Perhaps stupid question.
Question: Can "almost supermartingale" theorem be equally applicable to prove the convergence of some algorithms solving non-random optimization problems?
Attempt for a non-random version of "almost supermartingale" theorem (without any proof):
Let the non-negative sequences be $\{V^k\}$, $\{S^k\}$, and $\{ U^k \}$ for $k=0,1,2,\ldots$. Let $\beta_0, \beta_1$ be non-negative scalars satisfying $\sum_{k=0}^\infty \beta_k < \infty$. Assume $$V^{k+1} \leq \left( 1 + \beta_k \right) V^k - S^k + U^k$$ and $$\sum_{i={\color{red}{0 \text{ or } 1?}}}^\infty U^i < \infty.$$ Then,
- $V^k \rightarrow V^\infty$
- $\sum_{k=0}^\infty S^k < \infty$.
EDIT:
Additionally, let $V^k := \| x^k - x^\star \|_P^2$ where $P$ is a positive semidefinite matrix. Then, using the above theorem, can one say for a subsequence $n_k$ $\lim_{k \rightarrow \infty} x^{n_k} = x^\star$? or do we need more information to prove this?
For the completeness, I have also copied the "almost supermartingale" theorem from [Theorem 30, page 300].
Theorem: Let $V^k$, $S^k$, and $ U^k$ be $\mathcal{F}_k$-measurable random variables satisfying $V^k \geq 0$, $S^k \geq 0$, and $ U^k\geq 0$ for $k=0,1,2,\ldots$. Let $\beta_0, \beta_1$ be non-negative scalars satisfying $\sum_{k=0}^\infty \beta_k < \infty$. Assume $$\mathbb{E}\left[V^{k+1} \mid \mathcal{F}_k \right] \leq \left( 1 + \beta_k \right) V^k - S^k + U^k$$ and $$\sum_{i={\color{red}{0 \text{ or } 1?}}}^\infty U^i < \infty$$ almost surely. Then,
- $V^k \rightarrow V^\infty$
- $\sum_{k=0}^\infty S^k < \infty$.