1: \begin{definition}[Local iteration counter]

2: For each agent $v_i$, and all integers $k \geq 0$, define the local iteration counter

3: \[

4: 	c_i[k] \defeq \sum^{k}_{\ell = 0} \delta_i[\ell]

5: \]

6: to be the number of updates performed by agent $v_i$ in the time-interval $(t[0], t[k]]$.

7: By convention, for all $i \in [n]$, we take $\delta_i[0] \defeq 1$, and thus $c_i[0]= 1$.

8: 

9: \begin{corollary}[Convergence of Asynchronous Gradient Push for Known Update Rates]

10: \label{cor:const_congergence_opt_known_udpate_rates}

11: Suppose the assumptions made in Theorem~\ref{th:const_convergence_opt} hold, and suppose that each agent $v_i$ has prior knowledge of $c_i[K - 1]$, the number of local iterations it will have completed before time $t[K]$.

12: If each agent $v_i$ sets its local step-size scaling factor

13: \[

14: 	w_i \defeq \frac{K}{c_i[K - 1]} \geq 1,

15: \]

16: then

17: \begin{align*}

18: 	\frac{1}{K} \sum^{K -1}_{k=0} \norm{\xbar[k] - \xstar}^2 \leq& + \frac{1}{ K^\theta} \left( \frac{n (A_1 + A_3)}{2 \mu B} \right) +  \frac{1}{ K} \left( \frac{n A_2}{2 \mu B} \right) \\

19: 	& + \frac{1}{K^{1 - \theta}} \left(\frac{n \left( \norm{ \xbar[0] - \xstar }^2 \right)}{2 \mu B } \right),

20: \end{align*}

21: where $\xstar$ is the minimizer of~\eqref{eq:problem}.

22: \end{corollary}

23: Corollary~\ref{cor:const_congergence_opt_known_udpate_rates} states that if the agents know one another's update rates, then they can set their step-sizes to guarantee convergence to the unbiased global minimizer, even in the presence of persistent, but bounded, processing and message delays.

24: In particular, slower agents can simply scale up their step-size to compensate for their slower update rates.

25: 

26: We also provide guarantees for a version of the algorithm using diminishing step sizes.

27: 

28: \end{definition}

29: