# Prove that if $\lim_{n}\mu(A)=\mu(A)$ uniformly for all measurable set, then $\lim_{n}||\mu_n-\mu||=0$.


Let $$\mu$$ and $$\mu_1,\,\mu_2,\cdots$$ be finite signed or complex measures on $$(X,\,\scra)$$. Show that $$\limm{n}{\iy}\pl\mu_n-\mu\pl=0$$ holds if and only if $$\mu_n(A)$$ converges to $$\mu(A)$$ uniformly in $$A$$ as $$n$$ approaches infinity.

The proof of "$$\Longrightarrow$$" is easy. But how to prove the reversed assertion? By definition, $$|\mu-\mu_n|(X)=\sup\lbrace\sum_{i=1}^{k}\left|(\mu-\mu_n)(A_i)\right|\{A_i\}_{i=1}^k\;\text{is a partition of}\;X\rbrace.$$ The difficulty is that $$k$$ varies as soon as $$n$$ does.

• I don't know how to use the Ullrich hint, nor do I know what a "signed or complex" measure is. However, if you let $m(\cdot)$ be the real part of such a measure then I suspect you have the property that if $A_1, ..., A_k$ are disjoint and measurable then $$m(\cup_{i=1}^kA_i) = \sum_{i=1}^k m(A_i) \quad (Fact1)$$ Now you can break the sum $\sum_{i=1}^k |m(A_i)-m_n(A_i)|$ into a sum over indices $i$ for which $m(A_i)\geq m_n(A_i)$, and the opposite, which allows you to remove the absolute values. Then use Fact 1 on the separate sums. This basically turns the unbounded $k$ into $2$. – Michael May 17 at 17:31
• So likely you could use this method on the real part. Then on the imaginary part. If this comment helps, perhaps you can do me a favor and let me know the axioms and applications for a "complex measure." I am only familiar with real-valued and nonnegative measures. – Michael May 17 at 17:42

Lemma: Let $$(X,\mathcal{M})$$ be a measurable space. Let $$\mu$$ be a complex measure on $$(X,\mathcal{M})$$, then there exists $$E\subseteq X$$ such that $$||\mu||\leq8|\mu(E)|$$.

Proof: If $$||\mu||=0$$, we are done. Suppose that $$||\mu||>0$$. For $$\varepsilon=\frac{1}{2}||\mu||>0$$, there exists a partition $$X=\cup_{j=1}^{n}E_{j}$$, where $$E_{j}\in\mathcal{M}$$ such that $$||\mu||-\varepsilon<\sum_{j=1}^{n}|\mu(E_{j})|$$. For each $$j$$, write $$\mu(E_{j})=\alpha_{j}+i\beta_{j}$$, where $$\alpha_{j},\beta_{j}\in\mathbb{R}$$. Define subsets $$I_{1},I_{2},I_{3},I_{4}$$ of $$\{1,\ldots,n\}$$ by $$I_{1}=\{j\mid\alpha_{j}\geq0\}$$, $$I_{2}=\{j\mid\alpha_{j}<0\}$$, $$I_{3}=\{j\mid\beta_{j}\geq0\}$$, and $$I_{4}=\{j\mid\beta_{j}<0\}$$. Note that $$I_{1}\cap I_{2}=I_{3}\cap I_{4}=\emptyset$$. Further write $$\alpha_{j}=\alpha_{j}^{+}-\alpha_{j}^{-}$$, $$\beta_{j}=\beta_{j}^{+}-\beta_{j}^{-}$$, where for any $$x\in\mathbb{R}$$, $$x^{+}=\max(x,0)$$ and $$x^{-}=\max(-x,0)$$. Let $$s_{1}=\sum_{j=1}^{n}\alpha_{j}^{+}$$, $$s_{2}=\sum_{j=1}^{n}\alpha_{j}^{-}$$, $$s_{3}=\sum_{j=1}^{n}\beta_{j}^{+}$$, and $$s_{4}=\sum_{j=1}^{n}\beta_{j}^{-}$$.

Observe that $$\begin{eqnarray*} & & \sum_{j=1}^{n}|\mu(E_{j})|\\ & = & \sum_{j=1}^{n}|(\alpha_{j}^{+}-\alpha_{j}^{-})+i(\beta_{j}^{+}-\beta_{j}^{-})|\\ & \leq & \sum_{j=1}^{n}\alpha_{j}^{+}+\alpha_{j}^{-}+\beta_{j}^{+}+\beta_{j}^{-}\\ & = & s_{1}+s_{2}+s_{3}+s_{4}\\ & \leq & 4\max(s_{1},s_{2},s_{3},s_{4}). \end{eqnarray*}$$ Consider the case that $$s_{1}=\max(s_{1},s_{2},s_{3},s_{4})$$. Define $$E=\cup_{j\in I_{1}}E_{j}$$, then $$\begin{eqnarray*} |\mu(E)| & = & |\sum_{j\in I_{1}}\mu(E_{j})|\\ & = & |s_{1}+i\gamma_{1}|\\ & \geq & s_{1}\\ & = & \max(s_{1},s_{2},s_{3},s_{4}), \end{eqnarray*}$$ for some $$\gamma_{1}\in\mathbb{R}$$ (which is not important).

If $$s_{2}=\max(s_{1},s_{2},s_{3},s_{4})$$, we define $$E=\cup_{j\in I_{2}}E_{j}$$, then $$|\mu(E)|=|-s_{2}+i\gamma_{2}|\geq s_{2}=\max(s_{1},s_{2},s_{3},s_{4})$$.

If $$s_{3}=\max(s_{1},s_{2},s_{3},s_{4})$$, we define $$E=\cup_{j\in I_{3}}E_{j}$$, then $$|\mu(E)|=|\gamma_{3}+is_{3}|\geq s_{3}=\max(s_{1},s_{2},s_{3},s_{4})$$.

If $$s_{4}=\max(s_{1},s_{2},s_{3},s_{4})$$, we define $$E=\cup_{j\in I_{4}}E_{j}$$, then $$|\mu(E)|=|\gamma_{4}-is_{4}|\geq s_{4}=\max(s_{1},s_{2},s_{3},s_{4}).$$

Hence, we obtain $$\begin{eqnarray*} & & \frac{1}{2}||\mu||\\ & = & ||\mu||-\varepsilon\\ & < & \sum_{j=1}^{n}|\mu(E_{j})|\\ & \leq & 4\max(s_{1},s_{2},s_{3},s_{4})\\ & \leq & 4|\mu(E)|. \end{eqnarray*}$$ That is, $$||\mu||\leq8|\mu(E)|$$.

/////////////////////////////////////////////////////////////////////

Now, we go back to the original question. Suppose that $$\mu_{n}$$ and $$\mu$$ are complex measures. For each $$n$$, define $$\nu_{n}=\mu_{n}-\mu$$. Then $$\nu_{n}(A)\rightarrow0$$ uniformly on $$A\in\mathcal{M}$$, in the sense that: For any $$\varepsilon>0$$, there exists $$N$$ such that $$|\nu_{n}(A)|<\varepsilon$$ whenever $$n\geq N$$ and $$A\in\mathcal{M}$$.

Let $$\varepsilon>0$$ be given. Choose $$N$$ such that $$|\nu_{n}(A)|<\varepsilon/8$$ whenever $$n\geq N$$ and $$A\in\mathcal{M}$$. For each $$n$$, choose $$E_{n}\in\mathcal{M}$$ such that $$||\nu_{n}||\leq8|\nu_{n}(E_{n})|$$. Now, for any $$n\geq N$$, we have $$\begin{eqnarray*} & & ||\nu_{n}||\\ & \leq & 8|\nu_{n}(E_{n})|\\ & < & 8\cdot\frac{\varepsilon}{8}\\ & = & \varepsilon. \end{eqnarray*}$$ Therefore $$||\nu_{n}||\rightarrow0$$.

• Well, this proof seems essentially the same as my first comment. Under the assumption that the subtraction of two (complex) measures is also a (complex) measure then the original Ullrich hint makes sense (though I agree $1/8$ is a better/easier statement than $1/4$). I usually don't +1 full answers to homework questions (there seems nothing left for the asker to do) but I have +1'd your answer. I guess one could argue that my original comment didn't leave much more for the asker to do. – Michael May 17 at 22:22
• Very clear, thanks. – Tao X May 18 at 4:06

Hint:

Prop If $$\mu$$ is a complex measure on $$X$$ then there exists $$E\subset X$$ with $$|\mu(E)|\ge ||\mu||/\pi$$.

Getting $$|\mu(E)|\ge||\mu||/c$$ for $$c>\pi$$ is trivial from Lemma 6.3 in Rudin, which says

Lemma If $$z_1,\dots z_n\in\Bbb C$$ there exists $$S\subset\{1,\dots,n\}$$ with $$\left|\sum_{j\in S}z_j\right|\ge \frac 1\pi\sum_{j=1}^n|z_j|$$.

It's not so clear how to get the Prop as stated from the lemma. But one can simply adapt the proof of the lemma:

Proof of the Prop, cribbed from 6.3 in Rudin:

There exist a finite positive measure $$\nu$$ and a real-valued function $$\phi$$ such that $$d\mu=e^{i\phi}d\nu.$$For $$\theta\in\Bbb R$$ let $$E(\theta)=\{x\in X:\cos(\phi(x)-\theta)\ge0\}$$and define $$F(\theta)=|\mu(E(\theta))|.$$Note that $$F(\theta)=\left|\int_{E(\theta)}e^{-i\theta}d\mu\right|\ge\Re \int_{E(\theta)}e^{i(\phi-\theta)}d\nu=\int_X\cos^+(\phi(x)-\theta)\,d\nu(x).$$ Hence $$F$$ is continuous, so it has a maximum $$F(\theta_0)$$. And Fubini shows that $$F(\theta_0)\ge\frac1{2\pi}\int_0^{2\pi}F(\theta)\,d\theta =\frac1{2\pi}\int_X\int_0^{2\pi}\cos^+(\phi(x)-\theta)\,d\theta d\nu(x)=\frac1\pi\int_Xd\nu=\frac1\pi||\mu||.$$

QED.

And $$1/\pi$$ is best possible:

Define a complex Borel measure $$\mu$$ on $$[-\pi/2,\pi/2)$$ by $$\mu(E)=\int_Ee^{it}\,dt,$$so $$||\mu||=2\pi$$. It's not hard to show that $$|\mu(E)|\le 2$$ for every $$E$$.

• I have +1'd this answer in view of the Danny answer that suggests that the subtraction of two (complex) measures is also a measure. Yet, it seems nobody will care enough to tell me the use of (or axioms for) complex measures as I ask in comments above. I suspect that complex measures are a bit of a distraction for this problem and the same result could have been illuminated by asking the question with respect to ordinary (real-valued and nonnegative) measures. – Michael May 17 at 22:16
• A minor issue is that I would agree on existence of $E$ that gets $|\mu(E)|\geq (||\mu||/4)\rho$ for any $\rho$ that satisfies $0<\rho<1$, though the case $\rho=1$ (perhaps true) seems to be more involved to prove, at least it is not obvious to me... – Michael May 17 at 22:19
• @Michael I didn't think about the lemma carefully; maybe $1/4$ is wrong, the exact constant doesn't matter. You can find the definition of "complex measure" in more or less any book on real analysis. Or see en.wikipedia.org/wiki/Complex_measure – David C. Ullrich May 17 at 22:25
• Oh, I see from the link that the main application is likely integration over $\mathbb{C}$. I was wondering because the structure of $\mathbb{C}$ only seems useful if you do multiplications (else we may as well just take a vector of measures), and I didn't think of a reason why you would ever multiply $\mu(A)\mu(B)$. But for integration it makes sense I guess to multiply a "small" "$d\mu(A)$" by some complex function. – Michael May 17 at 22:29
• Here is also a proof of the lemma (with the claim that the factor $\frac 1\pi$ is optimal if $n$ is not restricted). – Martin R May 18 at 15:43