Let X‾\underline{X} be a random sample of size nn from the population PX; θ\mathbb{P}_{X;\,\theta} and T=t(X‾)T=t\left(\underline{X}\right) be a statistic. The statistic TT is a sufficient statistic for θ\theta if and only if the conditional distribution of the random sample given TT does not depend on θ.\theta.

P(X‾∣T,θ)=P(X‾∣T)\mathbb{P}\left(\underline{X}\mid T,\theta\right)=\mathbb{P}\left(\underline{X}\mid T\right)

A sufficient statistic contains all of the information in the random sample about θ.\theta. A statistic cannot contain more information about θ\theta than the random sample that it summarises.

Let X‾\underline{X} be a random sample of size nn from a continuous population PX; θ\mathbb{P}_{X;\,\theta} with joint pdf fX‾(x‾∣θ)f_{\underline{X}}\left(\underline{x}\mid\theta\right) and T=t(X‾)T=t\left(\underline{X}\right) be a statistic with pdf fT(t∣θ).f_T\left(t\mid\theta\right). The statistic TT is a sufficient statistic for θ\theta if and only if the ratio of the joint pdf of the random sample to the pdf of the statistic does not depend on θ\theta for all x‾∈Sn.\underline{x}\in S^n.

fX‾(x‾∣θ)fT(t(x‾)∣θ)\frac{f_{\underline{X}}\left(\underline{x}\mid\theta\right)}{f_T\left(t\left(\underline{x}\right)\mid\theta\right)}

Let X‾\underline{X} be a random sample of size nn from a continuous population PX; θ\mathbb{P}_{X;\,\theta} with joint pdf fX‾(x‾∣θ)f_{\underline{X}}\left(\underline{x}\mid\theta\right) and T=t(X‾)T=t\left(\underline{X}\right) be a statistic. The factorisation theorem states that TT is a sufficient statistic for θ\theta if and only if there exist functions u(t∣θ)u\left(t\mid\theta\right) and v(x‾)v\left(\underline{x}\right) such that the joint pdf of the random sample factorises into a function of the statistic and the parameter and a function of the data alone.

fX‾(x‾∣θ)=u(t(x‾)∣θ)v(x‾)  ∀ x‾∈Sn, θ∈Θf_{\underline{X}}\left(\underline{x}\mid\theta\right)=u\left(t\left(\underline{x}\right)\mid\theta\right)v\left(\underline{x}\right)\,\,\forall\,\underline{x}\in S^n,\,\theta\in\Theta

The ratio test and the factorisation theorem also hold for discrete populations with the corresponding pmfs in place of pdfs.