We now validate the proposition above, that solving
for the maximum entropy
multivariate truncated exponential (MaxEnt-MTE) distribution
approximates the asymptotic distribution of MCMC-UMS
for doubly-bounded data.
In this experiment, we used a data size of .
The matrix in (5.1)
computed the first Coefficients of the
length- DCT of .
We created a data sample using a raised sine-wave plus
Gaussian noise, clipped it to the interval [0,1], then computed the feature .
The original was then discarded. We then got a starting point
for
using a linear programming solver
as previously explained, with the lower and upper bounds of 0 and .
We then maximized (7.6) over
subject to (5.12)
using the method above. Figure 7.2 shows the results.
See software/test_lin01.m.
Figure:
The Maximum entropy
asymptotic mean (dark line) and the sample mean
(circles). One random MCMC-UMS sample is shown (light jagged line).
The pseudo-inverse solution (dotted lin) is seen to
have values outside .
|
The sample-mean of MCMC-UMS after 10000 samples matches the
Maximum entropy mean solution as close as could be determined.