Implement the ability to analytically simplify common combinations of distributions, for example the sum of two normal distributions is normal and the product of two log-normal distributions is log-normal.
The analytic simplification is performed by Distribution.simplify() which calls out to a helper class FlatTree.
I also modified sq.sample() to automatically call simplify() before sampling.
This PR is most useful in conjunction with #61. Performing analytic reductions before falling back to numeric methods increases both speed and accuracy. Without support for numeric methods, analytic reduction does increase the speed of Monte Carlo sampling, but probably only by a modest constant factor (eg if half the distributions in your model can be combined analytically, it doubles the speed), and it doesn't increase accuracy at all.
This docstring lists all the simplifications that it knows how to do.
Implement the ability to analytically simplify common combinations of distributions, for example the sum of two normal distributions is normal and the product of two log-normal distributions is log-normal.
The analytic simplification is performed by
Distribution.simplify()which calls out to a helper classFlatTree.I also modified
sq.sample()to automatically callsimplify()before sampling.This PR is most useful in conjunction with #61. Performing analytic reductions before falling back to numeric methods increases both speed and accuracy. Without support for numeric methods, analytic reduction does increase the speed of Monte Carlo sampling, but probably only by a modest constant factor (eg if half the distributions in your model can be combined analytically, it doubles the speed), and it doesn't increase accuracy at all.
This docstring lists all the simplifications that it knows how to do.