Skip to content

Instantly share code, notes, and snippets.

View thoma5B's full-sized avatar

Thomas Bunke thoma5B

  • Bad Homburg, Germany
View GitHub Profile
View file2
We are interested in a probability distribution $\rho:X \rightarrow \mathbb R_+$, such that
1. (Constraints) $ d_i^{-1} (\rho d\mathcal L ) = \mathbb P_i$, i.e. $\forall B \subseteq \mathbb R $ we have $\int 1_{\{d_i(x) \in B \} }\rho(x) d \mathcal L(x) = \mathbb P_i(B)$.
2. (Optimization problem) $\int \rho \log \rho d\mathcal L$ extremal - this expression maximizes entropy under the constraints. This takes account to the fact that less is known about the actual distribution of the probability of presence.
To assure positivity we set $\rho = e^\varphi$ and obtain $\rho \log \rho = \varphi e^\varphi$ for the entropy. For simplicity we assume for $\mathbb P_i$ only discrete probabilities $\mathbb P_i = \sum_k p_{i,k} \delta_{x_{i,k}}$
The Gâteaux differential $D^{Gâteaux}$ in direction $\psi$ is given by $\partial_h F(\varphi + h\psi)\mid_{h=0}$
View discussion
#A probability distribution (updated)
### suggested by Tilman J. Bohl
We are interested in a probability distribution $\rho:X \rightarrow \mathbb R_+$, such that
1. (Constraints) $ d_i^{-1} (\rho d\mathcal L ) = \mathbb P_i$, i.e. $\forall B \subseteq \mathbb R $ we have $\int 1_{\{d_i(x) \in B \} }\rho(x) d \mathcal L(x) = \mathbb P_i(B)$.
2. (Optimization problem) $\int \rho \log \rho d\mathcal L$ extremal - this expression maximizes entropy under the constraints. This takes account to the fact that less is known about the actual distribution of the probability of presence.
To assure positivity we set $\rho = e^\varphi$ and obtain $\rho \log \rho = \varphi e^\varphi$ for the entropy. For simplicity we assume for $\mathbb P_i$ only discrete probabilities $\mathbb P_i = \sum_k p_{i,k} \delta_{x_{i,k}}$