Say we have a univariate Gaussian distribution $p(x)=\mathcal N(x|\mu, \sigma^{2})$. Then suppose we have a data set of observations $X = \{x_1, x_2, \cdots ,x_N\}^{T}$ that are drawn independently from this distribution.

By performing MLE, which means by maximizing $$\ln\prod_{n = 1}^{N} \mathcal N(x_n|\mu, \sigma^{2})$$ with respect to $\mu$ and $\sigma$ respectively, I can get the maximum likelihood solution for $\sigma$, which is $$\sigma_{ML} = \frac{1}{N}\sum_{n = 1}^{N}(x_n-\mu_{ML})^{2}$$

Clearly, in front of the summation, we have a $\frac{1}{N}$. However, I also read that the maximum likelihood solution for $\sigma$ is $$\sigma_{ML} = \frac{1}{N-1}\sum_{n = 1}^{N}(x_n-\mu_{ML})^{2}$$, in which $N$ still denotes for the size of the dataset.

Why are the two solutions different? How do we have $N-1$ on the denominator?


1 Answer 1


While $\frac{1}{N}\sum_{n = 1}^{N}(x_n-\bar x)^{2}$ is the maximum likelihood estimator, it is biased and using $\frac{1}{N-1}$ instead corrects this bias. So the first equation (a.k.a. population variance) is the maximum likelihood estimator, while the second one (a.k.a. sample variance) is the bias-corrected one.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.