Let X‾\underline{X} be a random sample of size nn from the population PX; θ\mathbb{P}_{X;\,\theta} with state space S,S, and gL,gU:Sn→Rg_L,g_U:S^n\rightarrow\mathbb{R} be statistics with gL(x‾)≤gU(x‾)g_L\left(\underline{x}\right)\leq g_U\left(\underline{x}\right) for all x‾∈Sn.\underline{x}\in S^n. An interval estimator of θ\theta is the random interval [gL(X‾),gU(X‾)],\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right], which is intended to contain the true parameter θ.\theta.

Let [gL(X‾),gU(X‾)]\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right] be an interval estimator of θ\theta and x‾\underline{x} be an observed random sample. An interval estimate of θ\theta is the realised interval [gL(x‾),gU(x‾)],\left[g_L\left(\underline{x}\right),g_U\left(\underline{x}\right)\right], which is equivalent to the statement gL(x‾)≤θ≤gU(x‾).g_L\left(\underline{x}\right)\leq\theta\leq g_U\left(\underline{x}\right).

Let gLg_L and gUg_U be statistics. A one-sided interval estimate of θ\theta is an interval estimate of the form (−∞,gU(x‾)]\left(-\infty,g_U\left(\underline{x}\right)\right] or [gL(x‾),∞),\left[g_L\left(\underline{x}\right),\infty\right), which is equivalent to the statement θ≤gU(x‾)\theta\leq g_U\left(\underline{x}\right) or θ≥gL(x‾)\theta\geq g_L\left(\underline{x}\right) respectively.

Let [gL(X‾),gU(X‾)]\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right] be an interval estimator of θ.\theta. The coverage probability of the interval estimator is the probability that θ\theta lies within the random interval, which is the joint probability that the lower statistic is at most θ\theta and the upper statistic is at least θ.\theta.

P(θ∈[gL(X‾),gU(X‾)])=P(gL(X‾)≤θ,gU(X‾)≥θ)\begin{align*} \mathbb{P}\left(\theta\in\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right]\right)=\mathbb{P}\left(g_L\left(\underline{X}\right)\leq\theta,g_U\left(\underline{X}\right)\geq\theta\right) \end{align*}

The coverage probability is a function of θ.\theta.

Let [gL(X‾),gU(X‾)]\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right] be an interval estimator of θ.\theta. The confidence coefficient of the interval estimator is the infimum of the coverage probability over all possible values of the true parameter.

inf⁡θ∈ΘP(θ∈[gL(X‾),gU(X‾)])\inf_{\theta\in\Theta}\mathbb{P}\left(\theta\in\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right]\right)

The coverage probability is often constant as a function of θ,\theta, in which case the coverage probability and the confidence coefficient are equal.

Let [gL(X‾),gU(X‾)]\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right] be an interval estimator of θ\theta and α∈(0,1).\alpha\in\left(0,1\right). The interval estimator is a 100(1−α)%100\left(1-\alpha\right)\% confidence interval for θ\theta if and only if its confidence coefficient is at least 1−α.1-\alpha.

P(θ∈[gL(X‾),gU(X‾)])≥1−α  ∀ θ∈Θ\mathbb{P}\left(\theta\in\left[g_L\left(\underline{X}\right),g_U\left(\underline{X}\right)\right]\right)\geq 1-\alpha\,\,\forall\,\theta\in\Theta

A confidence interval has probability at least 1−α1-\alpha of containing θ\theta over repeated samples before the data are observed. A realised confidence interval either contains θ\theta or does not, and the confidence coefficient is not a probability statement about θ\theta given the observed data.