Given an absolutely integrable function $f:\mathbb R^n \to \mathbb R$, let $R[f]$ be its Radon transform defined for every $(w,b) \in (\mathbb R^n \setminus \{0\}) \times \mathbb R$ by

$$ R[f](w,b) := \int_H f(x)ds(x) = \int_{\mathbb R^n}f(x)\delta(b-x^\top w)\,dx, $$ where $\delta$ is the Dirac distribution and $ds$ is area element on the hyperplane $$ H:=\{x \in \mathbb R^n \mid x^\top w = b\}. $$ In particular, if $P$ is a probability distribution on $\mathbb R^n$ with density $f$ and $z \sim P$, then we can inteprete $R[f](w,b)$ as the density of the random variable $z^\top w$ evaluated at the point $b$.

Now, let $g:\mathbb R^n \to \mathbb R$ be the density of the multivariate Gaussian distribution $N(\mu,\Sigma)$ with mean $\mu \in \mathbb R^n$ and covariance matrix $\Sigma \in \mathbb R^{n \times n}$. Since $z^\top w \sim N(\mu^\top w,w^\top \Sigma w)$ for $z \sim N(\mu,\Sigma)$, it is clear by virtue of the previous remark that, $$ R[g](w,b) = \varphi\left(\frac{b-\mu^\top w}{\|w\|_\Sigma}\right), \tag{1} $$ where $\varphi$ is the density of the standard Gaussian distribution $N(0,1)$, and $\|w\|_\Sigma := (w^\top \Sigma w)^{1/2}$.

Consider the function $h:\mathbb R^n \to \mathbb R$ defined by $h(x)=x_1 g(x)$ for every $x=(x_1,\ldots,x_n) \in \mathbb R^n$.

Question. In the spirit of (1), what is an analytic formula for $R[h](w,b)$ ?

In the special case where $d=1$ so that $\Sigma=\sigma>0$ is just a scalar, a simple computation gives

$$ R[h](w,b) = \int_{-\infty}^{+\infty} x\varphi(x)\delta(b-wx)\,dx = \frac{b}{w}\varphi(\frac{b-w\mu}{2\sigma w}) = \frac{b}{w}R[g](w,b). $$


The accepted answer shows that

$$ R[h](w,b) = (\alpha(b-\mu_2)+\mu_1)R[g](w,b), $$

where $\mu_1 := e_1^\top \mu$, $e_1=(1,0,\dots,0) \in \mathbb R^n$, $\mu_2 := w^\top \mu$, $\alpha := \dfrac{e_1^\top \Sigma w}{w^\top \Sigma w}$.

Of course, this can be generalized by replacing $e_1$ with any unit-vector $v$.


1 Answer 1


$\newcommand{\si}{\sigma}\newcommand{\Si}{\Sigma}\newcommand{\ep}{\varepsilon}\newcommand{\vpi}{\varphi}\newcommand{\R}{\mathbb R}$In this "Gaussian" setting especially, it is convenient to approximate the delta function by the normal distribution $N(0,\ep^2)$ with $\ep\downarrow0$, so that
\begin{equation*} R[f](w,b)=\lim_{\ep\downarrow0}R_\ep[f](w,b), \tag{1}\label{1} \end{equation*} where \begin{equation*} R_\ep[f](w,b):=\int_{\R^n}dx\,\vpi_\ep(w^\top x-b)f(x), \end{equation*} \begin{equation*} \vpi_\ep(t):=\frac1\ep \vpi\Big(\frac t\ep\Big), \end{equation*} and $\vpi$ is the standard normal density.

Now, for $h(x)\equiv x_1 g(x)$ and $g$ the density of $N(\mu,\Si)$, we can write \begin{equation*} R_\ep[h](w,b) =\int_{\R^n}dx\,g(x) x_1 \vpi_\ep(w^\top x-b) =E e_1^\top X\, \vpi_\ep(w^\top X-b), \end{equation*} where $e_1:=[1,0,\dots,0]^\top\in\R^{n\times1}=\R^n$ and $X\sim N(\mu,\Si)$ .

Note that the joint distribution of $e_1^\top X$ and $w^\top X$ is bivariate normal with respective means \begin{equation*} \mu_1:=e_1^\top\mu\quad\text{and}\quad\mu_2:=w^\top\mu, \tag{2}\label{2} \end{equation*} respective standard deviations \begin{equation*} \si_1:=\sqrt{e_1^\top\Si e_1} \quad\text{and}\quad \si_2:=\sqrt{w^\top\Si w}, \tag{3}\label{3} \end{equation*} and correlation \begin{equation*} \rho:=\frac{e_1^\top\Si w}{\si_1\si_2}. \tag{4}\label{4} \end{equation*}

So, straightforward calculations yield \begin{equation*} R_\ep[h](w,b) = \frac{ \rho \si _1 \si _2 (b-\mu _2)+\mu _1 (\si _2^2+\ep ^2)}{\sqrt{2 \pi } (\si _2^2+\ep ^2){}^{3/2}}\, \exp\Big\{-\frac{(b-\mu _2){}^2}{2 (\si _2^2+\ep ^2)}\Big\}. \end{equation*} Finally, by \eqref{1}, \begin{equation*} R[h](w,b) = \frac{\rho\si_1(b-\mu_2)+\mu_1\si_2}{\sqrt{2\pi}\,\si _2^2}\, \exp\Big\{-\frac{(b-\mu_2)^2}{2 \si _2^2}\Big\}, \end{equation*} with $\mu_1,\mu_2,\si_1,\si_2,\rho$ given by \eqref{2}--\eqref{4}.

  • $\begingroup$ I always learn something new from your answers. Thanks for the generocity. $\endgroup$
    – dohmatob
    Mar 25 at 13:59
  • $\begingroup$ @dohmatob : You are very welcome. $\endgroup$ Mar 25 at 14:23

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.