fednova

**FedNova** (Federated Normalized Averaging) is a **federated learning algorithm that normalizes client updates to account for different numbers of local steps** — fixing the objective inconsistency in FedAvg where clients performing different amounts of local work contribute disproportionately to the global model. **How FedNova Works** - **Problem**: In FedAvg, a client doing 10 local steps has 10× more influence than one doing 1 step. - **Normalization**: Divide each client's update by its number of local steps: $Delta_k / au_k$. - **Effective Learning Rate**: Normalize out the accumulated learning rate from multiple local SGD steps. - **Aggregation**: Server aggregates normalized updates: $w_{t+1} = w_t - eta_g sum_k p_k (Delta_k / au_k)$. **Why It Matters** - **Objective Consistency**: FedNova provably converges to the correct solution, unlike FedAvg with heterogeneous local steps. - **System Heterogeneity**: Clients with different compute power can run different numbers of local steps without biasing the result. - **Drop-In**: Simple modification to FedAvg — just divide by local step count. **FedNova** is **fair averaging across unequal work** — normalizing client updates to prevent faster clients from dominating the global model.