commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Fran Lattanzio <>
Subject [math] Numerical derivatives in Commons Math
Date Thu, 11 Aug 2011 21:27:29 GMT

I have a proposal for a numerical derivatives framework for Commons
Math. I'd like to add the ability to take any UnivariateRealFunction
and produce another function that represents it's derivative for an
arbitrary order. Basically, I'm saying add a factory-like interface
that looks something like the following:

public interface UniverateNumericalDeriver {
 public UnivariateRealFunction derive(UnivariateRealFunction f, int derivOrder);

For an initial implementation of this interface, I propose using
finite differences - either central, forward, or backward. Computing
the finite difference coefficients, for any derivative order and any
error order, is a relatively trivial linear algebra problem. The user
will simply choose an error order and difference type when setting up
an FD univariate deriver - everything else will happen automagically.
You can compute the FD coefficients once the user invokes the function
in the interface above (might be expensive), and determine an
appropriate stencil width when they call evaluate(double) on the
function returned by the aformentioned method - for example, if the
user has asked for the nth derivative, we simply use the nth root of
the machine epsilon/double ulp for the stencil width. It would also be
pretty easy to let the user control this (which might be desirable in
some cases). Wikipedia has decent article on FDs of all flavors:

There are, of course, many other univariate numerical derivative
schemes that could be added in the future - using Fourier transforms,
Barak's adaptive degree polynomial method, etc. These could be added
later. We could also add the ability to numerically differentiate at
single point using an arbitrary or user-defined grid (rather than an
automatically generated one, like above). Barak's method and Fornberg
finite difference coefficients could be used in this case:

It would also make sense to add vectorial and matrix-flavored versions
of interface above. These interfaces would be slightly more complex,
but nothing too crazy. Again, the initial implementation would be
finite differences. This would also be really easy to implement, since
multivariate FD coefficients are nothing more than an outer product of
their univariate cousins. The Wikipedia article also has some good
introductory material on multivariate FDs.


To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message