What Is the Difference Between the Rician and Gaussian Distributions?

The key difference is what each distribution models. A Gaussian (normal) distribution describes a real-valued quantity that can lie anywhere on the real line. A Rician distribution describes a magnitude: it is the length of a two-dimensional Gaussian vector, so it cannot be negative.

This connection explains why a Rician model appears in applications such as radio signal amplitudes. The signal’s two components can each contain Gaussian noise, even though the measured amplitude itself is not Gaussian.

How the Rician distribution arises

Let $X$ and $Y$ be independent Gaussian random variables with the same variance. Their means can be viewed as coordinates of a vector with length $\nu$:

$$X \sim N(\nu,\sigma^2), \qquad Y \sim N(0,\sigma^2).$$

The magnitude

$$R=\sqrt{X^2+Y^2}$$

has a Rician distribution with parameters $\nu$ and $\sigma$. The direction of the mean vector does not affect the distribution of the magnitude, so it is convenient to place that mean along the $X$-axis as above. Here, $\nu\geq 0$ is the length of the mean vector, and $\sigma$ is the standard deviation of either Gaussian component.

For $r\geq 0$, its probability density function is

$$f_R(r)=\frac{r}{\sigma^2}\exp\left(-\frac{r^2+\nu^2}{2\sigma^2}\right) I_0\left(\frac{r\nu}{\sigma^2}\right),$$

and $f_R(r)=0$ for $r<0$. The function $I_0$ is the modified Bessel function of the first kind of order zero. It appears because the joint Gaussian density is being converted from Cartesian coordinates to a radial magnitude.

Main differences at a glance

Feature	Gaussian variable	Rician magnitude
Values it can take	Any real number	Nonnegative numbers only
Typical construction	One real-valued measurement	Length of a two-component Gaussian vector
Parameters in this model	Mean and standard deviation	Mean-vector length $\nu$ and component standard deviation $\sigma$
Shape	Symmetric about its mean	Generally asymmetric; shape depends on $\nu/\sigma$

A Gaussian distribution may describe either component, $X$ or $Y$, in the construction. It does not generally describe their magnitude $R$.

Special cases and useful relationships

When $\nu=0$, both components are centered at zero. The Rician distribution then reduces to the Rayleigh distribution, with density

$$f_R(r)=\frac{r}{\sigma^2}\exp\left(-\frac{r^2}{2\sigma^2}\right), \qquad r\geq 0.$$

More generally, the Rician distribution is a noncentral chi distribution with two degrees of freedom. This provides a useful link to the chi and chi-squared distributions.

The squared magnitude also has a simple expectation:

$$\mathbb{E}[R^2]=\mathbb{E}[X^2+Y^2]=\nu^2+2\sigma^2.$$

This is the mean square magnitude, not the square of the mean magnitude: in general, $\mathbb{E}[R^2]\ne (\mathbb{E}[R])^2$.

Worked example

Suppose $X\sim N(3,1)$ and $Y\sim N(0,1)$ independently, and measure $R=\sqrt{X^2+Y^2}$. The component standard deviation is $\sigma=1$, and the mean vector has length $\nu=3$. Therefore, $R$ is Rician with parameters $(3,1)$.

At magnitude $r=3$, its density is

$$\begin{aligned} f_R(3) &=3\exp\left(-\frac{3^2+3^2}{2}\right)I_0(9)\\ &=3e^{-9}I_0(9)\\ &\approx 0.405. \end{aligned}$$

The density is zero at $r=0$ for these parameters, and it is zero for every negative value. By contrast, the Gaussian component $X$ can be negative: $P(X<0)=P(Z<-3)\approx 0.00135$ for a standard normal variable $Z$. These are different random variables, so their probability distributions need not match.

Also, the mean square magnitude is $\mathbb{E}[R^2]=3^2+2(1)^2=11$, giving a root-mean-square magnitude of $\sqrt{11}\approx 3.32$.

When can a Rician variable be approximated by a Gaussian?

If $\nu$ is much larger than $\sigma$, most of the vector’s length comes from its nonzero mean rather than the noise. In that regime, the magnitude is often approximately normal, centered near $\nu$ with spread near $\sigma$. The approximation is less suitable when $\nu$ is comparable to $\sigma$, and a Gaussian approximation can still assign probability to negative values even though a magnitude cannot be negative.

Common mistakes

• Calling the magnitude Gaussian because its components are Gaussian. A nonlinear transformation such as $\sqrt{X^2+Y^2}$ changes the distribution.
• Treating a Rician variable as able to be negative. It represents a length, so its support is $[0,\infty)$.
• Assuming every pair of Gaussian components gives a Rician magnitude. The standard Rician model requires independent components with equal variance. Unequal variances or dependence generally lead to a different model.
• Mixing parameter conventions. Here, $\sigma$ is the standard deviation of each component and $\nu$ is the mean-vector length. Check definitions when using a source or software package, since notation and parameterizations can vary.