Suppose the sequence $\{\mu_n : n \in \mathbb{N} \}$ of signed measures on $(X, \mathcal{F})$ converges setwise to the signed measure $\mu$, by which we mean that $\lim_{n \to \infty}\mu_n(A) = \mu(A)$ for every $A \in \mathcal{F}$.

Under what further assumptions does $\{\mu_n : n \in \mathbb{N} \}$ converge in total variation to $\mu$?

Recall that the total variation norm $\| \cdot \|$ is defined by $\|\mu \| = \sup \sum_i |\mu(E_i)|$, where the supremum is over countable partitions of $X$, and $\{\mu_n : n \in \mathbb{N} \}$ converges in total variation to $\mu$ if $\|\mu_n - \mu \| \to 0$ as $n \to \infty$.

I had thought this would be a standard topic, but after googling and looking through some textbooks I still haven't been able to find any results.

  • $\begingroup$ what is $\mu_\infty$? $\endgroup$ – Filburt Mar 4 '17 at 14:34
  • $\begingroup$ A typo. Thanks for pointing it out. $\endgroup$ – grndl Mar 4 '17 at 16:55
  • $\begingroup$ @aduh I myself don't have the answer off the top of my head, but if you are desperate for an answer I can place a bounty on your question. $\endgroup$ – Clement C. Mar 6 '17 at 14:18
  • 1
    $\begingroup$ Well, it's not like these points have any actual meaning anyway. $\endgroup$ – Clement C. Mar 6 '17 at 15:33
  • 1
    $\begingroup$ Mildly relevant (although it's rather about when weak convergence implies setwise (strong) convergence, with one detour about when weak convergence implies TV convergence): hindawi.com/journals/ijsa/1997/614526/abs See Lemma 4.1 and Corollary 4.2. $\endgroup$ – Clement C. Mar 8 '17 at 20:33

I'm going to try to mimick the proof of Corollary 4.2 in the paper linked by @Clement C. Comments and generalizations are appreciated.

We first suppose that $\{\mu_n \}$ is a sequence of probability measures on $(X, \mathcal{F})$ that converges setwise to the probability measure $\mu$. Moreover, suppose $\mu_n(A) = \int f_n d \lambda$ and $\mu(A) = \int_A f d \lambda$, where $\lambda$ is a $\sigma$-finite measure on $(X, \mathcal{F})$ and $f_n \to f$ a.e. ($\lambda$).

Now, setwise convergence implies $\int f_n d\lambda \to \int f d\lambda$. By Scheffe's Lemma, $\int |f_n - f|d\lambda \to 0$. And this in turn implies $\|\mu_n - \mu \| \to 0$.

In order to upgrade setwise convergence to convergence in total variation, we have assumed that $\mu, \mu_1, \mu_2,...$ are probabilities (although this should work for any finite, non-negative measures) that are absolutely continuous with respect to $\lambda$, and that the densities $f_n = d\mu_n/d\lambda$ converge almost everywhere to $f = d\mu/d\lambda$.


I don't have an answer to your question, just a few comments. If you want $\Vert\mu_{n}-\mu\Vert(X)\rightarrow0$ then a necessary condition is that $\sup_{n}\Vert\mu_{n}\Vert(X)<\infty$. So let's assume that.

Now If you assume that all the measures $\mu_{n}$ and $\mu$ are absolutely continuous with respect to a positive measure $\nu$, then $\mu_{n}(E)=\int _{E}f_{n}\,d\nu$ and $\mu(E)=\int_{E}f\,d\nu$ and $\int_{E}f_{n}% \,d\nu\rightarrow\int_{E}f\,d\nu$ for every $E$. Together with the fact that $\sup_{n}\Vert\mu_{n}\Vert(X)=\sup_{n}\int_{X}|f_{n}|\,d\nu<\infty$, this condition is equivalent to weak convergence in $L^{1}(X,\nu)$ (it's a theorem of Dunford-Pettis). Thus $f_{n}\rightharpoonup f$ in $L^{1}(X,\nu)$. It follows (again by a theorem of Dunford-Pettis) that for every $\varepsilon>0$ there is $\delta>0$ such that for all $n$, $$ \int_{E}|f_{n}|\,d\nu\leq\varepsilon $$ for every measurable set $E$ with $\nu(E)\leq\delta$. Also for every $\varepsilon>0$ there exists $E_{\varepsilon}$ measurable, with $\nu (E_{\varepsilon})<\infty$ such that for all $n$, $$ \int_{X\setminus E}|f_{n}|\,d\nu\leq\varepsilon. $$

Next by Vitali's convergence theorem a sequence $\{f_{n}\}$ converges strongly in $L^{1}(X,\nu)$ if and only if $\{f_{n}\}$ converges to $f$ in measure and satisfies the previous two conditions. In other words a sequence $\{f_{n}\}$ converges strongly to $f$ in $L^{1}$ if and only if it converges to $f$ weakly in $L^{1}$ and in measure.

This is a very special case, very far from replying to your original question.

Concerning your answer and the paper linked by @Clement C., for real-valued functions it becomes $$ \int_{X}|f_{n}|\,d\nu\rightarrow\int_{X}|f|\,d\nu. $$

So going back to the original question, the analog would be $\Vert\mu_{n} \Vert(X)\rightarrow\Vert\mu\Vert(X)$. If you assume this, you can prove that $\Vert\mu_{n}\Vert(E)\rightarrow\Vert\mu\Vert(E)$.

To see this, fix $\varepsilon>0$ and find countable partition of $E$ such that $$ \Vert\mu\Vert(E)\leq\sum_{k}|\mu(E_{k})|+\varepsilon. $$ Since the series is convergent, we can find $l$ such that $\sum_{k=1}^{\infty }|\mu(E_{k})|\leq\sum_{k=1}^{l}|\mu(E_{k})|+\varepsilon$. Then \begin{align*} \Vert\mu\Vert(E) & \leq\sum_{k=1}^{l}|\mu(E_{k})|+2\varepsilon=\lim _{n\rightarrow\infty}\sum_{k=1}^{l}|\mu_{n}(E_{k})|+2\varepsilon\leq \liminf_{n\rightarrow\infty}\sum_{k=1}^{\infty}|\mu_{n}(E_{k})|+2\varepsilon \\ & \leq\liminf_{n\rightarrow\infty}\Vert\mu_{n}\Vert(E)+2\varepsilon. \end{align*} Now let $\varepsilon\rightarrow0$. To prove the other inequality, use the previous inequality for $X\setminus E$ to write \begin{align*} \Vert\mu\Vert(E) & =\Vert\mu\Vert(X)-\Vert\mu\Vert(X\setminus E)\geq\Vert \mu\Vert(X)-\liminf_{n\rightarrow\infty}\Vert\mu_{n}\Vert(X\setminus E)\\ & =\lim_{n\rightarrow\infty}\Vert\mu_{n}\Vert(X)+\limsup_{n\rightarrow\infty }(-\Vert\mu_{n}\Vert(X\setminus E))\\ & =\limsup_{n\rightarrow\infty}(\Vert\mu_{n}\Vert(X)-\Vert\mu_{n} \Vert(X\setminus E))=\limsup_{n\rightarrow\infty}\Vert\mu_{n}\Vert(E). \end{align*} So now you have $\Vert\mu_{n}\Vert(E)\rightarrow\Vert\mu\Vert(E)$ for every $E$. This is tight convergence. Unfortunately you are still far from $\Vert \mu_{n}-\mu\Vert(X)\rightarrow0$. Wish I could help more but this is all I can say....

  • $\begingroup$ Why is my application of Scheffe's Lemma wrong? Since $f_n$ is a probability density, $|f_n| = f_n$ a.s. $\endgroup$ – grndl Mar 12 '17 at 15:40
  • $\begingroup$ I was commenting to " (although this now seems inessential) ". You need probability measures. If $f_n\ge 0$ then your proof works fine $\endgroup$ – Gio67 Mar 12 '17 at 15:45
  • $\begingroup$ Ah, I see. Thanks. What I had in mind with "although this now seems inessential" was that the proof would work for any finite ($\sigma$-finite?) measure. I agree it does not work for general signed measures, however. $\endgroup$ – grndl Mar 12 '17 at 15:47
  • $\begingroup$ I edited my answer. You are right. My comment was misleading $\endgroup$ – Gio67 Mar 12 '17 at 15:49
  • $\begingroup$ I will correct my definition of total variation. Thanks for pointing that out. I had confused it with the common definition of total variation distance. $\endgroup$ – grndl Mar 12 '17 at 15:49

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.