Let us partition the [[Sample Space]] into finite number of [[Events]], e.g. $\{A_1, A_2, A_3\}$.
The total probability theorem (”TPT”) states that we can obtain an unconditional probability of $\mathbf P(B)$, by summing the [[Conditional Probability]] of each partition $\mathbf P(B \vert A_i)$ and weighting them by size size of the respective partition $P(A_i)$. ^83d664
$
\begin{aligned}
\mathbf P(B)
&= \mathbf P(B\vert A_1)*\mathbf P(A_1) \\
&+ \mathbf P(B\vert A_2)*\mathbf P(A_2) \\
&+ \mathbf P(B\vert A_3)*\mathbf P(A_3)
\end{aligned}
$
Obviously this extends to any $n$ partitions.
$ \mathbf P(B) = \sum_{i=1}^n \mathbf P(B\vert A_i)*\mathbf P(A_i) $
We can derive TPT by adding disjoint parts of $B$ from each partition (e.g. $B \cap A_1$). Each of these pieces can be deconstructed via the [[Probability Multiplication Rule]] into $\mathbf P(B \vert A_1)*\mathbf P(A_1)$.
![[total-probability-theorem.png]]
## Extension to Discrete Random Variables
To prove that this theorem has to work also for discrete [[Random Variable|random variables]] ("r.v."), we simply say that event $B$ is when r.v. $X$ equals some realization $x$. This lets us translate the probability of an event $\mathbf P (B)$ into a [[Probability Mass Function#^ac7640|PMF]].
$
\begin{align}
\mathbf P(B) &:\mathbf P(X=x) =p_X(x) \\
\mathbf P(B\vert A_i) &:\mathbf P(X=x \vert A_i) =p_{X\vert A_i}(x)
\end{align}
$
We perform the substitution and have established the TPT for PMF’s.
$ p_X(x) = \sum_{i=1}^n p_{X \vert A_i}(x)*\mathbf P(A_i) $ ^881fd7
## Extension to Continuous Random Variables
For continuous r.v’s. we say that event $B$ is when $\{X \le x\}$, which lets us translate the probability of event $\mathbf P(B)$ into a [[Cumulative Density Function]].
$
\begin{align}
\mathbf P(B)&:\mathbf P(X \le x) =F_X(x) \\
\mathbf P(B\vert A_i)&:\mathbf P(X \le x \vert A_i) =F_{X\vert A_i}(x)
\end{align}
$
We perform the substitution and have established the TPT for CDF’s. By taking the derivative on both sides we transform CDF’s to PDF’s.
$
\begin{align}
F_X(x)&=F_{X\vert A_1}(x)_\mathbf P(A_1)+ \dots +F_{X\vert A_n}(x)_\mathbf P(A_n) \\
f_X(x)&=f_{X\vert A_1}(x)_\mathbf P(A_1)+ \dots +f_{X\vert A_n}(x)_\mathbf P(A_n) \\
\end{align}
$