Thanks so much, now I have the article.
But a question remains open: How to choose
the order of the polynomial to
represent the lag weights?
In the article it is just assumed.
The order of a PDL, relative to the lag length, indicates the
strength of constraints placed on an unrestricted DL specification.
I.e. 12 lags would require 12 coefficients, unconstrained, but only
three if a 2d order PDL is fit (depending on endpoint constraints). A
3d order PDL would be less restrictive. As a 2d degree polynomial can
have only one inflection point, this restricts the shape of lags you
can get from a quadratic PDL. A cubic allows two inflection points,
and so on. It would probably be sensible to apply a relatively high
order PDL (not imposing many constraints on the DL form) and then
"test down" to a more parsimonious degree if the data are compliant.