message computation and aggregation are written as

$M$4

$M$5

followed by an update

$M$6

The mapping $M$7 is described as typically an MLP with sigmoidal activation to bound output in $M$8 (Chen et al., 4 Jun 2025).

The architectural emphasis differs from the earlier JCPGNN. JCPGNN used a shared message computation layer plus separate heads for one-hot channel allocation and scalar power allocation (Chen et al., 2024). JCPGNN-M instead uses message computation, aggregation, and post-processing to output a multi-channel power vector per user (Chen et al., 4 Jun 2025). The change is not merely quantitative; it aligns the architecture with a setting in which channel reuse and simultaneous multi-channel occupancy are first-class design objectives.

4. Lagrangian formulation and primal-dual training

A defining feature of JCPGNN-M is its use of a Lagrangian objective rather than a heuristic penalty-only loss. In the power-constrained version, the resource-allocation problem is written as

$M$9

with $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$0 enforcing the per-user power constraints (Chen et al., 4 Jun 2025). In the QoS-constrained extension, the Lagrangian becomes (Chen et al., 8 Sep 2025)

$\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$1

where $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$2 are dual variables for the QoS constraints.

Training alternates between updating GNN parameters and updating dual variables. The primal update is stated as

$\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$3

while the projected dual updates are

$\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$4

$\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$5

with $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$6 (Chen et al., 8 Sep 2025).

JCPGNN-M also includes a post-processing normalization step to enforce the global power constraint: $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$7 The papers describe this as ensuring per-user sum-power constraints by normalization (Chen et al., 8 Sep 2025).

The later QoS-constrained paper explicitly contrasts this construction with “reward+penalty” training, stating that standard deep learning with constraint penalties offers no theoretical guarantee of constraint satisfaction or convergence, whereas training the GNN within the Lagrangian framework ensures satisfaction of QoS constraints and convergence to a stationary point (Chen et al., 8 Sep 2025). That claim defines the principal theoretical distinction of the QoS-aware JCPGNN-M formulation.

5. Empirical properties

The empirical evaluations report several recurring properties: competitive sum rate relative to iterative optimization, substantially lower inference time, generalization to larger graphs, and robustness under imperfect CSI. The exact numerical statements differ slightly between the two JCPGNN-M papers and should be read in that context.

In the multi-channel formulation without QoS constraints, simulation results show that JCPGNN-M achieves better data rate compared to eWMMSE, and that the inference time of JCPGNN-M is much lower than eWMMSE (Chen et al., 4 Jun 2025). The detailed summary states that JCPGNN-M achieves higher or comparable sum rate than eWMMSE, especially as the number of channels increases, and that inference time is only about $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$8 of that of eWMMSE (Chen et al., 4 Jun 2025). The same source states that the network trained on small-scale cases performs near-optimally, within $\begin{aligned} \underset{\mathbf{C},\,\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M} \sum_{i=1}^{D} \alpha_{i} R_{i}^m(\mathbf{C},\mathbf{P}) \ \text{subject to}\qquad & p_i^m\geq0,\quad\forall i,m\ & c_{i}^m \in \{0,1\},\quad\forall i,m\ & \sum_{m=1}^{M} c_{i}^m p_{i}^m\leq P_{\max },\quad\forall i . \end{aligned}$9–$\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$0, on much larger topologies.

In the QoS-constrained formulation, the comparison target is the enhanced WMMSE (eWMMSE) algorithm, which is described as provably convergent to a locally optimal solution when the problem is feasible (Chen et al., 8 Sep 2025). There, JCPGNN-M is reported to achieve comparable or even slightly better average sum rates than eWMMSE, with less than $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$1–$\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$2 degradation in normalized sum rate even when generalizing to much larger networks. For QoS satisfaction, the violation probability gap to eWMMSE is stated to be less than $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$3 for most practical QoS loads. Convergence is reported at approximately $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$4 steps for JCPGNN-M versus $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$5 for prior penalty-based GNNs, while eWMMSE requires $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$6 block coordinate updates per test instance. Runtime is reported as up to approximately $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$7 faster than eWMMSE for large-scale scenarios. Generalization is illustrated by training on $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$8 and evaluating at $\begin{aligned} \underset{\mathbf{P}}{\operatorname{maximize}\quad & \sum_{m=1}^{M}\sum_{i=1}^{D} \alpha_i R_i^m(\mathbf{P}) \ \text{subject to } & p_i^m \ge 0,\quad \forall i,m \ & \sum_{m=1}^M p_i^m \leq P_{\max},\quad \forall i . \end{aligned}$9, where performance remains $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$0 of best possible. Under imperfect CSI, JCPGNN-M is reported to exhibit a sum-rate drop of less than $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$1 even if $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$2 of CSI entries are missing (Chen et al., 8 Sep 2025).

These experiments place JCPGNN-M in a specific empirical niche. It is not presented as an exact optimizer; rather, it is presented as a learned surrogate whose forward pass approximates the quality of iterative solvers while changing the latency regime from repeated per-instance optimization to one-shot inference.

6. Relationship to JCPGNN and to prior resource-allocation methods

The contrast with JCPGNN clarifies what JCPGNN-M adds. JCPGNN addressed heterogeneous wireless networks in which each pair uses at most one channel, and its architecture used a shared message computation layer plus separate task-specific layers for channel and power allocation (Chen et al., 2024). It achieved $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$3 of the optimal sum-rate for $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$4 or $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$5 with $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$6 channels, was reported as $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$7 times faster than exhaustive optimal for $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$8, and was $$\begin{aligned} &\underset{\mathbf{P}}{\max}\quad \sum_{m=1}^{M}\sum_{i=1}^{D}\alpha_{i} R_{i}^m(\mathbf{P}) \ &\text{s.t.}\;\; \sum_{m=1}^{M} p_{i}^{m}\leq P_{\max},\; p_i^m\geq0,\; \sum_{m=1}^{M} R_{i}^{m}(\mathbf{P}) \geqslant R_i^{\min}. \end{aligned}$$9 faster than the WMMSE-based Closest method. It also retained over $p_i^m>0$0 of the performance of traditional optimization methods when tested on larger or denser networks and retained over $p_i^m>0$1–$p_i^m>0$2 of its optimal sum-rate with up to $p_i^m>0$3 corrupted channel measurements (Chen et al., 2024).

JCPGNN-M inherits the graph-based inductive bias and the emphasis on computational efficiency, but changes the feasible set. The papers explicitly state that prior GNN-based methods limited each user to a single channel, while JCPGNN-M supports simultaneous multi-channel allocation for each user (Chen et al., 4 Jun 2025). They also compare against ICP-GNN, which allocates each channel independently and distributes user power as $p_i^m>0$4 across $p_i^m>0$5 channels, and against heuristic or single-channel methods that do not exploit full spectrum reuse (Chen et al., 4 Jun 2025).

In the QoS-constrained work, the principal methodological comparison is with penalty-based unsupervised GNNs. The stated criticism is that heuristic penalty terms require empirical hyperparameter tuning and offer no theoretical convergence guarantees; the stated alternative is a principled framework that integrates the GNN with a Lagrangian-based primal-dual optimization method (Chen et al., 8 Sep 2025). This is the main conceptual dividing line in the later literature on JCPGNN-M: not only whether a GNN is used, but whether the learning procedure is coupled to an explicit constrained optimization framework.

7. Significance and limitations of interpretation

Within the literature represented by these papers, JCPGNN-M serves as a graph-based approximation framework for constrained radio resource management in multi-channel interference networks. Its recurring claims are scalability, support for spectrum reuse, generalization to larger networks than seen during training, and robustness under incomplete or inaccurate CSI (Chen et al., 4 Jun 2025). The QoS-constrained version adds the explicit claim that the method ensures satisfaction of QoS constraints and convergence to a stationary point through the primal-dual Lagrangian formulation (Chen et al., 8 Sep 2025).

At the same time, the method’s guarantees are tied to the precise optimization framework stated in the papers. The locally optimal solution guarantee belongs to eWMMSE when the problem is feasible, whereas JCPGNN-M is described as matching eWMMSE while being substantially faster (Chen et al., 8 Sep 2025). Accordingly, the literature positions JCPGNN-M as a learned constrained optimizer with strong empirical fidelity to iterative baselines rather than as a proof of global optimality.

A plausible implication is that JCPGNN-M is most significant where per-instance optimization latency is itself the bottleneck: dense networks, dynamic topologies, or real-time wireless control. In those settings, the combination of message passing, post-processed power normalization, and primal-dual training gives the method a distinctive role within GNN-based wireless resource allocation.