Jr. Macdonald, Comparison of parametric and nonparametric methods for the analysis and inversion of immittance data: Critique of earlier work, J COMPUT PH, 157(1), 2000, pp. 280-301
Recently, two methods for the estimation of discrete and/or continuous dist
ributions of relaxation times from small-signal electrical frequency-respon
se data have been compared. For discrete-line distributions, the parametric
method used was found to be inferior in some ways to the nonparametric one
, which involved Tikhonov regularization, and it was concluded that the par
ametric one could not be employed to estimate continuous distributions at a
ll. Here it is shown by Monte Carlo simulation that both conclusions are in
correct. The same data situations analyzed in the earlier work were reanaly
zed using a complex nonlinear least-squares parametric method that has been
employed to estimate discrete-line distributions since 1982 and continuous
ones since 1993. Quite different results from those presented earlier were
obtained, and the original parametric method was shown to be far superior
to the nonparametric one for the estimation of discrete-line distributions,
since inversion is unnecessary and resolution is far greater. For continuo
us or mixed distribution inversions, the parametric method was again superi
or, and it allows unambiguous distinction between discrete-line points and
those associated with a continuous distribution, while the nonparametric in
version method does not allow such distinction and approximates all distrib
utional points as continuous-distribution ones. The parametric method used
and described here is also valuable for other data analysis tasks other tha
n those involving inversion. Some of its error characteristics are investig
ated herein, and the importance of matching the weighting error-model to th
e form of the errors in the data is illustrated. It was found that with nor
mally distributed random errors added to exact data, the distributions of e
stimated parameters were not normal but were closer to normal for proportio
nal errors than for additive ones. (C) 2000 Academic Press.