neonforestmist/improved-dimension-bco-gradient-variation-repro-artifacts / source /tex /Appendices /technical.tex
| \section{Technical Lemmas} | |
| \label{app:technical} | |
| \begin{Lemma}[{Lemma 4.8 of \citet{pogodin2019first}}] | |
| \label{lem:sum_cvx} | |
| Let $a_1, a_2, \ldots, a_T$ be non-negative real numbers. Then | |
| $$ | |
| \sum_{t=1}^T \frac{a_t}{\sqrt{1+\sum_{s=1}^{t-1} a_s}} \leq 4 \sqrt{1+\sum_{t=1}^T a_t}+\max _{t \in[T]} a_t . | |
| $$ | |
| \end{Lemma} | |
| \begin{Lemma}[{Lemma 9 of \citet{yan2023universal}}] | |
| \label{lem:sum_scvx} | |
| For a sequence of $\{a_t\}_{t=1}^T$ and $b$, where $a_t, b > 0$ for any $t \in [T]$, denoting by $a_{\max} \define \max_t a_t$ and $A \define \ceil{b \sumT a_t}$, we have | |
| \begin{equation*} | |
| \sumT \frac{a_t}{bt} \le \frac{a_{\max}}{b} (1 + \log A) + \frac{1}{b^2}. | |
| \end{equation*} | |
| \end{Lemma} | |
| \begin{Lemma}[{Lemma 9 of \citet{zhao2024adaptivity}}] | |
| \label{lem:small-loss-sqrt} | |
| For any $x, y, a, b>0$ satisfying $x-y \le \sqrt{a x}+b$, it holds that | |
| \begin{equation*} | |
| x-y \le \sqrt{a y+ab}+a+b. | |
| \end{equation*} | |
| \end{Lemma} | |
| \begin{Lemma}[{Lemma 16 of \citet{orabona2012beyond}}] | |
| \label{lem:small-loss-log} | |
| If $a, b, c, x, y>0$ satisfy $x-y \le a \log (b x+c)$, then it holds that | |
| \begin{equation*} | |
| x-y \le a \log \sbr{2 a b \log \frac{2 a b}{e} + 2by + 2c}. | |
| \end{equation*} | |
| \end{Lemma} |
Xet Storage Details
- Size:
- 1.29 kB
- Xet hash:
- e8dbbc65f58c297e36ca5aa1530baf3a8678ef0d0c5cff23886a0a5ac725bc51
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.