Flat-top Gaussian distributions I want to simulate microscope images (2D gray-scale pixel arrays) of fluorescent beads, which are tiny balls of polymer doped with fluorescent dye. I assume that an ideal image (neglecting out-of-focus contributions/depth information) would closely resemble a rectangle function - so the 2D intensity profile would be a rectangle, the 3D profile a cylindrical flat-top distribution. But due to the resolution-limited optics, the true image will be blurred.
For most practical cases, this blurring can be approximated by a Gaussian blur, i.e. by convolution with a Gaussian kernel that has a width equivalent to the Airy disk. Since the rectangle function is discontinuous, I was looking for a smooth alternative which approximates the blurred rectangle function and found the super-Gaussian or higher-order Gaussian:
$\displaystyle f{{\left({x}\right)}}={{f}_{{0}}{e}^{{-{\left(\frac{x}{{x}_{{0}}}\right)}^{{{2}{n}}}}}}$
However, there is one feature that is missing: Since I want to simulate fluorescent beads of different sizes, the model function should be scalable in width while keeping the slopes constant. The super-Gaussian will spread out with increasing width, which would corresponds to an unphysical increase in blur with larger bead sizes.
The question therefore is: How to modify the super-Gaussian so the slopes are constant for a wide range of width $\displaystyle{x}_{{0}}$? If possible, the modification should be as simple as possible in order to facilitate further image processing computations (e.g. further convolution operations; my laptop is kinda slow...).
 A: It is important to understand what each of the free parameters does to $f$.  Specifically, $f_0$ scales the height of $f$; the maximum value attained corresponds to $(x,y) = (0, f_0)$.  The parameter $x_0$ is a scaling factor in the $x$-direction.  And for $n \ge 1$, the larger the value, the sharper the transition from $f_0$ to $0$ as $x$ increases.
What we would like to do is compute the slope of $f$ at its inflection point as a function of these parameters; that is to say, $f'(x_c)$ where $x_c > 0$ satisfies $f''(x_c) = 0$.  This corresponds to a extremum for $f'$; i.e., where the slope is steepest.  Logarithmic differentiation yields
$$\frac{d}{dx} \left[\log f(x)\right] = \frac{f'(x)}{f(x)} = -\frac{2n}{x_0} \left(\frac{x}{x_0}\right)^{2n-1} = g(x),$$ hence $$f''(x) = f'(x) g(x) + f(x) g'(x) = f(x) (g^2(x) + g'(x)).$$  Since $f > 0$ for all $x$, we need only consider the nontrivial critical points satisfying $0 = g^2 + g'$, which leads to the equation $$0 = 1 + 2n\left(\left(\frac{x}{x_0}\right)^{2n} - 1\right),$$ or $$x_c = x_0 \left(1 - \frac{1}{2n}\right)^{1/(2n)}.$$  Consequently, $$f'(x_c) = f(x_c)g(x_c) = -\frac{2f_0}{x_0} n e^{-1 + 1/(2n)} \left(1 - \frac{1}{2n}\right)^{1 - 1/(2n)} = -\frac{f_0}{f_c} (2n-1) e^{-1 + 1/(2n)}$$  If we regard $f_0$ as fixed, this establishes a relationship between $x_0$ and $n$ for which we have a desired slope at the nontrivial point of inflection of the curve.
For instance, if we want the peak to be $f_0 = 1$, and we originally have $n = 3$ and $x_0 = 1$, but we want to widen this curve so that we have the same slope when $x_0 = 2$, then we can numerically solve for $n$ that does this, yielding $n \approx 6.06782$.  In general, the choice $n_{\text{new}} \approx x_0 n$ seems to be a good approximation for $x_0 > 1$.
A: Complementing heropup's very nice solution, I just wanted to add that the convolution of a rectangle with a Gaussian is actually a known function. It's just the difference of two Error functions:

Solution on Wolfram Alpha
