Sampling only from the uniform distribution $U(0,1)$, I am hoping to use transformations to create random values distributed uniformly around the perimeter of an ellipse. Eventually, I'd like to do the same on the surfaces of ellipsoids and other problematic objects.

My first idea was as follows. We can easily get $\Theta \sim U(0,2\pi)$. Then, from the parametric form of the ellipse,

$$X \equiv a \cos \Theta \\ Y \equiv b \sin \Theta $$

is a random point on the ellipse's perimeter.

Similarly, if we independently sample another angle $\Phi \sim U(0,\pi)$, we could use

$$X \equiv a \sin \Theta \cos\Phi \\ Y \equiv b \sin \Theta \sin\Phi\\ Z \equiv c \cos \Theta $$

The problem with these approaches is that they are uniformly distributed with respect to theta, not along the surface. They are equivalent to taking a uniform distribution on a circle and then projecting about the radius to the perimeter of the ellipse, so the density of points is higher near the major axis, as you can see here:

enter image description here

(This is itself counterintuitive to me: One would expect the points to be denser about the minor axis since they are being "sprayed" over a more concentrated region, right?)

How can I generate points distributed uniformly about the perimeter of the ellipse?

Here is an example of what I am trying to do but using a circle instead. The transformation used there doesn't work for the ellipse because it creates the same bunching behavior.

  • 1
    $\begingroup$ You want $L(\theta)=p\cdot L(2\pi)$ where $L(x)=\int\limits_0^x\sqrt{a^2\cos^2\theta+b^2\sin^2\theta}\,\mathrm{d}\theta$ for uniformly-distributed $p\in[0,1]$? $\endgroup$ – Alexey Burdin Jun 8 '20 at 2:22
  • $\begingroup$ I don't think that's what I'm looking for. That seems like a definition of $U(0,P)$ where P is the perimeter of the ellipse. I am looking for random points $(x,y)$ that satisfy the ellipse equation $x^2 / a^2 + y^2 / b^2 = 1$. $\endgroup$ – Max Jun 8 '20 at 2:28
  • 2
    $\begingroup$ Regarding the "counterintuitive" part: it is not generating uniformly from a circle and "projecting along the radius to the perimeter of the ellipse," rather it is generating uniformly from a circle and stretching the circle to the shape of an ellipse, so the points near the minor axis are stretched out and become less dense. $\endgroup$ – angryavian Jun 8 '20 at 2:48
  • 2
    $\begingroup$ Rephrasing Alexey's suggestion: pick a number $p \sim U(0, P)$ (where $P$ is the perimeter of the ellipse). Then choose the unique point $(x,y)$ on the ellipse such that the (counter-clockwise) arc length from $(a,0)$ to this point $(x,y)$ is $p$. This calculation involves elliptic integrals so you will need to approximate it numerically. $\endgroup$ – angryavian Jun 8 '20 at 2:53
  • 2
    $\begingroup$ Related to post 1 and post 2. $\endgroup$ – Ng Chung Tak Jun 8 '20 at 8:01

The following may be of some use in this question. (Note: Some of these points have also been sketched out in the above comments-included here for completeness). In particular, the below code computes the transformation based on the following derivation:

The points on the ellipse are assumed to have the coordinates defined by $$ x=a\cos{\theta} \\ y=b\sin{\theta} \\ $$

The arclength differential $\mathrm{d}s$ along the perimeter of the ellipse is obtained from

$$ {\mathrm{d}s}^{2}={\mathrm{d}x}^{2}+{\mathrm{d}y}^{2} $$

$$ {\mathrm{d}s}^{2}=a^{2}\sin^{2}{\theta}{\mathrm{d}\theta}^{2}+b^{2}\cos^{2}{\theta}{\mathrm{d}\theta}^{2} $$

$$ {\mathrm{d}s}^{2}=\left(a^{2}\sin^{2}{\theta}+b^{2}\cos^{2}{\theta}\right){\mathrm{d}\theta}^{2} $$

$$ {\mathrm{d}s}=\sqrt{a^{2}\sin^{2}{\theta}+b^{2}\cos^{2}{\theta}}{\mathrm{d}\theta} $$

$$ \frac{{\mathrm{d}s}}{\mathrm{d}\theta}=\sqrt{a^{2}\sin^{2}{\theta}+b^{2}\cos^{2}{\theta}} $$

Now, the probability function is taken to be

$$ p\left(\theta\right)=\frac{{\mathrm{d}s}}{\mathrm{d}\theta} $$

with the interpretation that when the rate-of-change of arclength increases, we want a higher probability of sample points in that interval to keep the density of points uniform.

We can then set up the following expression:

$$ p\left(\theta\right){\mathrm{d}\theta}=p\left(x\right){\mathrm{d}x} $$

and assuming a uniform distribution for $x$:

$$ \int p\left(\theta\right){\mathrm{d}\theta}=x+K $$.

Some plots of the uncorrected and corrected ellipses are shown in the below figure, using the above derivation and code implementation below. I hope this helps.

enter image description here

Python code below:

import math
import matplotlib.pyplot as plt

# ellipse major (a) and minor (b) axis parameters

# num points for transformation lookup function
npoints = 1000


# integrated probability density

for iTheta in range(1,npoints+1):
    # ds/d(theta):
    delta_s_val=math.sqrt(a**2*math.sin(iTheta*delta_theta)**2+ \

    # do integral
    integ_delta_s_val = integ_delta_s_val+delta_s_val*delta_theta
# normalize integrated ds/d(theta) to make into a scaled CDF (scaled to 2*pi)
integ_delta_s_norm = []
for iEntry in integ_delta_s:
#print('theta= ', theta)
#print('delta_theta = ', delta_theta)
#print('delta_s= ', delta_s)
#print('integ_delta_s= ', integ_delta_s)
#print('integ_delta_s_norm= ', integ_delta_s_norm)

# Plot tranformation function
plt.xlim(-0.2, x_axis_range)
plt.ylim(-0.2, y_axis_range)
# overplot reference line which are the theta values.

# Reference ellipse without correction.

# Create corrected ellipse using lookup function


for theta_index in range(npoints_new):
    theta_val = theta_index*delta_theta_new
#    print('theta_val = ', theta_val)
# Do lookup:
    for lookup_index in range(len(integ_delta_s_norm)):
#        print('doing lookup: ', lookup_index)
#        print('integ_delta_s_norm[lookup_index]= ', integ_delta_s_norm[lookup_index])
        if theta_val >= integ_delta_s_norm[lookup_index] and theta_val < integ_delta_s_norm[lookup_index+1]:
#            print('value found in lookup table')
#            print('theta_prime = ', theta_prime)
#            print('---')
    # ellipse without transformation applied for reference
    # ellipse with transformation applied

# Plot reference and transformed ellipses  
plt.xlim(-x_axis_range, x_axis_range)
plt.ylim(-y_axis_range, y_axis_range)
plt.gca().set_aspect('equal', adjustable='box')
plt.plot(ellip_x, ellip_y, '.')

plt.xlim(-x_axis_range, x_axis_range)
plt.ylim(-y_axis_range, y_axis_range)
plt.gca().set_aspect('equal', adjustable='box')
plt.plot(ellip_x_prime, ellip_y_prime, '.')
  • $\begingroup$ Great! I've been working on my own Python script which works much the same way: codereview.stackexchange.com/questions/243590/… Using interpolation instead of lookups, you can pick the arc length on $(0,P)$ and map it to a $\theta$-value. $\endgroup$ – Max Jun 14 '20 at 0:20

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.