This is somewhat a basic question, but I'm having difficulty proceeding with a certain part of the proof. I was reading Billingsley "Convergence of Probability Measures", and I encountered the following question:

Question: Given a metric space $X$ and $\mathscr{B}$ the Borel sigma algebra, prove that a probability measure $P$ on $X$ is tight iff $$\forall A \in \mathscr{B},\quad \sup \{P(K): K\subset A, K \mbox{ compact}\} = P(A)$$

My Proof:

($\Rightarrow$) Put $A=X$. Then we have $$1 = P(X) = \sup \{P(K): K\subset X, K \mbox{ compact}\}$$ By definition of sup, we get $\forall \epsilon > 0$, $\exists K$ compact such that $$P(K) \geq 1-\epsilon$$ Thus P is tight.


Billingsley suggests using the following theorem:

Probability measures on metric spaces are regular. That is $$\forall A \in \mathscr{B}, \forall \epsilon > 0, \exists F \mbox{ closed}, G \mbox{ open such that}$$ $$F \subset A \subset G, \quad P(G \setminus F) < \epsilon$$

I do not know how to use compactness here. I reasoned that if I could prove that $\forall \epsilon > 0$, $\exists K \subset A$ compact such that $P(A\setminus K) < \epsilon$, I would be done. Since the compact sets in "tightness", needn't be subsets of A, I don't know how to use it. I was able to show the easy inequality which is:

$$\sup \{P(K): K\subset A, K \mbox{ compact}\} \leq P(A)$$ (since $K\subset A$)

I would appreciate any ideas, hints and tips (if not answers) to this. References are also welcome. I searched for "Tight Regular Measures" and "Tight Probability" but to no useful results.


First, we will see how Billingsley's claim helps to conclude. For each $n$, consider $K_n$ compact such that $\mu(K_n)>1-n^{-1}$. We can assume the sequence $(K_n,n\geqslant 1)$ to be increasing. Take $B\in\mathcal B(X)$ and $\varepsilon>0$. By the claim, there is $F\subset B$ such that $\mu(B\setminus F)<\varepsilon$. Fix $n$ such that $\varepsilon>n^{-1}$. Then $$\mu(B\cap K_n)=\mu(B)-\mu(B\setminus (B\cap K_n))=\mu(B)-\mu(B\setminus K_n)\\\geqslant \mu(B)-\mu(K_n^c)\geqslant \mu(B)-n^{-1}\geqslant \mu(B)-\varepsilon.$$ We thus get $\mu(F\cap K_n)>\mu(B)-2\varepsilon$, and $F\cap K_n$ is a compact subset of $B$.

Now we shall see the claim is true. First, note that in a metric space, an closed set is a countable intersection of open sets, and using the fact that the measure is finite, we deduce that the assertion is true for $B$ open.

Then the remaining task is to show that the collection of the $B$ for which the property holds is a $\sigma$-algebra.

  • $\begingroup$ Thanks Davide. So it doesn't directly follow... Also I wasn't aware that a closed set was a countable intersection of open sets. Are you sure this doesn't require separability of the space?, since Billingsley quotes it for the general metric space. $\endgroup$ – Gautam Shenoy May 8 '13 at 9:20
  • 1
    $\begingroup$ If $F$ is closed, then $F:=\bigcap_{n\geqslant 1}O_n$, where $O_n:=\{x,d(x,O)< n^{-1}\}$, and there is no need of separability. $\endgroup$ – Davide Giraudo May 8 '13 at 9:58
  • $\begingroup$ No wait. I got it. Your statement "We thus get ... compact subset of B" completed the proof. $\endgroup$ – Gautam Shenoy May 8 '13 at 10:09
  • $\begingroup$ Hi Davide, did you just prove that any probability measure on Polish space is tight? For example here in this note, they have proved it but I am not understanding from the line $\because P$ is a probability measure then there is a $k$ for some $n_k$....., could you explain me ''why'' for this particular line is true? Here Thm 2.49 stat.columbia.edu/~porbanz/teaching/G6106S16/… $\endgroup$ – Marso Jul 3 at 10:14

Thanks to Davide Giraudo for this:

The remaining steps are as follows. Let $K_n$, be a collection of compact sets such that $\forall n \geq 1$ $$P(K_n) > 1-\frac{1}{2n}$$

Now as per the provided theorem, given $A \in \mathscr{B}$, and $n \geq 1$ we have $F\subset A$ closed such that $$P(A) - P(F) = P(A \setminus F) < \frac{1}{2n}$$

Now as Davide Giraudo showed, $P(A \cap K_n) \geq P(A) - \frac{1}{2n}$ Add the previous two statements, and use the fact that $F \subset A$ to get $$P(A) - \frac{1}{n} \leq P(A \cap K_n) - P(A) + P(F)$$ $$ = P(F) - P(A \cap K_n^c) = P(F \cap K_n) + P(F \cap K_n^c)- P(A \cap K_n)$$ $$\leq P(F \cap K_n)$$

Now $F \cap K_n$ is a closed subset of a compact set in a metric space and hence is compact. Additionally $F \cap K_n \subset A$ !!.

Therefore we have $$P(A) \leq P(F \cap K_n) + \frac{1}{n} \leq \sup\{P(K): K \subset A, \mbox{K compact}\} + \frac{1}{n}$$ $$\Rightarrow P(A) \leq \sup\{P(K): K\subset A, \mbox{K compact}\} + \frac{1}{n}$$ Take $n \to \infty$ to get the desired inequality.



An alternative approach to prove $P(A) \leq \sup\{P(K): K \subset A, K \hspace{1mm} compact\}$.

For any given $n$, we can find $F_n$ closed such that $P(A) \leq P(F_n) + \frac{1}{n}$. (By regularity of the measure P)

Also, we can find $K_n$ compact such that $P(K_n) > 1 - \frac{1}{n}$ (since we assume P is tight). Now $F_n \cap K_n$ is compact with $P(F_n \cap K_n) = P(F_n) + P(K_n) - P(F_n \cup K_n)$.

Hence we have, $P(F_n) = P(F_n \cap K_n) - P(K_n) + P(F_n \cup K_n) \leq P(F_n \cap K_n) - 1 + \frac{1}{n} + 1$.

We thus have $P(A) \leq P(F_n \cap K_n) + \frac{1}{2n}$.

Therefore, for any $n$, we have $J_n := F_n \cap K_n \subset A$ and $J_n$ compact such that $P(A) \leq P(J_n) + \frac{1}{2n}$.

Since the above is true for all $n$, we have $P(A) \leq \sup\{P(J): J \subset A, J \hspace{1mm} compact\}$.


Let $\epsilon>0$ be given. There are closed F (By Theorem. 1.1 of billingsly) and compact $K$ such that $$P(K)>1-\frac{\epsilon}{2},$$ i.e. $$P(K')<\epsilon/2,$$ and $$P(A-F)<\epsilon/2,$$ Then we have $P(A-(F\cap K ))<\epsilon$, which proposes $F\cap K$ as desired compact set.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.