Let $\sigma>0$, $$\phi(x):=\frac1{\sqrt{2\pi\sigma^2}}e^{-\frac{x^2}{2\sigma^2}}\;\;\;\text{for }x\in\mathbb R$$ and $$\mathcal N(x,B):=\int_B\varphi(y-x)\:{\rm d}y\;\;\;\text{for }(x,B)\in\mathbb R\times\mathcal B(\mathbb R).$$ If $$\iota:\mathbb R\to[0,\infty)\;,\;\;\;x\mapsto x-\lfloor x\rfloor,$$ then the wrapped normal distribution kernel on $[0,1)$ with variance $\sigma^2$ is defined by $$\mathcal W(x,\;\cdot\;):=\mathcal N(x,\;\cdot\;)\circ\iota^{-1}\;\;\;\text{for }x\in[0,1).$$ If $$\psi(x):=\sum_{k\in\mathbb Z}\phi(x+k)\;\;\;\text{for }x\in\mathbb R,$$ we can show that $$\mathcal W(x,B)=\int_B\psi(y-x)\:{\rm d}y\;\;\;\text{for all }(x,B)\in[0,1)\times\mathcal B([0,1))\tag1.$$

Let $x\in[0,1)$. How can we compute the expectation of a random variable $Y\sim\mathcal W(x,\;\cdot\;)$?

Clearly, $$\operatorname E[Y]=\int_0^1\psi(y-x)y\:{\rm d}y\tag2.$$


The wrapped Normal distribution can be written as an infinite mixture of truncated Normal distributions since, if $\phi_\sigma$ and $\Phi_\sigma$ denote respectively the density and cdf of the Normal $\mathcal N(0,\sigma^2)$ distribution, then $$\psi(y-x)=\sum_{k\in\mathbb Z}\phi_\sigma(y-x-k)\mathbb I_{(0,1)}(y)$$ can be rewritten as \begin{align}\psi(y-x)&=\sum_{k\in\mathbb Z}[\Phi_\sigma(1-x-k)-\Phi_\sigma(-x-k)]\times\dfrac{\phi_\sigma(y-x-k)\mathbb I_{(0,1)}(y)}{\Phi_\sigma(1-x-k)-\Phi_\sigma(-x-k)}\\ &=\sum_{k\in\mathbb Z}\underbrace{[\Phi_\sigma(x+k)-\Phi_\sigma(x+k-1)]}_{p_k(x)}\times\dfrac{\phi_\sigma(y-x-k)\mathbb I_{(0,1)}(y)}{\Phi_\sigma(x+k)-\Phi_\sigma(x+k-1)}\end{align} Therefore its mean is the weighted sum of the means of the truncated Normal distributions $\mathcal N_0^1(x+k,\sigma^2)$, given by $$x+k-\sigma^2\dfrac{\phi_\sigma(1-x-k)-\phi_\sigma(-x-k)}{\Phi_\sigma(x+k)-\Phi_\sigma(x+k-1)}$$ That is, $$\mathbb E_x[Y]=x+\sum_{k\in\mathbb Z} kp_k(x)-\sigma^2 \sum_{k\in\mathbb Z}[\phi_\sigma(1-x-k)-\phi_\sigma(-x-k)]=x+\sum_{k\in\mathbb Z} kp_k(x)$$ since the last sum is equal to zero. Except when $x=0,1/2,1$, where the expectation is $1/2$, I do not know whether or not the mean has a closed form expression. A vague simplification using the symmetry properties of the Normal cdf is $$\sum_{k\in\mathbb Z} kp_k(x)=\Phi(-x)+\sum_{k=1}^\infty[\Phi(x-k)+\Phi(-x-k)]$$

Here is the Monte Carlo approximation to the mean when $\sigma=.1$ (although there is no visible difference in the graph when moving $\sigma$):

enter image description here

| cite | improve this answer | |
  • $\begingroup$ I see. However, the relation to the truncated normal distribution might be useful. I've got a related question, you might be able to answer as well: math.stackexchange.com/q/3594286/47771. $\endgroup$ – 0xbadf00d Mar 25 at 10:41

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.