Lévy's continuity theorem (source code)

= Lévy's continuity theorem
{c}
{wiki}

= Characteristic-function convergence theorem
{synonym}

Let $\mu_n$ be <probability measure>[probability measures] on the <real line> with <characteristic function>[characteristic functions] $\phi_n(t)=\int e^{itx}\,\mu_n(dx)$. <Lévy's continuity theorem> says that if $\phi_n(t)$ converges for every real $t$ to a <function> $\phi$ continuous at zero, then $\phi$ is the <characteristic function> of a unique <probability measure> $\mu$ and $\mu_n$ converges <weak convergence of probability measures>[weakly] to $\mu$. Conversely, <weak convergence of probability measures>[weak convergence] implies <pointwise convergence> of <characteristic function>[characteristic functions] to the <characteristic function> of the limit.

The converse follows immediately by testing against the bounded <continuous function>[continuous functions] $e^{itx}$. The difficult direction uses continuity at zero to control <probability> escaping to infinity, and the uniqueness of <characteristic function>[characteristic functions] to identify every <subsequence>[subsequential limit]. The theorem turns convergence of <probability law>[probability laws] into convergence of ordinary <functions>, which is especially useful when <independence> makes <characteristic function>[characteristic functions] multiply.

Continuity at zero is essential. For $X_n\sim N(0,n^2)$, $\phi_n(t)=e^{-n^2t^2/2}$ tends to zero for every nonzero $t$ but remains one at $t=0$. This discontinuous limit is not a <characteristic function>, and the distributions have no weak <probability> limit: their mass in every fixed bounded <interval> tends to zero.