Maximum likelihood estimation

One-line definition

For a parametric family $p(x\mid \theta )$ and observed data $\{x_1,\dots ,x_n\}$, the maximum likelihood estimate (MLE) is

$${\hat{\theta }}_{\text{MLE}}=\arg \underset{\theta }{\max }\prod _{i=1}^{n}p(x_i\mid \theta )=\arg \underset{\theta }{\max }\sum _{i=1}^n\log p(x_i\mid \theta ).$$

Why it matters

MLE underlies almost every modern ML loss function:

• Cross-entropy for classification = MLE under a categorical model.
• Mean-squared error = MLE under a Gaussian noise model.
• Negative log-likelihood for language models = MLE.

When you read “minimize the negative log-likelihood,” you’re reading MLE.

Properties

Under regularity conditions (smooth log-likelihood, identifiable model, true parameter in interior), MLE is:

• Consistent: $\hat{\theta }\to {\theta }^{\ast }$ as $n\to \infty$.
• Asymptotically normal: $\sqrt{n}(\hat{\theta }-{\theta }^{\ast })\to \mathcal{N}(0,I({\theta }^{\ast }{)}^{-1})$ where $I(\theta )$ is the Fisher information.
• Asymptotically efficient: achieves the Cramér–Rao lower bound. No consistent estimator has lower asymptotic variance.

Common cases

Model	MLE solution
Gaussian mean (known ${\sigma }^2$)	$\hat{\mu }=\bar{x}$
Gaussian variance	${\hat{\sigma }}^2=\frac{1}{n}\sum (x_i-\bar{x}{)}^2$ (biased; sample-variance uses $n-1$)
Bernoulli ($p$)	$\hat{p}=\text{fraction of successes}$
Categorical	empirical class frequencies
Linear regression with Gaussian noise	OLS: $\hat{\beta }=(X^{\top }X{)}^{-1}{X}^{\top }y$
Logistic regression	no closed form; iterative (Newton, gradient methods)

Connection to cross-entropy and KL

For categorical $p_{\theta }(x)$ and empirical distribution ${\hat{p}}_{\text{data}}$, the negative log-likelihood divided by $n$ is

$$-\frac{1}{n}\sum _i\log p_{\theta }(x_i)=H({\hat{p}}_{\text{data}},p_{\theta })=H({\hat{p}}_{\text{data}})+\mathrm{KL}({\hat{p}}_{\text{data}}\Vert p_{\theta }).$$

Maximizing likelihood = minimizing cross-entropy = minimizing KL from the empirical distribution to the model.

MLE vs MAP

MAP (maximum a posteriori) adds a prior: ${\hat{\theta }}_{\text{MAP}}=\arg {\max }_{\theta }\log p(\theta )+\sum \log p(x_i\mid \theta )$. MAP equals MLE when the prior is uniform (improper). Common choices:

• Gaussian prior $\Rightarrow$ L2 regularization on $\theta$.
• Laplace prior $\Rightarrow$ L1 regularization.

Common pitfalls

• Treating sample variance as MLE. MLE for Gaussian variance divides by $n$ (biased); the sample variance divides by $n-1$ (unbiased). Different estimators.
• Stopping at MLE without checking identifiability. If two parameter values yield identical likelihoods, MLE is non-unique.
• Trusting MLE on small samples. Asymptotic guarantees can be misleading when $n$ is small relative to dimension; use cross-validation or Bayesian methods.
• Forgetting that MLE on a misspecified model is still well-defined. It converges to the parameter that minimizes KL to the (mismatched) true distribution within the model family.