The Probability Integral Transform states that if we evaluate the inverse [[Cumulative Density Function|CDF]] of any continuous distribution $X$, denoted as $F^{-1}_X$, and evaluate it at points that are generated by a standard [[Continuous Uniform Distribution]] $Y \sim \mathcal U(0,1)$,we will get data points, following a distribution of $X$.
**Key Steps:**
1. *Forward transform:* Sampling from $X$ and applying its CDF $F_X$ transforms the data into a standard Uniform distribution $Y\sim \mathcal U(0,1)$.
$ Y \sim F_X(X) $
2. *Reverse transform:* Sampling from a Uniform distribution $\mathcal U (0,1)$ and applying the inverse CDF $F_X^{-1}$ generates samples that follow the original distribution $X$.
$ X \sim F_X^{-1}(Y) $
**Proof:**
To prove that the reverse transform $F^{-1}_X(Y)$ generates samples from $X$, consider the following:
$
\begin{align}
F_Y(y) &= \mathbf P(Y \le y) \tag{1}\\[2pt]
&= \mathbf P(F_X(X) \le y) \tag{2}\\[2pt]
&= \mathbf P(X \le F^{-1}_X(y)) \tag{3}\\[2pt]
&= F_X(F^{-1}_X(y)) \tag{4}\\[2pt]
&=y \tag{5}
\end{align}
$
where:
- (2) Substitute $F_X(X)$ for $Y$.
- (3) Use the monotonic property of $F_X$, where $F_X(X) \le y \iff X \le F_X^{-1}(y)$.
- (4) Recognize that this can be expressed as a CDF of $X$.
- Since $F_X^{-1}$ is the inverse of $F_X$, they offset each other, leaving $y$ as the remainder.
By showing that $F_Y(y)=y$ we have proven that it is a standard Uniform distribution.
![[probability-integral-transform.png|center|400]]