Higher floating point precision than double

[DanielDoehring]

Higher floating point precision than double

[adomahidi]

That's an interesting direction... From an ECOS point of view there shouldn't be a fundamental issue with higher precision types, but I am not sure about the linear solver under the hood (SparseLDL). All code is there so you might give it a try and see where it fails? :)

Just curious, from which field is the optimization problem requiring more than double precision arithmetics?

[DanielDoehring]

Alright, I will give it a shot :)

Essentially, I am trying to optimize the coefficients $ \boldsymbol \alpha \in \mathbb R^{N+1} $ of a complex polynom
$$ \sum_{i=0}^N \alpha_i z^i = \sum_{i=0}^N \frac{\widetilde \alpha_i}{i!} z^i , \quad z \in \mathbb C $$
where $ \vert z^i \vert \sim \mathcal O (10)$. Already for $N \approx 20 $ I run into stability issues since the $z^i$ term explodes. This can be somewhat mitigated by the normalization by $i!$. However, then you have to divide the $ \tilde \alpha_i$ by $i!$ which gives you in any case $\alpha_i$ very small to zero. Of course, this problem is inherently ill-conditioned, but I want to see how far I can go with this formulation.

[DanielDoehring]

Switching from doubleto long double is actually not that difficult, it can be done with moderate changes. However, some adjustments have to be made, i.e., fabs becomes fabsl and printing requires usually also an "L", e.g. %f becomes %fL.

Thus, a fully type-generic version would demand some more effort.

[adomahidi]

Ok great to hear, so did it run with long double after those changes and did you see an improvement on numerical stability?

[DanielDoehring]

I called ECOS through the Julia interface where there is, unfortunately, within the Julia MOI no support for long double or other types (like Double64) but only for the standard C double.

I will now take a look at the Matlab & Python interfaces.

[DanielDoehring]

Based on the EiCOS project which allows fairly easy specification of constraints I forked a long double version.