Let $A$ be a parametric family of probability distributions that include all distributions in the form of $\phi(X)$ where $X\sim\mathcal{N}(0,\mathbf{I})$ is jointly Gaussian and $\phi:\mathbb{R}^d\to \mathbb{R}^d$ belongs to a parametric set of functions $G$.

To uniquely determine a distribution in $A$, how many moments do we need to know? For example, if $G$ is linear, A includes all Gaussian distributions. Thus, we need first and second moments to uniquely determine a distribution in the set $A$. I wonder if such a result can be extended for other $G$ such as polynomials with degree $k$.

  • $\begingroup$ in the multidimensional case many polynomials will give the same distribution so you cannot determine the polynomial from moments, e.g., $(x_1x_2, x_1^2)$ and $(x_1x_2, x_2^2)$ $\endgroup$ – user83457 Jan 18 '18 at 7:51
  • $\begingroup$ @michael Even in the 1-dimensional case: there you cannot distinguish $-x$ from $x$ $\endgroup$ – Bjørn Kjos-Hanssen Jan 18 '18 at 8:00
  • $\begingroup$ Thanks guys. Just to clarify, the goal is not to uniquely determine the $\phi$ functions, but to uniquely determine the distribution. I.e., if $m_k(P_1)=m_k(P_2)$ for 1\leq k\leq K, we can conclude that $P_1=P_2$. $\endgroup$ – Soheil Feizi Jan 18 '18 at 17:30
  • $\begingroup$ Are you also interested in knowing how many statistics you need to know (not necessarily moments)? This would probably be a simpler question. $\endgroup$ – S.Surace Jan 19 '18 at 23:51
  • $\begingroup$ yes, that'd be helpful too. $\endgroup$ – Soheil Feizi Jan 21 '18 at 1:20

Looks complicated in general, but here's a derivation that for $d=1$ and $k=2$, in the slightly further restricted case where $\phi(X)=aX^2+b$, three moments suffice.

Let $m_i=E(\phi(X)^k)$. Recall $E(X^4)=3$ and $E(X^6)=15$. We calculate $$m_1=a+b$$ $$m_2=3a^2+2ab+b^2$$ with solution: $$m_2=3a^2+2a(m_1-a)+(m_1-a)^2$$ $$m_2=a^2+2am_1+m_1^2-2am_1 +a^2$$ $$a^2=\frac12(m_2-m_1^2) =: u$$ So far we don't know $a$ exactly. But $$m_3=15a^3+9a^2b+3ab^2+b^3$$ so we can use our value for $a^2$ to reduce $$m_3=15a^3+9a^2(m_1-a)+3a(m_1-a)^2+(m_1-a)^3$$ to a linear equation in $a$: $$m_3=15ua+9um_1-9ua+3am_1^2-6um_1+3ua+m_1^3-3am_1^2+3um_1-ua$$ $$m_3=8ua+6um_1+m_1^3$$ and we get the somewhat appealing solution

$$\begin{eqnarray*}a&=&\frac14\frac{m_3-m_1^3}{m_2-m_1^2}-\frac34m_1\\ b&=&m_1-a\end{eqnarray*}$$

| cite | improve this answer | |
  • $\begingroup$ Thanks. yes I did this calculation as well. Any ideas for a more general case? $\endgroup$ – Soheil Feizi Jan 18 '18 at 17:32
  • 1
    $\begingroup$ @SoheilFeizi yes I suppose you could work with a formula for $E(X^k)$ as a function of $k$, and the multinomial theorem $\endgroup$ – Bjørn Kjos-Hanssen Jan 19 '18 at 2:59

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.