Canonical: https://knowlpedia.com/large-deviations/log-moment-generating-function/

# Log moment generating function

The logarithm of the moment generating function, viewed as a convex functional of the parameter.

A **log moment generating function** (log-MGF) of an \(\mathbb R^d\)-valued [random variable](https://knowlpedia.com/probability/random-variable/) \(X\) is the function \(\Lambda:\mathbb R^d\to(-\infty,\infty]\) defined by
\[
\Lambda(\theta)=\log \mathbb E\big[e^{\langle \theta, X\rangle}\big],
\]

where \(\langle\theta,X\rangle\) is the Euclidean [inner product](https://knowlpedia.com/linear-algebra/inner-product/) and the expectation is taken in the sense of [expectation](https://knowlpedia.com/probability/expectation/).

## Equivalent characterizations

Equivalently, if \(\mu\) is the [law](https://knowlpedia.com/probability/distribution-law/) of \(X\), then
\[
\Lambda(\theta)=\log\int_{\mathbb R^d} e^{\langle\theta,x\rangle}\,\mu(dx),
\]
where the integral is a special case of the [Lebesgue integral](https://knowlpedia.com/measure-theory/lebesgue-integral/).

## Remarks

The log-MGF is a central object in large deviations: it is convex and encodes exponential moment growth, and its convex dual gives the [Cramér transform](https://knowlpedia.com/large-deviations/cramer-transform/). In particular, for sums of an [i.i.d. sequence](https://knowlpedia.com/probability/iid-sequence/), the log-MGF is the starting point for [Cramér's theorem](https://knowlpedia.com/large-deviations/cramers-theorem/) and the [Gärtner–Ellis theorem](https://knowlpedia.com/large-deviations/gartner-ellis-theorem/).

## Examples

- If \(X\sim \mathcal N(0,\sigma^2)\) on \(\mathbb R\), then \(\Lambda(\theta)=\frac{\sigma^2\theta^2}{2}\) for all \(\theta\in\mathbb R\).
- If \(X\sim \mathrm{Bernoulli}(p)\) on \(\{0,1\}\), then \(\Lambda(\theta)=\log\big((1-p)+p e^{\theta}\big)\) for all \(\theta\in\mathbb R\).
