Proof of inversion formula in Probability

1.1k Views Asked by At

Theorem : Let $X$ be a real random variable such that $\phi_X \in L^1$ i.e $\int_{\mathbb{R}} \vert \phi_X(t) \vert < \infty$, then $X$ has density $f_X(x) \in C_b(\mathbb{R})$ given by $$f_X(x) = \frac{1}{2\pi}\int_{\mathbb{R}}\phi_X(t)e^{-itx}dt$$

(Where $\phi_X$ denotes the characteristic function of $X$)

The proof articulates in two parts : The first where we assume we already know $X$ has a density $f$ and we prove the inequality above, the second where we don't know such $f$ exists and we lead back to the first case using the trick of taking $X+\epsilon N$ where $N \sim N(0,1)$.

The sketch of the first part of the proof, which is where my problems are is the following : We consider a $g \geq 0, g \in C_{b}(\mathbb{R})$, such that $g = 0$ outside a compact, i.e $g \in C_{c}(\mathbb{R})$ and we apply the isometry lemma to $f_X + g$ which leads us to (Using Fubini-Tonelli and the fact that $\frac{1}{2\pi}\int_{\mathbb{R}}\phi_X(t)e^{-itx}dt$ is real)

$$\int g(x)f_X(x)dx = \mathbb{E}[g(X)] = \int g(x)[\frac{1}{2\pi}\int_{\mathbb{R}}\phi_X(t)e^{-itx}dt]dx$$

Which concludes the proofs for $g$ assumed as mentioned at the beginning. But how to extend the result to just continuos and bounded functions ?

During this next proof the following three observation are made :

$1) \frac{1}{2\pi}\int_{\mathbb{R}}\phi_X(t)e^{-itx}dt$ is continuos in $x$ (and I was able to prove it as a consequence of dominated convergence theorem, prooving continuity by sequences).

$2) \frac{1}{2\pi}\int_{\mathbb{R}}\phi_X(t)e^{-itx}dt \geq 0$

3)$\hspace{0.1cm} \exists \hspace{0.1cm} g_n \uparrow 1 $ with $g_n \in C_b(\mathbb{R})$

I was able to prove step $2)$ in the following way with an hint which I was unable to prove (so this should be the way to prove it). The proof goes like this : if it was strictly negative, by continuity (prooved in the first observation) exists an open set $U$ where the function is still strictly negative, and then I can find a $g > 0, g \in C_b(\mathbb{R})$ defined on $U$ and conclude by contradiction thanks to the fact that in this case we would have $0 > \frac{1}{2\pi}\int_{\mathbb{R}}\phi_X(t)e^{-itx}dt = E[g(X)] > 0$, thanks to $g$.

As far concerned the third point I don't know where to start, I thought it could be useful to use a simplified version of Urysohn Lemma but unsuccefully. I was unable to prove the highlighted sentences, there are any way to explicit such functions ? Any direct proof, explicit or not, would be appreciated, and some reference of using Fourier trasfrom in Probability as well, those seemed nice tricks to know.

1

There are 1 best solutions below

0
On BEST ANSWER

You do not need to assume that $g$ has compact support in the first part, I believe.

You can use a trick very similar to what you propose in the beginning to show that \begin{align} \lim_{n\to\infty}f_{X + \frac{1}{n}N}(t) = \frac{1}{\sqrt{2\pi}}\int_{R}e^{-\text{i}ts}\varphi_{X}(s)\text{d}s = f(t) \end{align} where $N\sim N(0,1)$. Thus it follows, that $f$ is strictly positive and measurable, and we can therefore consider the measure $\nu$ with density $f$.

Now, if we can show that \begin{align} \int_{\mathbb{R}}\psi \text{d}\nu = \int_{\mathbb{R}}\psi \text{d}\mathbb{P}_{X}, \quad \psi \in C_{c}(\mathbb{R}) \end{align} then $\nu = \mathbb{P}_{X}$, which would imply that $f = f_{X}$, as desired. But showing that the integrals are equal is simply using dominated convergence on the integrals wrt. $\mathbb{P}_{X+\frac{1}{n}N}$.