1

I do not know what method is appropriate to fit (say, a Log-Normal distribution) over an observed distribution. I say observed to make it sound generic because I am not sure if this qualifies as an Empirical distribution.

Anyways, after my simulation the data looks as follows:

Note that the density is very choppy but the CDF is much more smoother.

enter image description here

I would very much appreciate some advise how to approach this problem. I am looking for a method name, or some sort of sudo code. (I have to use Java to code it but that is not too important)

Edv Beq
  • 721
  • 1
  • 8
  • 21
  • The advice [here](http://stats.stackexchange.com/questions/157292/what-are-the-general-methods-for-parameter-estimation-in-statistics/157307#157307) briefly describes methods of parameter estimation. Most popular would generally be maximum likelihood. – Glen_b Sep 10 '16 at 15:45
  • @Glen_b Hi Glen, note that I do not have any random values from the simulation. All I have is the probability density as seen above. I thought one needs random variates to use the maximum likelihood method. – Edv Beq Sep 10 '16 at 15:59
  • Sorry to have misunderstood. So how what do you actually have -- a series of $(x,f(x))$ pairs? Or something else? How was the density obtained? – Glen_b Sep 10 '16 at 16:09
  • @Glen_b Yes that is correct. All I have is $ (t, f(t)) $. This is the density of a Wiener process functional. The original paper is titled ON DISTRIBUTIONS OF CERTAIN WIENER FUNCTIONALS - by M. KAC. I am doing a numerical simulation though - therefore, I do not get a nice closed form. – Edv Beq Sep 10 '16 at 16:32
  • The simulations themselves are giving you the $(t,f(t))$ pairs? – Glen_b Sep 10 '16 at 23:57
  • @Glen_b Yes, I can give you more background privately if you want. – Edv Beq Sep 11 '16 at 02:11

1 Answers1

0

A very straightforward option is ordinary least squares.

You have $(x_i, y_i)$ pairs, and want to fit a function $y=f(x, \mu, \sigma)$, $f$ depends on some parameters (like $\mu$ and $\sigma$ for Log-normal).

The game is to find the best set of $\mu$ and $\sigma$. To do so you have to find the $\mu$ and $\sigma$ that minimize $\sum_i(y_i-f(x_i, \mu, \sigma))^2$.

The hard part is to find a algorithm that will find the best set of $\mu$ and $\sigma$. You could try Gauss–Newton algorithm. It is available in many libraries, including The Cognitive Foundry in java.

Adrien Renaud
  • 670
  • 3
  • 8