FERCOP: Communication-Aware Optimization

• FERCOP is a family of optimization problems that treat communication as a primary variable, integrating network rounds, power allocation, and reliability constraints.
• It unifies multiple formulations by addressing trade-offs between communication cost, error rates, and resource usage across federated, edge, and wireless networks.
• Practical implementations employ event-triggered transmission, Lyapunov control, and local multi-step updates to enhance efficiency and fairness in diverse settings.

Searching arXiv for the cited works to ground the article in current metadata. {"query":"(Wu et al., 2023) Federated Conditional Stochastic Optimization", "max_results": 5} {"query":"(Gatsis, 2021) Federated Reinforcement Learning at the Edge", "max_results": 5} {"query":"(Dakić et al., 2023) Frame Error Rate Prediction for Non-Stationary Wireless Vehicular Communication Links", "max_results": 5} {"query":"(Askin et al., 2024) Federated Communication-Efficient Multi-Objective Optimization", "max_results": 5} {"query":"(Perazzone et al., 1 Mar 2025) Communication-Efficient Device Scheduling for Federated Learning Using Lyapunov Optimization", "max_results": 5} The FER-Communication Optimization Problem (FERCOP) denotes a class of communication-aware optimization problems in which communication decisions are treated as first-class variables rather than as fixed overhead. In federated conditional stochastic optimization, it can be interpreted as the design of algorithms and parameter schedules that minimize communication rounds and total communication cost needed to achieve a target optimality for conditional federated objectives (Wu et al., 2023). In edge reinforcement learning, an equivalent viewpoint appears through a scalarized objective that jointly penalizes communication and final learning error (Gatsis, 2021). In imperfect and energy-constrained underwater wireless sensor networks, FERCOP is stated explicitly as a joint link-scheduling and power-allocation problem that maximizes a weighted utility of spatial reuse, fairness, and ineffective communication under energy, lifetime, SINR, and malfunction constraints (Zhang et al., 11 Aug 2025). Related federated optimization work further shows that communication-efficient formulations can also be organized around objective compression, client scheduling, and topology-aware message passing (Askin et al., 2024, Perazzone et al., 1 Mar 2025, Mota, 2013).

1. Terminology, scope, and conceptual variants

The term is not used with a single universal meaning across the literature. One usage arises in a federated / edge / remote setting, where communication efficiency is framed around rounds, local steps, and gradient-oracle complexity for distributed learning (Wu et al., 2023). A second usage appears in underwater networking, where “FER” means “fair, efficient, and reliable” communication and the optimization target is an explicitly weighted long-term network utility (Zhang et al., 11 Aug 2025). A third, distinct acronym occurs in vehicular wireless communications, where FER means frame error rate; that work does not formulate FERCOP directly, but it provides a learned FER predictor that can serve as a surrogate reliability oracle inside broader communication-optimization pipelines (Dakić et al., 2023).

This terminological overlap matters because the objects being optimized differ. In federated learning, the dominant communication variables are local-step frequency, client participation, gradient aggregation, and per-round upload size. In wireless networking, the dominant variables are link scheduling, transmit power, reliability, and network lifetime. In vehicular channel modeling, the central object is a mapping from time-varying channel transfer functions to FER classes, which is useful for coverage or reliability constraints rather than a standalone communication-control policy (Dakić et al., 2023).

Context	Core optimization target	Main communication variables
Federated conditional optimization	Target stationarity with minimal rounds and oracle use	$q$, $m$, $b$, $B$, gradient-state exchange
Edge reinforcement learning	Joint communication and value-function error	$\alpha_k^i$, trigger thresholds
Wireless FL with Lyapunov control	Joint convergence term and communication time	$q_n^t$, $\omega_n^t$, $P_n^t$
Imperfect UWSNs	Weighted utility of fairness, efficiency, reliability	Scheduling indicators, discrete powers

2. Canonical mathematical formulations

A central federated formulation is the conditional stochastic objective

$$\min_{x \in \mathcal{X}} F(x), \qquad F(x) := \frac{1}{N}\sum_{n=1}^N F^n(x),$$

with local conditional objective

$$F^n(x)=\mathbb{E}_{\xi^n}\Big[f_{\xi^n}^n\big(\mathbb{E}_{\eta^n\mid \xi^n} g_{\eta^n}^n(x,\xi^n)\big)\Big].$$

The distinctive feature is the conditional dependence $m$0, which separates federated conditional stochastic optimization from standard compositional objectives with independent inner and outer randomness. In this setting, a communication round occurs every $m$1 local steps, so the dominant communication metric is roughly $m$2, while computational cost is measured by first-order oracle calls (Wu et al., 2023).

In edge reinforcement learning, the problem is written directly as a joint performance metric. For two agents and communication indicators $m$3,

$m$4

blends average communication cost with final value-function fitting error. The associated trigger rule communicates only when the expected reduction in $m$5 is large enough relative to a threshold determined by $m$6 and $m$7, making communication an event-triggered decision variable rather than a fixed periodic action (Gatsis, 2021).

A wireless federated learning variant starts from a convergence bound under arbitrary participation probabilities $m$8 and constructs the per-round combined objective

$m$9

Here the first term is a convergence-related penalty derived from reciprocal participation probabilities, and the second is expected communication time under a Shannon-capacity uplink model. Long-term power constraints and sampling-with-replacement constraints convert this into a stochastic network optimization problem (Perazzone et al., 1 Mar 2025).

In imperfect and energy-constrained underwater wireless sensor networks, FERCOP is defined through the long-term utility

$b$0

subject to per-node energy constraints, a minimum lifetime $b$1, an SINR threshold $b$2, discrete power levels, and node-malfunction constraints. Spatial reuse measures the fraction of scheduled links that succeed, fairness is Jain’s index over successful deliveries, and ineffective communication penalizes transmissions that consume energy but fail (Zhang et al., 11 Aug 2025).

3. Algorithmic design patterns

Several methodological families recur across FERCOP formulations. In federated conditional optimization, communication reduction is achieved by combining biased but controllable conditional gradient estimators with local multi-step updates. FCSG uses synchronous averaging every $b$3 steps, FCSG-M adds local momentum on the gradient estimator, and Acc-FCSG-M introduces a SARAH-/SPIDER-like recursive correction to reduce variance without full-gradient computation. The design principle is a three-way trade-off among inner batch size $b$4, outer mini-batch size $b$5, and local-step count $b$6 (Wu et al., 2023).

In edge reinforcement learning, the dominant mechanism is event-triggered transmission. Each agent computes a local stochastic gradient estimate, evaluates the expected gain of communicating it, and transmits only if the gain crosses a threshold. The practical trigger replaces the unknown objective decrease by a quadratic approximation using the same mini-batch used for the gradient. This converts communication scheduling into a greedy local decision rule coupled to learning dynamics (Gatsis, 2021).

In wireless federated learning under constrained uplinks, the dominant mechanism is Lyapunov drift-plus-penalty. Virtual queues enforce long-term average power constraints, and the per-round optimizer jointly selects participation probabilities and transmit powers using instantaneous CSI rather than channel distributions. The inner power-allocation subproblem yields a Lambert-$b$7-based closed form, while the sampling distribution over devices is solved on the probability simplex using manifold optimization and Manopt (Perazzone et al., 1 Mar 2025).

In imperfect UWSNs, the dominant mechanism is deep multi-agent reinforcement learning. ICRL-JSA models joint scheduling and power allocation as a Dec-POMDP, uses DQN with GRU and VDN under centralized training with decentralized execution, and folds energy and malfunction robustness into reward shaping and episode termination. Its advanced training mechanism combines a performance evaluation module with a malfunction rate adjustment module, effectively creating a curriculum over failure rates (Zhang et al., 11 Aug 2025).

A broader communication-efficient optimization literature supplies additional structural templates. FedCMOO approximates the Gram matrix $b$8 of task gradients so that each client sends a single aggregated $b$9-dimensional update and a compressed Jacobian object with upload cost $B$0, independent of the number of objectives $B$1 (Askin et al., 2024). Earlier distributed optimization work based on D-ADMM exploits induced subgraphs, node coloring, and Steiner trees so that only variable coordinates shared by neighboring functions are communicated, which reduces message count in problems with sparse variable dependence (Mota, 2013).

4. Communication metrics, rates, and trade-offs

Communication is quantified differently across formulations, but the common theme is explicit coupling between task error and message exchange. In federated conditional optimization, the principal asymptotic metric is communication-round complexity. With $B$2, FCSG and FCSG-M attain communication complexity $B$3, while Acc-FCSG-M attains $B$4, along with sample complexity $B$5, matching the single-machine lower bound for conditional stochastic optimization (Wu et al., 2023).

In federated multi-objective optimization, the crucial quantity is per-round upload size rather than only the number of rounds. FedCMOO preserves the standard nonconvex $B$6 stationarity rate while keeping per-client upload cost $B$7, independent of $B$8; by contrast, prior FMOO approaches such as FedMGDA require $B$9 upload per client per round (Askin et al., 2024).

In the Lyapunov-based wireless FL setting, the communication cost is wall-clock communication time under TDMA uplink transmission,

$\alpha_k^i$0

and the learning-side penalty is the average reciprocal participation term $\alpha_k^i$1. The resulting trade-off parameter $\alpha_k^i$2 determines whether the system favors more uniform learning participation or more aggressive exploitation of devices with favorable channels (Perazzone et al., 1 Mar 2025).

In edge RL, the communication metric is the average number of transmissions,

$\alpha_k^i$3

embedded directly in the optimization objective. In underwater networks, communication quality is not reduced to round count alone; it is mediated by SINR-feasible simultaneous links, fairness of successful access, and ineffective transmissions that waste energy (Gatsis, 2021, Zhang et al., 11 Aug 2025).

5. Application domains and empirical evidence

The most developed federated-learning instantiation appears in conditional objectives such as MAML-type meta-learning, invariant logistic regression, and AUPRC maximization. In few-shot Omniglot experiments, FCSG, FCSG-M, and Acc-FCSG-M converge much faster than Local-SCGD and Local-SCGDM for the same communication budget. In federated online AUPRC on imbalanced MNIST and CIFAR-10, Acc-FCSG-M reports final average AP values of $\alpha_k^i$4 and $\alpha_k^i$5, compared with $\alpha_k^i$6 and $\alpha_k^i$7 for CODA+, with similar communication budgets (Wu et al., 2023).

The edge RL formulation is validated on grid-world and continuous-state value-function approximation. The theoretical trigger achieves a strong communication-performance trade-off, and the practical approximate-gain trigger remains substantially better than random communication. The same paper reports a scaling gain when moving from two agents to ten agents: learning becomes faster for similar average communication rate per agent (Gatsis, 2021).

Vehicular frame-error-rate prediction contributes a complementary reliability oracle rather than a scheduler. A fully connected DNN trained on $\alpha_k^i$8 labeled channel-response samples achieves $\alpha_k^i$9 classification accuracy on a held-out synthetic test set and $q_n^t$0 accuracy on measured urban V2V channels. The model maps a $q_n^t$1 complex CTF representation to four FER classes, which the paper explicitly connects to reliable coverage-area design and to optimization problems with FER constraints (Dakić et al., 2023).

Wireless federated learning with Lyapunov scheduling demonstrates the benefit of communication-aware participation. On CIFAR-10 with heterogeneous channels and negligible computation time, the proposed policy reaches $q_n^t$2 test accuracy in approximately $q_n^t$3 s, compared with approximately $q_n^t$4 s for uniform selection when $q_n^t$5, a reported $q_n^t$6 wall-clock speedup. The gain persists under heterogeneous data, though it is smaller when data become highly non-i.i.d. (Perazzone et al., 1 Mar 2025).

In underwater networks, ICRL-JSA is evaluated in both perfect and imperfect settings. In perfect UWSNs it achieves the highest combined utility among the reported baselines, with throughput up to approximately $q_n^t$7 kbps, fairness approximately $q_n^t$8, and delivery ratio approximately $q_n^t$9. In IC-UWSNs with malfunction rate $\omega_n^t$0, it maintains approximately $\omega_n^t$1 successful links per slot for five senders while preserving high fairness and delivery ratio (Zhang et al., 11 Aug 2025).

6. Assumptions, misconceptions, and related directions

A common misconception is that FERCOP refers to a single standardized optimization problem. The literature instead supports a family resemblance: communication is explicit in the objective or constraints, but the precise task loss varies from stationarity measures, to value-function error, to wall-clock communication time, to long-term fairness-efficiency-reliability utilities (Wu et al., 2023, Gatsis, 2021, Zhang et al., 11 Aug 2025). A second misconception is acronymic: “FER” may denote federated / edge / remote, fair / efficient / reliable, or frame error rate, and those meanings are not interchangeable (Dakić et al., 2023).

The main technical limitations are also formulation-specific. Federated conditional optimization is analyzed under synchronous, full-participation rounds and bounded heterogeneity, and its communication cost is measured primarily in rounds rather than bits (Wu et al., 2023). Edge RL theory is developed for one approximate value-iteration step with linear value-function approximation, positive definite feature covariance, and a two-agent theoretical focus (Gatsis, 2021). The Lyapunov FL framework requires strictly positive minimum participation probability and assumes channel-state knowledge at each round; its per-round manifold subproblem is nonconvex and solved only to a local optimum (Perazzone et al., 1 Mar 2025). The underwater formulation relies on a thresholded SINR-success model, discrete power levels, and empirically chosen curriculum parameters (Zhang et al., 11 Aug 2025).

Related work indicates two broader research directions. First, communication efficiency can be made structurally aware: D-ADMM and its extensions classify problems as global, star-shaped, mixed, connected, or non-connected through induced subgraphs, then reduce communication by exploiting sparsity and Steiner trees (Mota, 2013). Second, communication efficiency can be made objective-aware: FedCMOO shows that multi-objective federated learning can preserve $\omega_n^t$2 per-client upload independent of the number of objectives while retaining standard nonconvex convergence rates (Askin et al., 2024). This suggests that FERCOP is less a single model than a design principle: communication variables, statistical efficiency, and system constraints should be optimized jointly rather than in isolation.