In the paper of the Adam Optimizer, the author states in the section 2.1 that the effective stepsize has two upper bounds: $\alpha \cdot (1- \beta_1) \ / \sqrt{1 - \beta_2}$ in the case $1 - \beta_1 > \sqrt{1 - \beta_2}$, and $\alpha$, otherswise. So the question is how can we prove this?
1 Answer
Glad someone is looking at this too. First we can write the $\hat{m}$ and $\hat{v}$ as follows:
$$ \hat{m}_t = \frac{1 - \beta_1}{1 - \beta_1^t} \sum_{j=1}^t \beta_1^{t - j} g_j.$$ $$ \hat{v}_t = \frac{1 - \beta_2}{1 - \beta_2^t} \sum_{j=1}^t \beta_2^{t - j} g_j^2. $$
if $1 - \beta_1 > \sqrt{1-\beta_2}$, then $ 1 - 2 \beta_1 + \beta_1^2 > 1 - \beta_2 \Rightarrow \beta_2 > 2 \beta_1 - \beta_1^2$, from which $\beta_1 < \beta_2$, since $\beta_i \in [0, 1]$.
To get the first bound, we just need to show $$\frac{ \sum_{j=1}^t \beta_1^{t - j} g_j}{1 - \beta_1^t} \leq \sqrt{\frac{ \sum_{j=1}^t \beta_2^{t - j} g_j^2}{1 - \beta_2^t}}.$$ Without the square root on the right hand side, the inequality above is clear from the relation $\beta_1 < \beta_2$. With the square root, if the right hand side is $\leq 1$, it would also be clear. But we can simply scale down the $g_j$'s so that the right hand side is indeed $\leq 1$ without changing the conclusion.
I think the second bound is indeed false. Consider $t = 2$ and $g_2 = 0$, then the assertion reduces to showing that $1 - \beta_1 \leq \sqrt{1-\beta_2}$ implies
$$ \frac{\beta_1}{1 + \beta_1} < \sqrt{\frac{\beta_2}{1 + \beta_2}}.$$
But I can simply take $\beta_1 = 1$ and $\beta_2 = 0$, to satisfy the condition. However the inequality above is clearly false in this case.
-
$\begingroup$ Someone asked a clarifying question a while ago but I didn't get a chance to answer. I think the derivation above may indeed contain an error. Will get to that later. $\endgroup$ Commented Oct 29, 2021 at 0:31