diff --git a/tex/slides.tex b/tex/slides.tex index d07c51d1e6d79acdbcd0511a08371fab109a10d2..ee73817f2c189388e30ac9fb66b6af8698347f8b 100644 --- a/tex/slides.tex +++ b/tex/slides.tex @@ -125,7 +125,7 @@ \medskip - \item If we obtain $\{x_i\}_{i=1}^N$ from another distribution $q(x)$: + \item If we obtain $\{x_i\}_{i=1}^N$ from another distribution with density $q(x)$: \[ \mathbb E_p\left[ \varphi(X) \right] = \int q(x)\frac{p(x)}{q(x)}\varphi(x)\mathrm d x = \int q(x)\tilde{\omega}(x)\varphi(x)\mathrm d x = \mathbb E_q\left[ \tilde{\omega}(X)\varphi(X) \right] \] @@ -147,7 +147,7 @@ \begin{itemize} - \item IS works as long as the support of the targetted distribution ($p$) is included in the support of the instrumental distribution ($q$).\newline + \item IS works as long as the support of the targeted distribution ($p$) is included in the support of the instrumental distribution ($q$).\newline \item How do we choose distribution $q(x)$?\newline @@ -161,7 +161,7 @@ % ⇒ We are better off if the unormalized weight variance is low, i.e. if f/g is approximately constant. % The closer g is to f, the better we are. - \item Is it possible to find a reasonable intrumental distribution for a posterior distribution?\newline + \item Is it possible to find a reasonable intrumental distribution for a posterior distribution? The prior?\newline \item Probably better not to ``jump'' directly to the posterior distribution... @@ -184,14 +184,14 @@ \item[Target] $p\left(\theta|\mathcal Y_T\right) \propto p(\theta) p(\mathcal Y_T| \theta)$\newline - \item Consider the ''simplified`` object $p\left(\theta\right)p\left(\mathcal Y_T\right)^{\phi}$\newline + \item Consider the ''simplified`` object $p\left(\theta\right)p\left(\mathcal Y_T| \theta\right)^{\phi}$\newline - \item $p\left(\theta\right)$ is a good instrumental distribution for $p\left(\theta\right)p\left(\mathcal Y_T\right)^{\phi}$ for small values of $\phi$.\newline + \item $p\left(\theta\right)$ is a good instrumental distribution for $p\left(\theta\right)p\left(\mathcal Y_T| \theta\right)^{\phi}$ for small values of $\phi$.\newline \item Consider a sequence of $\{\phi_i\}_{i=1}^M$ with $\phi_i<\phi_j$ for all $j>i$ and $\phi_M = 1$.\newline - \item Use $p\left(\theta\right)$ as an intrumental distribution for $p\left(\theta\right)p\left(\mathcal Y_T\right)^{\phi_1}$ and $p\left(\theta\right)p\left(\mathcal Y_T\right)^{\phi_i}$ - as an instrumental distribution for $p\left(\theta\right)p\left(\mathcal Y_T\right)^{\phi_{i+1}}$.\newline + \item Use $p\left(\theta\right)$ as an intrumental distribution for $p\left(\theta\right)p\left(\mathcal Y_T| \theta\right)^{\phi_1}$ and $p\left(\theta\right)p\left(\mathcal Y_T| \theta\right)^{\phi_i}$ + as an instrumental distribution for $p\left(\theta\right)p\left(\mathcal Y_T| \theta\right)^{\phi_{i+1}}$.\newline \item Herbst and Schofheide consider $\phi_n = \left( \frac{n}{M} \right)^\lambda$ with $\lambda>1$ (convexity). @@ -296,9 +296,9 @@ posterior_sampling_options = ('particles', 20000, \begin{itemize} - \item More options.\newline + \item More options (number of mutation MH steps, resampling algorithm, ...).\newline - \item Complete posterior computations.\newline + \item Complete posterior computations (bayesian IRFs, forectasts, ...).\newline \item No reason to start from the prior distribution.\newline