Federated Learning (FL) offers a powerful paradigm for privacy-preserving distributed machine learning: training a global model across thousands of edge devices while keeping raw data localized. However, deploying FL in real-world mobile edge networks reveals a severe bottleneck: latency.
Standard FL algorithms, such as Federated Averaging (FedAvg), typically treat the training process as two isolated phases: a computation phase where devices run local training steps, and a communication phase where devices upload their updates over wireless channels. This disjointed view creates a fundamental inefficiency. If a device performs too few local updates, it sends noisy, premature gradients, requiring more global communication rounds to converge. Conversely, if it performs too many local updates, it spends excessive time and energy on computation, and the global model updates stall.
As highlighted by Pavana Prakash et al. in To Talk or to Work: Delay Efficient Federated Learning over Mobile Edge Devices, failing to jointly optimize these two domains leads to massive operational delays. The core insight of recent research is that communication and computation are not independent steps; they are deeply coupled "knobs" that must be co-optimized to minimize total training time and energy consumption.
The Mathematical Coupling of Talk and Work
To understand why joint optimization works, we can look at how the hyperparameters of machine learning interact directly with the physical constraints of wireless networks. Consider the total training time $T_{\text{total}}$ required to reach a target model accuracy. It can be modeled conceptually as:
$$T_{\text{total}} = N_{\text{rounds}} \times (T_{\text{comp}} + T_{\text{comm}})$$
Where $N_{\text{rounds}}$ is the number of global communication rounds needed for convergence, $T_{\text{comp}}$ is the local computation time per round, and $T_{\text{comm}}$ is the wireless transmission time per round. Tweaking a single variable ripples through this entire equation in opposing directions:
Local Iterations ($E$) and Client Participation ($K$)
Increasing the number of local iterations $E$ or adjusting the local batch size increases $T_{\text{comp}}$, but it generally decreases $N_{\text{rounds}}$ because each communication step carries more refined information. Similarly, sampling more active clients ($K$) per round can stabilize convergence (reducing $N_{\text{rounds}}$), but it increases the communication bottleneck because the server must wait for the slowest device (the straggler) to finish uploading.
Bing Luo 0002 et al. in Cost-Effective Federated Learning Design and Cost-Effective Federated Learning in Mobile Edge Networks analytically link $K$ and $E$ to both system costs and convergence bounds. The authors demonstrate that a sampling-based multivariate control approach can dynamically find the sweet spot between these coupled variables, achieving near-optimal performance that rivals exhaustive search methods.
Quantization and Compression Levels
Another critical coupling occurs between message precision and synchronization frequency. Compressing model updates (via quantization or sparsification) drastically reduces $T_{\text{comm}}$. However, highly compressed updates introduce noise, which increases $N_{\text{rounds}}$.
Milad Khademi Nori et al. in Fast Federated Learning by Balancing Communication Trade-Offs argue that isolating data compression from communication frequency fails because the optimal precision changes dynamically based on how often the global model synchronizes. To resolve this, Rui Chen et al. introduce a framework in Service Delay Minimization for Federated Learning over Mobile Devices that jointly optimizes local training iterations, weight quantization levels, and gradient quantization levels. By treating local computation and wireless transmission as coupled bottlenecks, the system balances the computational cost of local updates with the communication cost of transmitting parameters, reducing total service delay by up to 56%.
Physical Layer Integration and Hardware Awareness
In real-world mobile networks, edge devices are highly heterogeneous. A high-end smartphone computes faster and enjoys better channel conditions than a remote IoT sensor. Joint optimization frameworks bridge the gap between abstract machine learning hyperparameters and physical layer resources like CPU frequency ($f_i$), transmit power ($P_i$), and wireless channel states.
+-----------------------------------------------------------------+
| Central Server |
+-----------------------------------------------------------------+
^ ^
| Joint Optimization: |
| - Client Scheduling & Data Offloading |
| - Quantization & Compression Levels |
v v
+-----------------------+ +-----------------------+
| High-End Device | | Constrained Device |
| (High CPU, Good Ch) | | (Low CPU, Poor Ch) |
| | | |
| - High CPU Freq (f) | <--- D2D Offloading --- | - Low CPU Freq (f) |
| - High Transmit (P) | [Hosseinalipour et al.]| - Low Transmit (P) |
+-----------------------+ +-----------------------+
For instance, Zhiren Han et al. in Resource Allocation for Intelligent Reflecting Surfaces Assisted Federated Learning System with Imperfect CSI minimize total training time by jointly optimizing user selection, CPU frequency, transmit power, and Intelligent Reflecting Surface (IRS) phase shifts. By defining the per-round delay as the maximum of local computation and communication times across all users, they prevent stragglers from derailing the training process.
Furthermore, wireless channels are noisy and prone to packet errors. Mingzhe Chen et al. demonstrate in A Joint Learning and Communications Framework for Federated Learning Over Wireless Networks that accounting for packet errors during the joint optimization of FL parameters and wireless resources yields significant gains in model identification accuracy compared to methods that optimize only one aspect. To handle hardware diversity dynamically, Peichun Li et al. propose AnycostFL in AnycostFL: Efficient On-Demand Federated Learning over Heterogeneous Edge Devices, which dynamically addresses the specific bottleneck of a device—whether it is training time or data transmission—to maximize participation without exceeding strict latency and energy budgets.
Intelligent Scheduling and Data Offloading
Traditional random client selection ignores both the value of a client's data and its current wireless link quality. Intelligent scheduling policies must simultaneously consider wireless channel quality and the importance (e.g., gradient magnitude or loss) of local model updates.
As shown by Yinghui He et al. in Scheduling for Cellular Federated Edge Learning With Importance and Channel Awareness, balancing these two factors accelerates convergence, whereas relying on only one leads to suboptimal training performance. Data heterogeneity also plays a vital role; Jiaxiang Geng et al. argue in Adaptive Federated Learning in Heterogeneous Wireless Networks with Independent Sampling that true efficiency requires a joint optimization considering computation time, communication time, and data heterogeneity simultaneously, rather than just minimizing the number of rounds.
Taking this a step further, Seyyedali Hosseinalipour et al. show in Device Sampling and Resource Optimization for Federated Learning in Cooperative Edge Networks that device sampling can be jointly optimized with device-to-device (D2D) data offloading. By using a Graph Convolutional Network (GCN) to learn optimal sampling and offloading strategies, they allow resource-constrained devices to offload computational burdens to nearby peers, maximizing overall FL efficiency.
Advanced Paradigms: Multi-Service and Blockchain FL
The joint optimization paradigm extends naturally to more complex architectures, such as multi-task environments and decentralized blockchain networks:
- Multi-Service Edge Networks: When multiple FL tasks compete for the same pool of hardware resources, Minh N. H. Nguyen et al. in Toward Multiple Federated Learning Services Resource Sharing in Mobile Edge Networks introduce a mechanism that simultaneously optimizes hardware resources (CPU frequency, bandwidth) alongside training hyperparameters like the learning rate. By treating the learning rate as a controllable variable, the system minimizes the joint cost of energy and time across multiple competing services. For asynchronous or multi-model systems, Seyyedali Hosseinalipour et al. demonstrate in Asynchronous Multi-Model Dynamic Federated Learning Over Wireless Networks: Theory, Modeling, and Optimization that optimizing device scheduling is as critical as resource allocation to balance model performance with energy consumption under dynamic data conditions.
- Blockchain-Aided FL: When FL is integrated with blockchain to eliminate single points of failure, a third bottleneck emerges: mining and block generation delay. Jun Li et al. address this in Blockchain-Aided Wireless Federated Learning: Resource Allocation and Client Scheduling by using Lyapunov optimization to jointly optimize client computation frequency, mining frequency, and client scheduling under long-term energy budgets. To manage the massive communication overhead, Laizhong Cui et al. in A Fast Blockchain-based Federated Learning Framework with Compressed Communications formulate a bi-convex optimization problem that balances the model compression ratio against block generation frequency, minimizing total training loss within a fixed time budget.
Why Joint Optimization Works
Why does joint communication-computation optimization yield such massive speedups? Because it breaks the artificial wall between the machine learning algorithm and the underlying network fabric.
By recognizing that hyperparameter choices (like local iterations, batch size, and quantization) directly dictate network traffic, and that network allocations (like transmit power and CPU frequency) directly dictate training speed, these frameworks find a global equilibrium. Instead of letting a poor wireless channel stall an entire training round, the system adapts—either by adjusting the local work, compressing the payload, or reallocating hardware resources. The result is a highly adaptive, latency-aware federated learning ecosystem capable of training robust models on resource-constrained edge hardware.
Key Foundational Work
To Talk or to Work: Delay Efficient Federated Learning over Mobile Edge Devices (2021) — contributed 4 concept elements to this direction.
Performance Optimization of Federated Learning over Mobile Wireless Networks (2020) — contributed 3 concept elements to this direction.
Resource Allocation for Intelligent Reflecting Surfaces Assisted Federated Learning System with Imperfect CSI (2021) — contributed 2 concept elements to this direction.