This PR adds SmoothedLaplace, an approximation of Laplace, and its pdf is evaluated as
$$p(x) = \frac{1}{2b} \exp\left(-\frac{\sqrt{(x-\mu)^2 + \beta}}{b}\right),$$
where $\beta$ is a smoothing parameter.
Compared with Laplace, SmoothedLaplace comes with a gradient method and it enables the use of gradient-based samplers.
Smoothed Laplace was implemented earlier as a custom distribution in showcase notebooks on smoothed Laplace and super resolution.
fixed #451
This PR adds SmoothedLaplace, an approximation of Laplace, and its pdf is evaluated as $$p(x) = \frac{1}{2b} \exp\left(-\frac{\sqrt{(x-\mu)^2 + \beta}}{b}\right),$$ where $\beta$ is a smoothing parameter.
Compared with Laplace, SmoothedLaplace comes with a
gradient
method and it enables the use of gradient-based samplers.Smoothed Laplace was implemented earlier as a custom distribution in showcase notebooks on smoothed Laplace and super resolution.