= Normal mean and variance maximum-likelihood estimators
For an <IID sample> from $N(\mu,\sigma^2)$ with $\sigma^2>0$, a nonconstant observed sample has <maximum-likelihood estimators>
$$
\widehat\mu=\overline X,\qquad\widehat\sigma^2=\frac1n\sum_i(X_i-\overline X)^2.
$$
Completing the square first minimizes the residual sum at $\overline X$; differentiating the profiled <log-likelihood> in $\sigma^2$ gives the second formula. It uses denominator $n$, whereas the unbiased sample <variance> uses $n-1$: its expectation is $(n-1)\sigma^2/n$. A constant sample, including any one-observation sample, has an unbounded <likelihood> as the <variance> decreases to zero, so the displayed zero <variance> is only a boundary extension in that case.
Back to article page