The proof of Lemma 1 is given in Appendix.
Remark 1: Part 1 of Lemma 1 indicates that, within one iteration
of the proposed algorithm, if the stepsizes are sufficiently
small (meaning the dual variables ϕ(t + 1) and ϕ(t) are not
far apart), the primal variables q(t) and q(t) are not far apart.
Similarly in Part 2, with sufficiently small stepsizes, the real
part of the cross product (∇f(q(t))−∇f(w∗))H(q(t)−w∗)
will not be far below zero. Notice that if ϕ(t+1) = ϕ(t) then
q(t) = q(t), and Part 2 degenerates into the common result
(34) for convex functions. Hence Part 2 implies that ∇f(q(t))
will not be far away from ∇f(q(t)), either.
Now we are ready to present a brief proof for Theorem 1.
Proof for Theorem 1: First, a metric is needed to measure
the “distance” between the intermediate solution (ϕ(t),w(t))
and the stationary point (ϕ∗,w∗). For this purpose, we define
the Lyapunov function