Laplace algorithm computes Hessian at the mode, but it is not available for the user. The Hessian would be useful for
- computing the normalization term of the normal approximation, which would help to estimate marginal likelihood (we have a use case)
- computing normal approximation based sd and SE without drawing from the normal distribution, which would be useful for very big data and close to normal posterior
- for algorithm development for algorithms that start with mode and Hessian, like CCD