
Hi,
I was just playing around with the RadianceOfXAndYAndZAndThetaAndPhiDetectorInput and I (only now) noticed that the results of the MC simulations are exported as mean and second moments (and a derived standard deviation). Assuming that these are the means over
all runs, how should I interpret the standard deviations?
Am I correct that the variance should approach the mean as we are dealing with a Poisson distribution? Or does that not hold because of the weights associated with the photons along the trajectories?
Thanks,
Martijn



Hi Martijn,
The calculated mean is determined by mean = E[xi] = (1/N) sum_{i=1}^{N} xi(beta_i) where E is expectation, xi is the random variable or estimator, beta_i is the ith random walk, and N is the number of photons launched in the simulation. The Nsample variance
is var[xi] = (1/N) {E[xi^2](E[xi])^2} where E[xi^2] is the second moment estimate. The standard deviation sigma = sqrt(var[xi]). This sigma provides the 1sigma confidence interval about the mean assumed in a normal or Gaussian distribution. So the variance
goes to zero as O(1/N) and error as O(1/sqrt(N)).
I hope this helps.
Carole



Hi Carole,
Thanks for the explanation. I still do not completely get why the variance should go to zero, given the inherent variability between random walks, but hopefully it becomes clear when I played with it a bit more.
Thanks!
Martijn



The way I think about it, qualitatively, is that the variance and the standard deviation are quantities to understand the probability of the next MEAN value (expectation value of solution, averaging ALL photons' trajectories) to be similar to the one calculated
as part of this solution. So, more photons (or a better algorithm) allows you to reduce that uncertainty, down to zero at the limit.

David J. Cuccia, PhD.
CTO/Acting CEO, Modulated Imaging Inc.
Technology Incubator Office:
1002 Health Sciences Rd.
Irvine, CA 92612
Phone: (949) 8248367
Fax: (949) 8246969
[email removed]
This document may contain information that is privileged, CONFIDENTIAL and exempt from disclosure under applicable law. If you are not the intended recipient, please notify me immediately, as the use of this information is strictly prohibited.



Ah yes, I think what went wrong in my head is that I didn't realize that the result of the MC simulation is normalized. Then the figures Carole gave (O(1/N) for the variance and O(1/sqrt(N) for the SD) make perfect sense.
Thanks!
Martijn

