Let $f(x) = \log(\cosh(x))$, and define the kernel density:

$$p_r(\phi;\theta) = \Big(f\big(r\cos(\phi-\theta)\big) - f\big(r\cos(\phi+\theta)\big) \Big)\hspace{0.5pt} \frac{\sin(2\phi)}{\sin(2\theta)},\quad 0\le \phi,\theta \le \mbox{$\large\frac{\pi}{2}$},\;\,r>0$$

Show that for all $r>0$, $p_r(\phi;\theta)$ has a decreasing likelihood ratio. I.e., for $0\le\theta_1 < \theta_2\le\pi/2$, the function:

$$ h(\phi) = \frac{f\big(r\cos(\phi-\theta_2)\big) - f\big(r\cos(\phi+\theta_2)\big) }{f\big(r\cos(\phi-\theta_1)\big) - f\big(r\cos(\phi+\theta_1)\big)}$$

is decreasing on $[0,\pi/2]$, and:

$$ \frac{\partial}{\partial \phi} \frac{\partial}{\partial \theta}\hspace{2pt} \log p_r(\phi;\theta) \le 0,\quad 0 \le \phi,\theta\le \mbox{$\large\frac{\pi}{2}$}$$

Taking the derivatives, defining $x=r cos(\phi-\theta)$ and $y=r \cos(\phi+\theta)$, and collecting terms in $r^2$, we find that the inequality holds if the following two inequalities hold for $x>y>0$:

$$f(x)-f(y) \;\ge\; \frac{\tanh^2(x) - \tanh^2(y)}{2-\tanh^2(x)-\tanh^2(y)} $$
and:

$$f(x)-f(y)\; \le \;\frac{x^2\tanh^2(x) - y^2\tanh^2(y)}{x\tanh(x) + y\tanh(y) + x^2(1-\tanh^2(x)) + y^2(1-\tanh^2(y))}  $$

Subtracting to get two non-negative functions of $x,y$, we get functions that are monotonic in $x$ and $y$. Taking the derivative of the first with respect to $x$ and simplifying gives a simple obviously positive expression. However, the derivative of the second function does not simplify in an obvious way, and is a complicated function of powers of $x,y,\tanh(x)$ and $\tanh(y)$.

This result is important to prove uniqueness of stable optima in the unmixing and deconvolution of linear mixtures of independent random variables with strongly sub- and super-gaussian densities, using variation diminishing property of MLR densities.