First-Order Methods
Riemannian Gradient Descent (RGD)
Algorithm
\begin{algorithm}
\caption{Riemannian Gradient Descent}
\begin{algorithmic}
\ENSURE{$\tau,r \in (0,1)$}
\PROCEDURE{RGD}{$x_0 \in \mathcal{M}$}
\FOR{$k = 0,1,2\dots$}
\STATE{Pick a step size $\alpha_k > 0$}
\STATE{$x_{k+1} = R_{x_k}(s_k)$, with step $s_k = -\alpha_k \text{grad}f(x_k)$}
\ENDFOR
\RETURN{$x_{final}$}
\ENDPROCEDURE
\PROCEDURE{BacktrackingLineSearch}{$x \in \mathcal{M}, \hat{\alpha} > 0$}
\STATE{$\alpha \leftarrow \hat{\alpha}$}
\WHILE{$f(x) - f(R_x(-\alpha \text{grad}f(x))) < r\alpha\lVert\text{grad}f(x)\rVert^2$}
\STATE{$\alpha = \tau \alpha$}
\ENDWHILE
\RETURN{$\alpha$}
\ENDPROCEDURE
\end{algorithmic}
\end{algorithm}
Problem setting
- Lower Bound: \(\forall x\in \mathcal{M}, f(x) \ge f_{min}\)
- Good Step Picking Algorithm: \(f(x_k) - f(x_{k+1}) \ge c\lVert \text{grad}f(x_k) \rVert\)
- “Lipschitz Continuity”: For a given \(S \subset T\mathcal{M}, \exists L.\forall (x,s) \in S. f(R_x(s)) \le f(x) + \langle \text{grad}f(x), s\rangle + \frac12 \lVert s\rVert^2\)
Riemannian Stochastic Gradient Descent (RSGD)
Arxiv Link
Riemannian Accelerated Gradient Descent (RAGD)
Arxiv Link
Riemannian Langevin Diffusion (RLD)
Paper Link