Difference between revisions of "Convergence of measures"
(moved newcommand up from start of para (it adds unwanted whitespace)) |
Ulf Rehmann (talk | contribs) m (fixed: \text{} problem) |
||
(9 intermediate revisions by one other user not shown) | |||
Line 6: | Line 6: | ||
$\newcommand{\abs}[1]{\left|#1\right|}$ | $\newcommand{\abs}[1]{\left|#1\right|}$ | ||
− | A concept in measure theory, determined by a certain topology in a space of measures that are defined on a certain | + | A concept in measure theory, determined by a certain topology in a space of measures that are defined on a certain [[Algebra of sets|σ-algebra]] $\mathcal{B}$ of subsets of a space $X$ or, more generally, in a space $\mathcal{M} (X, \mathcal{B})$ of charges, i.e. countably-additive real (resp. complex) functions $\mu: \mathcal{B}\to \mathbb R$ (resp. $\mathbb C$), often also called $\mathbb R$ (resp. $\mathbb C$) valued or [[Signed measure|signed measures]]. The total variation measure of a $\mathbb C$-valued measure is defined on $\mathcal{B}$ as: |
\[ | \[ | ||
− | \abs{\mu}(B) :=\sup\left\{ \sum \abs{\mu(B_i)}: | + | \abs{\mu}(B) :=\sup\left\{ \sum \abs{\mu(B_i)}: \{B_i\}\subset\mathcal{B} \text{ is a countable partition of } B \right\}. |
\] | \] | ||
In the real-valued case the above definition simplifies as | In the real-valued case the above definition simplifies as | ||
\[ | \[ | ||
− | \abs{\mu}(B) = \sup_{A\in \mathcal{B}, A\subset B} \left(\abs{\mu (A)} + \abs{\mu ( | + | \abs{\mu}(B) = \sup_{A\in \mathcal{B}, A\subset B} \left(\abs{\mu (A)} + \abs{\mu (B\setminus A)}\right). |
\] | \] | ||
The total variation of $\mu$ is then defined as $\left\|\mu\right\|_v := | The total variation of $\mu$ is then defined as $\left\|\mu\right\|_v := | ||
\abs{\mu}(X)$. | \abs{\mu}(X)$. | ||
+ | |||
+ | ''Warning'': If $\mathcal{B}$ is the $\sigma$-algebra of Borel sets of a topological space $X$, we will then denote by $\mathcal{M}^b (X)$ the space of ''Radon'' signed measures, i.e. those signed measures with finite total variation such that $|\mu|$ is a [[Radon measure]]. This is actually not a restriction in many cases, for instance if $X$ is the euclidean space. | ||
+ | |||
+ | |||
+ | ==Notions of convergence== | ||
The space $\mathcal{M}^b (X, \mathcal{B})$ of $\mathbb R$ (resp. $\mathbb C$) valued measure with finite total variation is a [[Banach space]] and the following are the most commonly used topologies. | The space $\mathcal{M}^b (X, \mathcal{B})$ of $\mathbb R$ (resp. $\mathbb C$) valued measure with finite total variation is a [[Banach space]] and the following are the most commonly used topologies. | ||
− | + | ===The norm or [[strong topology]]=== | |
− | if $\left\|\mu_n-\mu\right\|_v\to 0$. | + | $\mu_n\to \mu$ if and only if $\left\|\mu_n-\mu\right\|_v\to 0$. This convergence is sometimes called ''convergence in variation''. |
− | + | ===The [[weak topology]]=== | |
+ | A sequence of measures $\mu_n \rightharpoonup \mu$ if and only if $F (\mu_n)\to F(\mu)$ | ||
for every bounded linear functional $F$ on $\mathcal{M}^b$. | for every bounded linear functional $F$ on $\mathcal{M}^b$. | ||
− | + | ===The narrow topology=== | |
+ | When $X$ is a [[topological space]] and $\mathcal{B}$ the corresponding $\sigma$-algebra of [[Borel set|Borel sets]], we can introduce on $X$ the narrow topology. In this case $\mu_n$ converges to $\mu$ if and only if | ||
\begin{equation}\label{e:narrow} | \begin{equation}\label{e:narrow} | ||
\int f\, \mathrm{d}\mu_n \to \int f\, \mathrm{d}\mu | \int f\, \mathrm{d}\mu_n \to \int f\, \mathrm{d}\mu | ||
\end{equation} | \end{equation} | ||
− | for every bounded continuous function $f:X\to \mathbb R$ (resp. $\mathbb C$) | + | for every bounded continuous function $f:X\to \mathbb R$ (resp. $\mathbb C$). The following is an important consequence of the narrow convergence when $X$ is a [[Locally compact space|locally compact]] [[Hausdorff space]]: if $\mu_n$ converges narrowly to $\mu$, then $\mu_n (A)\to \mu (A)$ for any Borel set such that $\abs{\mu}(\partial A) = 0$ (cp. with Theorem 1(iii) of Section 1.9 in {{Cite|EG}}). |
− | |||
− | + | ===The wide or weak$^\star$ topology=== | |
+ | When $X$ is a locally compact topological space and $\mathcal{B}$ the $\sigma$-algebra of Borel sets yet another topology can be introduced, the so-called wide topology, or sometimes referred to as [[weak-star topology|weak$^\star$ topology]]. A sequence $\mu_n\rightharpoonup^\star \mu$ if and only if \eqref{e:narrow} holds for continuous functions which are compactly supported. | ||
+ | This topology is in general weaker than the narrow topology but they do coincide when restricted to probability measures if $X$ is a Hausdorff space. | ||
− | + | A topology analogous to the weak$^\star$ topology is defined in the more general space $\mathcal{M}^b_{loc}$ of locally bounded measures, i.e. those measures $\mu$ such that for any point $x\in X$ there is a neighborhood $U$ with $\abs{\mu}(U)<\infty$. | |
− | + | '''Warning''' Sequences of measures converging in the narrow (or in the wide topology) are called ''weakly convergent'' sequences by several authors (cp. with {{Cite|Bi}}, {{Cite|Ma}} and {{Cite|EG}}). This is, however, inconsistent with the terminology of Banach spaces, see below. | |
− | ==== | + | ==Properties== |
+ | ===Relation with functional analysis=== | ||
+ | If $X$ is compact and Hausdorff the [[Riesz representation theorem]] shows that $\mathcal{M}^b (X)$ is the dual of the space $C(X)$ of continuous functions. Under this assumption the narrow and weak$^\star$ topology coincides with the usual [[weak-star topology|weak$^\star$ topology]] of the Banach space theory. Since in general $C(X)$ is not a [[Reflexive space|reflexive space]], it turns out that the narrow topology is in general weaker than the weak topology. | ||
+ | |||
+ | ===Metrizability of the weak$^*$ topology=== | ||
+ | On bounded subsets of $\mathcal{M}^b (X)$, the weak$^*$ topology is metrizable. If $X$ is compact, this follows directly from standard functional-analytic arguments, since $\mathcal{M}^b (X)$ is then the dual of a separable Banach space. The case of a $\sigma$-compact $X$ can be reduced to that of a compact space by exhaustion with compact sets. | ||
+ | |||
+ | The cone of nonnegative measures is metrizable without further restrictions on the size of the measures (see for instance Proposition 2.6 of {{Cite|De}}). | ||
+ | |||
+ | ===Compactness of the weak$^*$ topology=== | ||
+ | If $\{\mu_k\}$ is a sequence with $\sup_k \|\mu_k\|_v < \infty$ and $X$ is $\sigma$-compact then a subsequence converges weakly$^*$. This is again a consequence of standard Banach space theory if $X$ is compact (see [[Banach-Alaoglu theorem]]), whereas the locally compact case can easily reduced to the compact one by exhaustion. More general compactness statements are possible (cp. for instance with Theorem 2 in Section 1.9 of {{Cite|EG}}). | ||
+ | |||
+ | ==Probability measures== | ||
+ | On the space of probability measures one can get further interesting properties. | ||
+ | |||
+ | ===Narrow and wide topology=== | ||
+ | The narrow and wide topology coincide on the space of probability measures on a locally compact spaces. If $X$ is compact, then the space of probability measures with the narrow (or wide) topology is also compact. However, if $X$ is not compact, the compactness of the wide topology fails: as an example take the sequence of Dirac masses $\delta_n$ on $\mathbb R$, where $n\in \mathbb N$. This sequence converges, in the wide topology, to the measure $0$. However, if one assumes ''tightness'' of the sequence of measures $\{\mu_n\}$ (cp. with \ref{e:tight}), then the sequential (pre)compactness is reestablished. More precisely (cp. with Theorem 6.1 of {{Cite|Bi}}): | ||
+ | |||
+ | '''Theorem (Prohorov)''' | ||
+ | Let $X$ be a locally compact Hausdorff space and $\{\mu_k\}$ a sequence of Radon probability measures. If | ||
+ | \begin{equation}\label{e:tight} | ||
+ | \forall \varepsilon\; \exists K\, \mbox{compact such that }\; \mu_k (X\setminus K)<\varepsilon \; \forall k\, | ||
+ | \end{equation} | ||
+ | then a subsequence converges weakly$^*$ to a probability Radon measure $\mu$. | ||
+ | |||
+ | A sequence of probability measures converging in the narrow topology is often called a ''weakly converging sequence''. | ||
+ | See [[Weak convergence of probability measures]]. | ||
+ | |||
+ | ===[[Wasserstein metric|Wasserstein metrics]]=== | ||
+ | The space of probability measures on a [[Polish space]] can be endowed with several interesting metrics, called [[Wasserstein metric|Wasserstein]] or Monge-Kantorovich distances (see Section 7.1 of {{Cite|Vi}}) and related to the [[Mass transport]] problem. The $1$-Wasserstein distance (also called Kantorovich-Rubinstein distance) is defined as | ||
+ | \[ | ||
+ | W_1 (\mu, \nu) = \sup \left\{ \int \varphi d\mu - \int \varphi d\nu : \; \varphi\in C(X, \mathbb R)\; \mbox{ with }{\rm Lip}\, (\varphi)\leq 1 \right\} | ||
+ | \] | ||
+ | (here ${\rm Lip}\, (\varphi)$ denotes the [[Lipschitz condition|Lipschitz constant]] of $\varphi$). | ||
+ | |||
+ | ==References== | ||
{| | {| | ||
|- | |- | ||
|valign="top"|{{Ref|AmFuPa}}|| L. Ambrosio, N. Fusco, D. Pallara, "Functions of bounded variations and free discontinuity problems". Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, 2000. {{MR|1857292}}{{ZBL|0957.49001}} | |valign="top"|{{Ref|AmFuPa}}|| L. Ambrosio, N. Fusco, D. Pallara, "Functions of bounded variations and free discontinuity problems". Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, 2000. {{MR|1857292}}{{ZBL|0957.49001}} | ||
+ | |- | ||
+ | |valign="top"|{{Ref|Bi}}|| P. Billingsley, "Convergence of probability measures" , Wiley (1968) {{MR|0233396}} {{ZBL|0172.21201}} | ||
|- | |- | ||
|valign="top"|{{Ref|Bo}}|| N. Bourbaki, "Elements of mathematics. Integration" , Addison-Wesley (1975) pp. Chapt.6;7;8 (Translated from French) {{MR|0583191}} {{ZBL|1116.28002}} {{ZBL|1106.46005}} {{ZBL|1106.46006}} {{ZBL|1182.28002}} {{ZBL|1182.28001}} {{ZBL|1095.28002}} {{ZBL|1095.28001}} {{ZBL|0156.06001}} | |valign="top"|{{Ref|Bo}}|| N. Bourbaki, "Elements of mathematics. Integration" , Addison-Wesley (1975) pp. Chapt.6;7;8 (Translated from French) {{MR|0583191}} {{ZBL|1116.28002}} {{ZBL|1106.46005}} {{ZBL|1106.46006}} {{ZBL|1182.28002}} {{ZBL|1182.28001}} {{ZBL|1095.28002}} {{ZBL|1095.28001}} {{ZBL|0156.06001}} | ||
+ | |- | ||
+ | |valign="top"|{{Ref|De}}|| C. De Lellis, "Rectifiable sets, densities and tangent measures" Zurich Lectures in Advanced Mathematics. European Mathematical Society (EMS), Zürich, 2008. {{MR|2388959}} {{ZBL|1183.28006}} | ||
|- | |- | ||
|valign="top"|{{Ref|DS}}|| N. Dunford, J.T. Schwartz, "Linear operators. General theory" , '''1''' , Interscience (1958) {{MR|0117523}} | |valign="top"|{{Ref|DS}}|| N. Dunford, J.T. Schwartz, "Linear operators. General theory" , '''1''' , Interscience (1958) {{MR|0117523}} | ||
|- | |- | ||
− | |valign="top"|{{Ref| | + | |valign="top"|{{Ref|EG}}|| L.C. Evans, R.F. Gariepy, "Measure theory and fine properties of functions" Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1992. {{MR|1158660}} {{ZBL|0804.2800}} |
|- | |- | ||
|valign="top"|{{Ref|Ma}}|| P. Mattila, "Geometry of sets and measures in euclidean spaces. Cambridge Studies in Advanced Mathematics, 44. Cambridge University Press, Cambridge, 1995. {{MR|1333890}} {{ZBL|0911.28005}} | |valign="top"|{{Ref|Ma}}|| P. Mattila, "Geometry of sets and measures in euclidean spaces. Cambridge Studies in Advanced Mathematics, 44. Cambridge University Press, Cambridge, 1995. {{MR|1333890}} {{ZBL|0911.28005}} | ||
|- | |- | ||
+ | |valign="top"|{{Ref|Vi}}|| Villani, Cédric Topics in optimal transportation.Graduate Studies in Mathematics, 58. American Mathematical Society, Providence, RI, 2003. {{MR|1964483}} {{ZBL|1106.90001}} | ||
|} | |} |
Latest revision as of 09:33, 16 August 2013
2010 Mathematics Subject Classification: Primary: 28A33 [MSN][ZBL] $\newcommand{\abs}[1]{\left|#1\right|}$
A concept in measure theory, determined by a certain topology in a space of measures that are defined on a certain σ-algebra $\mathcal{B}$ of subsets of a space $X$ or, more generally, in a space $\mathcal{M} (X, \mathcal{B})$ of charges, i.e. countably-additive real (resp. complex) functions $\mu: \mathcal{B}\to \mathbb R$ (resp. $\mathbb C$), often also called $\mathbb R$ (resp. $\mathbb C$) valued or signed measures. The total variation measure of a $\mathbb C$-valued measure is defined on $\mathcal{B}$ as: \[ \abs{\mu}(B) :=\sup\left\{ \sum \abs{\mu(B_i)}: \{B_i\}\subset\mathcal{B} \text{ is a countable partition of } B \right\}. \] In the real-valued case the above definition simplifies as \[ \abs{\mu}(B) = \sup_{A\in \mathcal{B}, A\subset B} \left(\abs{\mu (A)} + \abs{\mu (B\setminus A)}\right). \] The total variation of $\mu$ is then defined as $\left\|\mu\right\|_v := \abs{\mu}(X)$.
Warning: If $\mathcal{B}$ is the $\sigma$-algebra of Borel sets of a topological space $X$, we will then denote by $\mathcal{M}^b (X)$ the space of Radon signed measures, i.e. those signed measures with finite total variation such that $|\mu|$ is a Radon measure. This is actually not a restriction in many cases, for instance if $X$ is the euclidean space.
Notions of convergence
The space $\mathcal{M}^b (X, \mathcal{B})$ of $\mathbb R$ (resp. $\mathbb C$) valued measure with finite total variation is a Banach space and the following are the most commonly used topologies.
The norm or strong topology
$\mu_n\to \mu$ if and only if $\left\|\mu_n-\mu\right\|_v\to 0$. This convergence is sometimes called convergence in variation.
The weak topology
A sequence of measures $\mu_n \rightharpoonup \mu$ if and only if $F (\mu_n)\to F(\mu)$ for every bounded linear functional $F$ on $\mathcal{M}^b$.
The narrow topology
When $X$ is a topological space and $\mathcal{B}$ the corresponding $\sigma$-algebra of Borel sets, we can introduce on $X$ the narrow topology. In this case $\mu_n$ converges to $\mu$ if and only if \begin{equation}\label{e:narrow} \int f\, \mathrm{d}\mu_n \to \int f\, \mathrm{d}\mu \end{equation} for every bounded continuous function $f:X\to \mathbb R$ (resp. $\mathbb C$). The following is an important consequence of the narrow convergence when $X$ is a locally compact Hausdorff space: if $\mu_n$ converges narrowly to $\mu$, then $\mu_n (A)\to \mu (A)$ for any Borel set such that $\abs{\mu}(\partial A) = 0$ (cp. with Theorem 1(iii) of Section 1.9 in [EG]).
The wide or weak$^\star$ topology
When $X$ is a locally compact topological space and $\mathcal{B}$ the $\sigma$-algebra of Borel sets yet another topology can be introduced, the so-called wide topology, or sometimes referred to as weak$^\star$ topology. A sequence $\mu_n\rightharpoonup^\star \mu$ if and only if \eqref{e:narrow} holds for continuous functions which are compactly supported. This topology is in general weaker than the narrow topology but they do coincide when restricted to probability measures if $X$ is a Hausdorff space.
A topology analogous to the weak$^\star$ topology is defined in the more general space $\mathcal{M}^b_{loc}$ of locally bounded measures, i.e. those measures $\mu$ such that for any point $x\in X$ there is a neighborhood $U$ with $\abs{\mu}(U)<\infty$.
Warning Sequences of measures converging in the narrow (or in the wide topology) are called weakly convergent sequences by several authors (cp. with [Bi], [Ma] and [EG]). This is, however, inconsistent with the terminology of Banach spaces, see below.
Properties
Relation with functional analysis
If $X$ is compact and Hausdorff the Riesz representation theorem shows that $\mathcal{M}^b (X)$ is the dual of the space $C(X)$ of continuous functions. Under this assumption the narrow and weak$^\star$ topology coincides with the usual weak$^\star$ topology of the Banach space theory. Since in general $C(X)$ is not a reflexive space, it turns out that the narrow topology is in general weaker than the weak topology.
Metrizability of the weak$^*$ topology
On bounded subsets of $\mathcal{M}^b (X)$, the weak$^*$ topology is metrizable. If $X$ is compact, this follows directly from standard functional-analytic arguments, since $\mathcal{M}^b (X)$ is then the dual of a separable Banach space. The case of a $\sigma$-compact $X$ can be reduced to that of a compact space by exhaustion with compact sets.
The cone of nonnegative measures is metrizable without further restrictions on the size of the measures (see for instance Proposition 2.6 of [De]).
Compactness of the weak$^*$ topology
If $\{\mu_k\}$ is a sequence with $\sup_k \|\mu_k\|_v < \infty$ and $X$ is $\sigma$-compact then a subsequence converges weakly$^*$. This is again a consequence of standard Banach space theory if $X$ is compact (see Banach-Alaoglu theorem), whereas the locally compact case can easily reduced to the compact one by exhaustion. More general compactness statements are possible (cp. for instance with Theorem 2 in Section 1.9 of [EG]).
Probability measures
On the space of probability measures one can get further interesting properties.
Narrow and wide topology
The narrow and wide topology coincide on the space of probability measures on a locally compact spaces. If $X$ is compact, then the space of probability measures with the narrow (or wide) topology is also compact. However, if $X$ is not compact, the compactness of the wide topology fails: as an example take the sequence of Dirac masses $\delta_n$ on $\mathbb R$, where $n\in \mathbb N$. This sequence converges, in the wide topology, to the measure $0$. However, if one assumes tightness of the sequence of measures $\{\mu_n\}$ (cp. with \ref{e:tight}), then the sequential (pre)compactness is reestablished. More precisely (cp. with Theorem 6.1 of [Bi]):
Theorem (Prohorov) Let $X$ be a locally compact Hausdorff space and $\{\mu_k\}$ a sequence of Radon probability measures. If \begin{equation}\label{e:tight} \forall \varepsilon\; \exists K\, \mbox{compact such that }\; \mu_k (X\setminus K)<\varepsilon \; \forall k\, \end{equation} then a subsequence converges weakly$^*$ to a probability Radon measure $\mu$.
A sequence of probability measures converging in the narrow topology is often called a weakly converging sequence. See Weak convergence of probability measures.
Wasserstein metrics
The space of probability measures on a Polish space can be endowed with several interesting metrics, called Wasserstein or Monge-Kantorovich distances (see Section 7.1 of [Vi]) and related to the Mass transport problem. The $1$-Wasserstein distance (also called Kantorovich-Rubinstein distance) is defined as \[ W_1 (\mu, \nu) = \sup \left\{ \int \varphi d\mu - \int \varphi d\nu : \; \varphi\in C(X, \mathbb R)\; \mbox{ with }{\rm Lip}\, (\varphi)\leq 1 \right\} \] (here ${\rm Lip}\, (\varphi)$ denotes the Lipschitz constant of $\varphi$).
References
[AmFuPa] | L. Ambrosio, N. Fusco, D. Pallara, "Functions of bounded variations and free discontinuity problems". Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, 2000. MR1857292Zbl 0957.49001 |
[Bi] | P. Billingsley, "Convergence of probability measures" , Wiley (1968) MR0233396 Zbl 0172.21201 |
[Bo] | N. Bourbaki, "Elements of mathematics. Integration" , Addison-Wesley (1975) pp. Chapt.6;7;8 (Translated from French) MR0583191 Zbl 1116.28002 Zbl 1106.46005 Zbl 1106.46006 Zbl 1182.28002 Zbl 1182.28001 Zbl 1095.28002 Zbl 1095.28001 Zbl 0156.06001 |
[De] | C. De Lellis, "Rectifiable sets, densities and tangent measures" Zurich Lectures in Advanced Mathematics. European Mathematical Society (EMS), Zürich, 2008. MR2388959 Zbl 1183.28006 |
[DS] | N. Dunford, J.T. Schwartz, "Linear operators. General theory" , 1 , Interscience (1958) MR0117523 |
[EG] | L.C. Evans, R.F. Gariepy, "Measure theory and fine properties of functions" Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1992. MR1158660 Zbl 0804.2800 |
[Ma] | P. Mattila, "Geometry of sets and measures in euclidean spaces. Cambridge Studies in Advanced Mathematics, 44. Cambridge University Press, Cambridge, 1995. MR1333890 Zbl 0911.28005 |
[Vi] | Villani, Cédric Topics in optimal transportation.Graduate Studies in Mathematics, 58. American Mathematical Society, Providence, RI, 2003. MR1964483 Zbl 1106.90001 |
Convergence of measures. Encyclopedia of Mathematics. URL: http://www.encyclopediaofmath.org/index.php?title=Convergence_of_measures&oldid=27239