Exporing the Lanczos' Generalized Derivative
Developed in the 1950's by the Hungarian
mathematician Cornelius Lanczos, the Lanczos' Generalized Derivative (LGD) the function f
at x is written as
,
provided that this limit exists. The motivation for
this formula comes from least squares regression. It can be shown that this formula yields
a proper extension of the derivative. In other words, if a function has a normal derivate
at a point, then the function possesses an LGD and the two quantities are equal. However,
a function can possess an LGD at a point even if the normal derivative is not defined. An
example is the piecewise linear function
,
whose graph is shown below.

For this function, LGD(f)(0)=.5, a fact that
allows one to draw a "pseudo-tangent" line of sorts at the origin. (The line in
red above.)
Since the LGD is based upon the ideas of linear
regression, one can derive a quantity known as "instantaneous correlation,"
which measures the ability of pseudo-tangent lines to approximate the original function
near the point at which the LGD is based. For example, the instantaneous correlation
of the pseudo-tangent line above works about to be approximately .554. Not surprisingly,
if a function has a normal, nonzero derivative at a point, the correlation is 1 or -1, a
fact demonstrating perfect positive or negative association, respectively.
Many interesting functions possess LGD's. For example,
the function shown below was constructed by Karl Weierstrass and is continuous but has a
normal derivative nowhere.

Thus function has period one and possesses a LGD at x
if x has a finite ternary expansion. If x has a repeating ternary expansion that cannot be
written as a finite one, then LGD(f)(x) does not exist.
Many open question remain concerning the LGD. Here are
but a few.
The LGD does not obey the usual differentiation
rules of calculus, e.g. the product and chain rules, even for continuous functions. How
well-behaved must a function be in order for the LGD of the function to obey these
calculus rules?
What are the weakest conditions on a function for
which the LGD exists?
Is there a multivariable version of the LGD? If so,
what is its formula?
Is there a version of the LGD for vector fields?