Error function must be a symmetrical non-negative function that
has second derivative. Moreover it must be convex near the zero.
A robust error function is an error function that grows slower than
the quadratic function, thus outliers do not influence the target
cost function much. There are many functions that satisfy the above
requirements. Intel IPP provides the Cauchy function (see “Intel
IPP Reference Manual” vol.1):
where C is a user-supplied parameter.