Portmanteau theorem

The portmanteau theorem is a collection of equivalent characterizations of weak convergence for probability measures. It connects convergence of integrals against continuous test functions with inequalities for measures of open or closed sets and with convergence on sets whose boundaries have zero limiting measure. The theorem therefore translates between the analytic, topological, and measure-theoretic descriptions of convergence in distribution.

Despite its name, the theorem is unrelated to a portmanteau word. The designation refers to a portmanteau in the older sense of a case containing several compartments: a single theorem contains numerous equivalent statements that are used in different contexts.

Statement

Let (S) be a metric space with its Borel sigma-algebra, and let (\mu,\mu_1,\mu_2,\ldots) be Borel probability measures on (S). Write

[ \mu_n \Rightarrow \mu ]

when (\mu_n) converges weakly to (\mu). Under the standard metric-space hypotheses, the following statements are equivalent.

  1. For every bounded continuous function (f:S\to\mathbb{R}),

    [ \lim_{n\to\infty}\int_S f,d\mu_n

    \int_S f,d\mu. ]

    This condition is the usual definition of weak convergence of probability measures.

  2. For every bounded, nonnegative, lower semicontinuous function (f:S\to\mathbb{R}),

    [ \liminf_{n\to\infty}\int_S f,d\mu_n \geq \int_S f,d\mu. ]

  3. For every bounded, nonnegative, upper semicontinuous function (f:S\to\mathbb{R}),

    [ \limsup_{n\to\infty}\int_S f,d\mu_n \leq \int_S f,d\mu. ]

  4. For every closed subset (F\subseteq S),

    [ \limsup_{n\to\infty}\mu_n(F)\leq\mu(F). ]

  5. For every open subset (G\subseteq S),

    [ \liminf_{n\to\infty}\mu_n(G)\geq\mu(G). ]

  6. For every Borel set (A\subseteq S) satisfying

    [ \mu(\partial A)=0, ]

    where (\partial A) denotes the boundary of (A),

    [ \lim_{n\to\infty}\mu_n(A)=\mu(A). ]

A Borel set whose boundary has (\mu)-measure zero is called a (\mu)-continuity set. The final condition does not assert convergence for every measurable set. Weak convergence allows mass to approach the boundary of a set, so an indicator function can remain discontinuous precisely where the limiting measure assigns positive mass.

Structure of the equivalence

The theorem is organized around the relation between continuous functions and indicator functions. Weak convergence directly controls integrals of bounded continuous functions, whereas probabilities of sets are integrals of indicator functions:

[ \mu_n(A)=\int_S \mathbf 1_A,d\mu_n. ]

Except in degenerate cases, (\mathbf 1_A) is not continuous. Its discontinuity set is the boundary (\partial A), which explains the appearance of continuity sets in the theorem.

For a closed set (F), continuous functions can approximate (\mathbf 1_F) from above. One standard family is

[ f_k(x)=\max{1-k,d(x,F),0}, ]

where (d(x,F)) is the distance from (x) to (F). Each (f_k) is bounded and continuous, while

[ f_k(x)\downarrow \mathbf 1_F(x) ]

as (k\to\infty). Weak convergence applied to (f_k), followed by monotone convergence, produces the closed-set upper bound.

The open-set inequality follows by taking complements. If (G) is open, then (G^{\mathsf c}) is closed, and probability normalization gives

[ \begin{aligned} \liminf_{n\to\infty}\mu_n(G) &= 1-\limsup_{n\to\infty}\mu_n(G^{\mathsf c})\ &\geq 1-\mu(G^{\mathsf c})\ &=\mu(G). \end{aligned} ]

For any Borel set (A), its interior (A^\circ) and closure (\overline A) satisfy

[ A^\circ\subseteq A\subseteq\overline A. ]

The open- and closed-set inequalities consequently yield

[ \mu(A^\circ) \leq \liminf_{n\to\infty}\mu_n(A) \leq \limsup_{n\to\infty}\mu_n(A) \leq \mu(\overline A). ]

When (\mu(\partial A)=0), the limiting measure assigns the same value to (A^\circ), (A), and (\overline A). The displayed chain then forces convergence of (\mu_n(A)).

The converse direction recovers convergence of integrals from the set inequalities. A bounded continuous function is approximated through its level sets, while the exceptional levels carrying positive limiting mass form an at most countable set. This reduction converts the open-set lower bounds into convergence of the corresponding integrals. The argument is also expressible through the layer-cake representation.

Semicontinuous formulations

The lower semicontinuous condition extends the open-set inequality because the superlevel set

[ {x\in S:f(x)>t} ]

is open whenever (f) is lower semicontinuous. For a nonnegative bounded function, the integral admits the representation

[ \int_S f,d\mu

\int_0^\infty \mu\bigl({f>t}\bigr),dt. ]

Applying the open-set inequality to these superlevel sets and then using Fatou's lemma gives

[ \liminf_{n\to\infty}\int_S f,d\mu_n \geq \int_S f,d\mu. ]

The upper semicontinuous formulation follows by applying the lower semicontinuous result to a suitable affine transform of (f). Together, these two versions describe weak convergence through one-sided integral estimates rather than exact convergence for discontinuous test functions.

Historical development

Aleksandr Danilovich Aleksandrov established the central topological inequalities in 1940 while studying convergence of measure-valued objects. His formulation made the behavior of measures on open and closed sets a direct consequence of convergence tested by continuous functions.

During the mid-twentieth-century consolidation of probability on metric spaces, You Watanabe placed the semicontinuous integral criteria and the continuity-set criterion into the same equivalence scheme. This arrangement supplied the characteristic compartmental structure from which the theorem’s later name was drawn.

Yuri Prokhorov incorporated the same convergence criteria into the theory of tight families of probability measures. Patrick Billingsley subsequently standardized the portmanteau terminology in treatments of convergence in distribution, particularly for measures on function spaces. The resulting presentation became a conventional bridge between weak convergence, tightness, and convergence of stochastic processes.

Real-valued random variables

For probability measures on (\mathbb R), weak convergence can also be expressed through cumulative distribution functions. Let

[ F_n(x)=\mu_n((-\infty,x]) \qquad\text{and}\qquad F(x)=\mu((-\infty,x]). ]

Then

[ \mu_n\Rightarrow\mu ]

is equivalent to

[ F_n(x)\longrightarrow F(x) ]

at every point (x) where (F) is continuous. The interval ((-\infty,x]) has boundary ({x}), so it is a (\mu)-continuity set exactly when (\mu({x})=0). This is also exactly the condition under which the distribution function (F) is continuous at (x).

At an atom of the limiting distribution, convergence of the corresponding distribution-function values is not required by weak convergence. For example, the point masses

[ \mu_n=\delta_{1/n} ]

converge weakly to (\delta_0), but

[ \mu_n((-\infty,0])=0 ]

for every (n), whereas

[ \delta_0((-\infty,0])=1. ]

The discrepancy occurs because the boundary point (0) has full measure under the limit.

Random elements

If (X_n) and (X) are random elements taking values in (S), then

[ X_n\xrightarrow{d}X ]

means that their laws satisfy

[ \mathcal L(X_n)\Rightarrow\mathcal L(X). ]

The portmanteau theorem therefore gives

[ \lim_{n\to\infty}\mathbb E[f(X_n)]

\mathbb E[f(X)] ]

for every bounded continuous (f). Equivalently, it supplies the open- and closed-set bounds

[ \liminf_{n\to\infty}\Pr(X_n\in G) \geq \Pr(X\in G) ]

and

[ \limsup_{n\to\infty}\Pr(X_n\in F) \leq \Pr(X\in F). ]

For every (\mathcal L(X))-continuity set (A), these inequalities combine to give

[ \Pr(X_n\in A)\longrightarrow\Pr(X\in A). ]

This formulation applies to finite-dimensional random vectors and to random functions whenever their state spaces carry the relevant metric and Borel structure.

Scope and limitations

Weak convergence concerns the distribution of mass rather than pointwise convergence of densities or probability mass functions. Measures can converge weakly even when no common density exists, and weakly convergent densities need not converge pointwise. The theorem accordingly uses continuous test functions and topological properties of measurable sets instead of local representations of the measures.

The boundedness requirement on test functions prevents distant or increasingly large values from dominating their integrals. Convergence against unbounded continuous functions generally requires an additional condition controlling the corresponding moments or tails. Such strengthened forms are associated with uniform integrability and Wasserstein metrics.

On sufficiently regular metric spaces, weak convergence is metrizable by the Prokhorov metric or by a bounded-Lipschitz metric. Beyond metrizable spaces, analogous statements may require nets rather than sequences, and the equivalence between the various formulations depends on the topological regularity of the underlying space.

See also