We assume the distribution of a [[Random Variable]] consists of two sub-populations, where each sub-population follows a separate [[Gaussian Distribution]]:
- $X_1 \sim \mathcal{N}(\mu_1, \sigma_1^2)$
- $X_2 \sim \mathcal{N}(\mu_2, \sigma_2^2)$
The mixing proportion $\pi^\star$ determines the contribution of each sub-population in the overall mixture:
- $\pi^\star \in [0,1]$: Proportion of first sub-population.
- $1-\pi^\star$: Proportion of second sub-population.
The [[Probability Density Function|PDF]] of the mixture is a linear combination of the individual Gaussian PDFs:
$
f(x)=\pi^\star* \frac{1}{\sigma_1\sqrt{2\pi}}* \exp\Big(-\frac{(x-\mu_1)^2}{2\sigma_1^2}\Big) +
(1-\pi^\star)* \frac{1}{\sigma_2\sqrt{2\pi}}* \exp\Big(-\frac{(x-\mu_2)^2}{2\sigma_2^2}\Big)
$
Using a compact notation for the individual Gaussian PDFs $f_1(x)$ and $f_2(x)$:
$ f(x) = \pi^\star *f_1(x) + (1- \pi^\star)*f_2(x) $
## Validity of the Mixture
To ensure $f(x)$ is a valid PDF, it must satisfy the conditions from the [[Probability Axioms]]:
- *Non-negativity:* All densities need to be greater equal to zero → $f(x)\ge0$
- *Normalization:* The integral over the entire domain must equal one → $f(x)=1$.
**Proof of Non-negativity:**
Since both $f_1(x), f_2(x)$ satisfy the condition of positive density separately, the mixture (a linear combination with positive factors) has to be satisfied as well.
$ \begin{rcases} f_1(x)\ge0\\ f_2(x)\ge0 \end{rcases} f(x)\ge0 $
**Proof of Normalization:**
Since both densities $f_1(x), f_2(x)$ integrate to $1$, the multiplication with $\pi^\star$ and $(1-\pi^\star)$ has to equal $1$ as well.
$
\begin{aligned}
\int f(x) \,dx&=
\pi^\star*
\overbrace{\int f_1(x) \,dx}^{=1} + (1-\pi^\star)*
\overbrace{\int f_2(x) \,dx}^{=1} \\
&=\pi^\star_1 + (1-\pi^\star)*1 \\[6pt] &=1
\end{aligned}
$
## Random Weights Representation
Instead of fixed weights $\{\pi^\star, 1-\pi^\star \}$, we can also model the weights of a mixture as a [[Bernoulli Distribution|Bernoulli]] r.v. $Z \sim \text{Ber}(\pi)$.
$ X=ZX_1+ (1-Z)X_2 $
So if $Z$ returns $1$, $X$ will be sampled only from $X_1 \sim \mathcal{N}(\mu_1, \sigma_1^2)$ and otherwise only from $X_2 \sim \mathcal{N}(\mu_2, \sigma_2^2)$.
$ Z = \begin{cases} 1 & \text{w.p. } \pi &&\text{then } X\sim N(\mu_1, \sigma_1^2)
\\ 0 & \text{w.p. } (1-\pi) &&\text{then } X\sim N(\mu_2, \sigma_2^2) \end{cases} $
For this representation to give the same result as with a fixed $\pi$, both $X_1, X_2$ have to be independent of $Z$. This means e.g. $f(X_1\vert Z=1)=f(X_1)$.