Non Gaussian Additive Channel Capacity If I have a additive, stationary, memoryless non-Gaussian noise channel
$$
Y_i = X_i + Z_i
$$
With the fixed mean and covariance on the noise $Z_i$
$$
\mathbb{E}(Z_i) = 0,\ Var(Z_i) = 1
$$
and a covariance constraint on $X_i$
$$
\frac{1}{n}\sum x_i^2 \leq P.
$$
How can I prove the following bound on the channel capacity
$$
\frac{1}{2}\log(1+P) \leq C(P)\leq \frac{1}{2}\log(1+P) + D(P_z||\mathcal{N}(0,1))? 
$$
I think I can get the lower bound as follows. Letting $X_g \sim \mathcal{N}(0,P)$
\begin{align}
C(P) =& \max_{P_x : Var(X) \leq P} I(X;Y)\\
=& \max_{P_x : Var(X) \leq P} I(X;X+Z)\\
\geq& I(X_g; X_g + Z)\\
\geq& \min_{P_z} I(X_g;X_g+Z)\\
=& I(X_g;X_g+Z^*)\\
=& \frac{1}{2}\log(1-P)
\end{align}
Because the noise distribution that minimizes mutual information is Gaussian $P_{Z^*} = \mathcal{N}(0,1)$.
But I'm having trouble with the upper bound.
 A: I am not sure I follow the proof for the lower bound. You can obtain a lower bound by setting $X$ to be Gaussian (which, is not optimal, in general), and then note that the worst distribution for $Z$ when $X$ is Gaussian is Gaussian as well! Therefore, the capacity is always lower bound by the capacity of the "standard" Gaussian input - Gaussian noise channel.
To prove the upper bound, start by writing 
$$
\begin{align}
I(X;Y)&= h(Y)-h(Y|X)\\
\tag{1}
&=h(Y)-h(Z),
\end{align}
$$
assuming that $X$ and $Z$ are independent. Now, the entropy of any variable with a given variance is upper bounded by the entropy of a Gaussian variable with the same variance, therefore, 
$$
\tag{2}
h(Y)\leq \frac{1}{2}\ln\left((1+P) \sqrt{2 \pi e} \right).
$$
In order to find a lower bound for $h(Z)$, let $p_Z(z)$ denote the probability distribution (pdf) of $Z$ and $p_G(z)\triangleq \frac{1}{\sqrt{2 \pi}} e^{-z^2/2}$. It holds 
$$
\begin{align}
D(p_Z\|p_G)&\triangleq \int p_Z(z) \ln\left(\frac{p_Z(z)}{p_G(z)} \right)dz\\
&=-h(Z)-\int p_Z(z) \ln\left(p_G(z)\right)dz\\
&\stackrel{(a)}{=} -h(Z)-\int p_G(z) \ln\left(p_G(z)\right)dz\\
&=-h(Z)+\frac{1}{2}\ln\left(\sqrt{2 \pi e} \right),
\end{align}
$$
where $(a)$ is due to $\ln p_G(z)$ being a quadratic function of $z$ and $p_G$,  $p_Z$ having the same variance (thanks to @user187815 for noting this!). Therefore,
$$
\tag{3}
h(Z) =  \frac{1}{2}\ln\left(\sqrt{2 \pi e} \right)- D(p_Z\|p_G) 
$$
Using (2) and (3) in (1), the upper bound follows.  
