https://cis.temple.edu/~jiewu/research/publications/Publication_files/FedCPD.pdf
0 2 2 2 L1 + E 0G2 + E2 0G2 2 + 2 L2E2 0G2 2 + 2 Theorem 2. (Non-convex FedCPD convergence). 0 < e < 0, e 2 f1 1; 2; : : : ; Eg, where represents the de- 2; cay factor for the learning rate. If the learning rate for each epoch satisfies the following condition, the loss function de-creases monotonically, leading to convergence:
https://cis.temple.edu/~jiewu/research/publications/Publication_files/Paper%206190%20Camera%20Ready%20Version.pdf
For notation, tindicates the communication round and e2 1=2;1;2;:::;Erefers to the local iterations, where Eis the total number of local updates. Thus, tE+ erepresents the e-th local iteration in the (t+ 1)-th round.