Used data debiasing log Gaussian

Remarkably, the largest partial posterior size was only $max_{r} left{n_{T_r}right}=2048$, leading to a maximum single replication cost of $sum_{t=1}^{max_r {T_r}}n_t= 4088$, and a median $n_{T_r}$ of only $16$. After $R=300$ replications, the number of data touched in total is $sum_{r=1}^{R} sum_{t=1}^{T_r} n_t=27,264$. Taking into account the $600$ MCMC iterations to estimate each partial posterior expectation, this sums up to $16,358,400$ likelihood evaluations, which is less than a quarter of a single full MCMC burn-in iteration ($N=2^{26}$), and less than $1/(4cdot 500)approx 0.0005$ times the number of likelihood evaluations required to complete the burn-in of full MCMC.

Leave a Reply

Your email address will not be published. Required fields are marked *