$X$ is a random variable such that there exist a sequence $(x_n)_n$ in $\mathbb{R}^*,\lim_nx_n=0,$ and for all $n \in \mathbb{N},\phi_X(x_n)=e^{-\sigma^2x^2_n}, \sigma \in \mathbb{R}.$

Prove that $X$ is normally distributed (possible degenerate).

The case where $\sigma=0$ (degenerate distribution) was asked here.

Any suggestions on how to deal with $\sigma \neq 0$?

  • $\begingroup$ Hint: define $X_n=X+x_n$. $\endgroup$ – Martín Vacas Vignolo Mar 4 at 7:02
  • $\begingroup$ $X_n$ converges in distribution to $X,$ and $\phi_{X_n}(x)=e^{ixx_n} \phi_X(x),\phi_{X_n}(x_n)=e^{x^2_n(i-\sigma^2)},$ why introducing $X_n$? $\endgroup$ – Kurt.W.X Mar 4 at 17:11
  • $\begingroup$ Any regularity assumption on $X$ (moments, analyticity of $\phi_X$, ...) ? $\endgroup$ – Gabriel Romon Mar 8 at 20:56
  • $\begingroup$ No, nothing is assumed on the moments, analyticity of $\phi_X.$ (Unless we succeed to prove them) $\endgroup$ – Kurt.W.X Mar 8 at 21:26

Let $Y$ be a random variable with characteristic function $\phi_Y:t\mapsto e^{-\sigma^2t^2}$, i.e. $Y$ is Gaussian (possibly degenerate).

The hypothesis in the problem states that $\phi_X$ and $\phi_Y$ coincide along a sequence $(x_n)$ of non-zero reals such that $x_n\to 0$. If $\phi_X$ were analytic in a neighborhood of $0$, so would be $\phi_X - \phi_Y$ and since the zeroes of an analytic function are isolated, this would imply $\phi_X - \phi_Y = 0$, as wanted.

Instead of showing directly that $\phi_X$ is analytic, we show that $X$ has moments of all orders. This will imply that $\phi_X$ is infinitely differentiable. We show next that $\forall n \geq 0$, $\phi_X^{(n)}(0)=\phi_Y^{(n)}(0)$. Since $Y$ is uniquely determined by its moments (see here or here), this will yield $\phi_X=\phi_Y$.

For $n\geq 1$, let $\displaystyle f_n:t\mapsto \frac{2-e^{ix_n t}-e^{-ix_n t}}{x_n^2}$ and note that it is real-valued, non-negative and by a Taylor expansion $\forall t\in \mathbb R, f_n(t) \to t^2.$ By Fatou's lemma, $$\begin{align} E(X^2) &\leq \liminf_n E(f_n(X)) \\ &=\liminf_n \frac{2-\phi_X(x_n)-\phi_X(-x_n)}{x_n^2} \\ &=\liminf_n \frac{2-\phi_X(x_n)-\overline{\phi_X(x_n)}}{x_n^2} \\ &=\liminf_n \frac{2-\phi_Y(x_n)-\phi_Y(-x_n)}{x_n^2} \\ &= -\phi_Y''(0) \end{align} $$ where the last equality follows from a Taylor expansion of $\phi_Y$. Hence $X$ has a second moment, and by a classical result $\phi_X$ is twice-differentiable over $\mathbb R$.

Since $\phi_Y$ is even we may assume WLOG that all the $x_n$ are positive, and also that $(x_n)$ is strictly decreasing (by taking a subsequence if needed). By Rolle's theorem applied to $\phi_X-\phi_Y$, there is some sequence $(y_n)$, strictly decreasing, with $y_n\to 0$ and $\forall n\geq 1, \phi_X'(y_n) = \phi_Y'(y_n)$. Another application of Rolle's provides some sequence $(z_n)$, strictly decreasing, with $z_n\to 0$ and $\forall n\geq 1, \phi_X''(z_n) = \phi_Y''(z_n)$.

Defining a new $f_n$ and performing a similar reasoning, we prove successively that $E(X^4)<\infty$, $\phi_X^{(3)}$ and $\phi_Y^{(3)}$ coincide along a sequence that goes to $0$ and similarly for $\phi_X^{(4)}$ and $\phi_Y^{(4)}$. This continues indefinitely. $\phi_X$ is therefore $C^\infty$, each $\phi_X^{(m)}$ is continuous and taking limits along the aforementioned sequences we find that $\forall m\geq 1$, $\phi_X^{(m)}(0)=\phi_Y^{(m)}(0)$.

Note that the proof follows through if $Y$ is replaced by any random variable that is uniquely determined by its moments. For example, it is sufficient that $\limsup_n \frac{E(|Y|^n)^{1/n}}{n} <\infty$.


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.