Problem description
Given data at many time instance $t$,
$$\min _{\alpha, \Lambda, \beta} \lVert y(t) - \alpha e^{\Lambda t} \beta \rVert_F$$ with $$ \lVert \alpha \rVert_2^F = 1 $$
where $y(t) \in \mathbb{R}^{n \times M}$, $\Lambda \in \mathbb{R}^{r\times r}$ is a diagonal matrix, $\alpha \in \mathbb{R}^{n \times r}$, $\beta \in \mathbb{R}^{r \times M}$
Special case
if $\alpha = I$, the problem becomes a classical problem in variable projection for nonlinear least square, which has been studied for decades.
The benefit of $\alpha = I$ is that, the nonlinear and linear part becomes separable through pseudo-inverse.
What I currently do
I can simply (blindly) use an optimization package to generally solve this problem by do naive gradient descent, it looks like a neural network training process, without leveraging the structure of the problem.
It works for toy problem, like when n,r less than 3.
The downside is that, since it doesn't assume any structure to leverage, the computational time is costly.
What I want to know
I want to know if any one know which category does this problem belongs to, and if there is better algorithm existing to solve this.
