Canonical: https://knowlpedia.com/large-deviations/contraction-principle-ldp/

# Contraction principle

How a large deviation principle transfers through a continuous mapping.

**Contraction principle:** Let \((\mu_n)\) be [probability measures](https://knowlpedia.com/probability/probability-measure/) on a space \(X\) that satisfy a [large deviation principle](https://knowlpedia.com/large-deviations/large-deviation-principle/) with speed \(a_n\to\infty\) and [rate function](https://knowlpedia.com/large-deviations/rate-function/) \(I\colon X\to[0,\infty]\). Let \(f\colon X\to Y\) be continuous, and let \(\nu_n=\mu_n\circ f^{-1}\) be the [pushforward measures](https://knowlpedia.com/measure-theory/pushforward-measure/) on \(Y\). Then \((\nu_n)\) satisfies an LDP on \(Y\) with the same speed \(a_n\) and rate function
\[
J(y)=\inf\bigl\{ I(x)\,:\, x\in X,\ f(x)=y\bigr\},
\]
with the convention \(\inf\varnothing=+\infty\).

In terms of [random variables](https://knowlpedia.com/probability/random-variable/), if \(Z_n\) satisfies an LDP on \(X\) and \(Y_n=f(Z_n)\) with \(f\) continuous, then \((Y_n)\) satisfies an LDP with rate obtained by minimizing \(I\) over the fiber \(\{x:f(x)=y\}\). This principle is routinely combined with [Sanov's theorem](https://knowlpedia.com/large-deviations/sanovs-theorem/) and [Cramér's theorem](https://knowlpedia.com/large-deviations/cramers-theorem/) to derive LDPs for many statistics.
