Given an algebraic expression, f, fracdiff computes the th derivative of f with respect to x, where  is not an integer. For integer (_nu_) order differentiation use diff - see also symbolic integer order differentiation.

By default, the direct method is used to calculate the fractional derivative using the Davison-Essex (D-E) definition, that is, first differentiate n times, then integrate n- times, where n = ceil(), using the standard formula for iterated integrals. The resulting D-E definition of fractional derivative is

diff(f(x),[x$nu]) = 1/GAMMA(n-nu)*Int((x-t)^(n-nu-1)*diff(f(t),[t$n]),t = 0 .. x);

where (see ceil)  <= n. This definition handles differentiation orders  > -1, and for integer values of  it is valid only in a limiting sense.

In the literature, the term fractional derivative is sometimes reserved for the Riemann-Liouville (R-L) fractional derivative, defined by

diff(f(x),[x$nu]) = 1/GAMMA(n-nu)*Diff(Int((x-t)^(n-nu-1)*f(t),t = 0..x),[x$n]);

The Davison-Essex and the Riemann-Liouville definitions are different in the following aspect: in the D-E formula, differentiation is performed first, then integration; in the R-L formula it is the other way around. The D-E definition implemented, thus, maps constants to zero, imitating integer order differentiation, while the R-L definition does not. This property of the D-E definition makes it suitable to work with initial value problems for fractional differential equations.

The  method calculates the Davison-Essex derivative via the Laplace transform. This method will not accept a symbol or name representing an unspecified real order of differentiation, unless assumptions are made about the symbol  that allow the routines to deduce the numerical value of n = ceil(). The routines will attempt to calculate the derivative for any non-integer real-valued order.

The series method will calculate the derivative by first expanding the expression into a power series. If any term in the returned series is not of the form constant*(name - a)^constant, an error message will be returned. The terms are then differentiated using a simplified case of the Davison-Essex formula (where the function being differentiated is a monomial). The optional parameters for this method, to be specified in mthopts, are  and , where a specifies the point about which to expand the series and o specifies the accuracy or order of the series.  defaults to  and  defaults to Order, the default order for series.

If  is a negative number, integration is performed on the terms in the series. This method accepts symbolic values of , and returns a formula for the -th derivative. However, the values returned by this series approximation will become less accurate as  increases, because terms that should normally be eliminated by differentiation will not be.

Error, (in fracdiff) parameter -6/5 is <= -1

Benghorbal, M. Power Series Solution of Fractional Differential Equations and Symbolic Derivatives. PhD Thesis, University of Western Ontario, Canada, 2004.

Davison, M., and Essex, G. C. "Fractional Differential Equations and Initial Value Problems." The Mathematical Scientist, (December 1998): 108-116.

Liouville, J. Collected Works.