So this problem comes from Achim Klenke's Probability Theory: A comprehensive course that I am going through on my own. It appears as Exercise 6.2.1 in a chapter on convergence theorems. I have tried some angles but have not been able to crack it.

Let $H \in \mathcal{L}^1(\mu)$ with $H > 0$ $\mu$- almost everywhere and let $(E,d)$ be a separable complete metric space. For measurable $f,g : \Omega \to E$, define:

$$ d_H(f,g) := \int_{\Omega} \min\{1 , d(f(\omega),g(\omega)) \} H(\omega) \ \mu(\mathrm{d}\omega)$$

i) Show that $d_H$ is a metric that induces convergence in measure

ii) Show that $d_H$ is complete if $(E,d)$ is complete.

I have been able to prove part (i) by splitting the integral up in various ways and using DCT and so on. But I have not been able to prove (ii). I thought the following Lemma proved earlier in the chapter could be useful,

Corollary 6.15 Let $(E,d)$ be a separable complete metric space. Let $(f_n)_{n\in \mathbb{N}}$ be a Cauchy sequence in measure in $E$, that is, for any $A \in \mathcal{A}$ with $\mu(A) < \infty$ and any $\epsilon > 0$ we have,

$$\mu(A \ \cap \ \{d(f_m, f_n) > \epsilon \}) \xrightarrow{\ m, n \to \infty } 0 $$

Then $(f_n)_{n\in \mathbb{N}}$ converges in measure.

My idea was to prove that being Cauchy in $d_H$ entails being Cauchy in measure, which would then allow me have that $(f_n)_{n\in\mathbb{N}}$ converges in measure and so by part (i) it should converge in $d_H$ as well and hence we have that $d_H$ is complete.

One way I thought I could try to achieve this is by supposing we have some Cauchy sequence in $d_H$, that is, taking $\epsilon > 0$ arbitrary we have some $N$ such that,

$$d_H(f_m, f_n) < \epsilon, \ \forall \ m, n \geq N $$

Taking $\delta > 0$ arbitrary we seek to show that (assuming WLOG that $\mu(\Omega) < \infty$),

$$\mu(\{d(f_m, f_n) > \delta \}) < \epsilon $$

Or at least, some one-to-one function of $\epsilon$ (so that we can then pick the appropriate one when invoking the assumption of $d_H$ being Cauchy). Taking for now $\delta \geq 1$ we can see that in fact,

$$d_H(f_m, f_n) = \int_{\{d(f_m, f_n) \geq 1 \}} \min \{1, d(f,g) \}H \ \mathrm{d}\mu + \int_{\{d(f_m, f_n) < 1 \}} \min \{1, d(f,g) \}H \ \mathrm{d}\mu \leq \epsilon \\ \\ \Rightarrow \ \ \int_{\{d(f_m, f_n) \geq 1 \}} H \ \mathrm{d}\mu + \int_{\{d(f_m, f_n) < 1 \}} H d(f,g) \ \mathrm{d}\mu \leq \epsilon \\ \\ \Rightarrow \ \ \int_{\{d(f_m, f_n) \geq 1 \}} H \ \mathrm{d}\mu \leq \epsilon$$

At this point I wanted to use what I know about $H$ to try to get this inequality into one that uses $\mu(\{d(f_m, f_n) \geq 1 \})$ and the fact that $H$ is in $\mathcal{L}^1$ but I cannot think of how to do it. Intuitively there should be something I can do manipulate to separate the measure of the set the integral is being done over from the function. For instance if $H = c$ was a constant then we simply need to divide by that constant to get an estimate for the measure and hence allow us to conclude Cauchyness in measure. But I cannot think of how to proceed here. Am I on the right track?


For convergence in measure, assume $d_H(f_n,f)\to 0$. For fixed $\epsilon \le 1$ let $E_n=\{\omega : d(f(\omega),g(\omega))\ge \epsilon\}$ so $\epsilon \chi_{E_n}\le \min(1,d(f(\omega),g(\omega)))$ and therefore $\int_{E_n} H\text{d}\mu \le \frac 1\epsilon d_H(f_n,f)$. As you mentioned, if $H\ge k$ on $\Omega$ we would be done, but this need not be the case, however, we can show a similar property that is sufficient for our purpose. To wit, define $A_k=\{\omega : \frac 1k \le H(\omega) \le \frac 1{k-1}\}$ so $\bigcup_{k >1} A_k \subset \Omega$ implies $\sum_{k>1} \mu(A_k)$ converges. Hence, for every $\delta>0$ there exists some $N$ s.t. $\sum_{k>N} \mu( A_k)<\delta$. Moreover, $H\ge \frac 1N$ on $A = \bigcup_{1<k\le N}A_k$ and $\mu(A^c)<\delta$. Therefore, $$\begin{align}\frac 1\epsilon d_H(f_n,f)\ge\int_{E_n} H\text{d}\mu&=\int_{A\cap E_n} H+\int_{A^c\cap E_n} H \\ &\ge \frac 1N\mu(A\cap E_n) \end{align}$$ It follows that $\mu(E_n)=\mu(E_n\cap A)+\mu(E_n\cap A^c)\le \mu(E_n\cap A)+\mu(A^c)\le \frac N\epsilon d_H(f_n,f)+\delta$. Now choose $n$ large enough so $d_H(f_n,f) \le \frac \epsilon N\delta$ to get $\mu(E_n)\le 2\delta$, which shows $d_H$ induces convergence in measure.
Now, if $f_n$ is Cauchy, it is Cauchy in measure, and you can essentially adapt this answer (although Cor. 6.13 in Klenke's book is also useful) to show $f_n \to f$ in measure because $d$ is complete. Now let $E = \{ \omega : d(f_n(\omega),f(\omega))\ge \epsilon\}$ and compute $$\begin{align}d_H(f_n,f) &\le \int_EH\text{d}\mu +\epsilon\int_{E^c} H\text{d}\mu \\ & \le\epsilon +\epsilon||H||_1 \end{align} $$ Whenever $\mu(E)$ is small enough because $H \in L^1$.

  • $\begingroup$ "Now, if $(f_n)$ is Cauchy, it is Cauchy in measure". This is the exact part where I am having trouble accepting. We did prove that convergence in $d_H$ induces convergence in measure, but why does this mean that if $(f_n)$ is Cauchy in $d_H$ then it is Cauchy in measure? Is this supposed to be obvious and I am missing it? Thanks. $\endgroup$ – symchdmath Dec 16 '18 at 7:30
  • 1
    $\begingroup$ @symchdmath yes, notice that when we prove that $\mu(E_n)\le \frac N\epsilon d_H(f_n,f)+\delta$, we could easily replace $f$ with $f_m$. So long as $n,m$ are large enough, if $f_n$ is Cauchy, $d_H(f_n,f_m)$ can be made less than $\frac \epsilon N \delta$, and Cauchy in measure still follows. Perhaps some relabeling of $E_n$ to $E_{n,m}$ would make it clearer, but the key fact remains that $d_H(f_n,f_m)$ can be made really small. $\endgroup$ – Guacho Perez Dec 16 '18 at 7:47

In any metric space if a Cauchy sequence has a convergent subsequence then the whole sequence converges. First choose an increasing sequence of integers $n_k$ such that $\mu (d_H(f_{n_k},f_{n_j})>\frac 1 k)< \frac 1 k$ for all $j \geq k$ and apply the corollary to the subsequence $(f_{n_k})$.

  • $\begingroup$ Do you mean that $\mu(\{d(f_{n_k}, f_{n_j} > 1/k\}) < 1/k$? How are you sure that you can find such a sequence like that given only the assumption that $(f_n)$ is Cauchy in $d_H$. Thanks. $\endgroup$ – symchdmath Dec 16 '18 at 5:20
  • 1
    $\begingroup$ @symchdmath because you proved that convergence in $d_H$ induces convergence in measure. $\endgroup$ – Guacho Perez Dec 16 '18 at 5:23
  • $\begingroup$ Sorry I am not sure I am following fully. Given a Cauchy sequence in $d_H$ we have not proven that it converges as that is the point of the question. So how can we use that convergence in $d_H$ induces convergence in measure? Thanks. $\endgroup$ – symchdmath Dec 16 '18 at 6:13

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.