[argyllcms] Re: profile <1 million dollar question!>

Roger Breton wrote:


My thinking is, I would like to demonstrate to the students the effect of
using a larger device sample in modeling its behavior. So, if I use
something like an IT8.7/3 928 sample set (as opposed to my lowly 182 sample
set) or an ECI2002 or an IT8.7/4 or, something like 3000 patches by Argyll,
can I reasonably expect those two statistics to go down?


It could well have the opposite effect, of going up. If the sample
points have various random errors in them, and/or the device has
a behaviour "shape" that the profile model doesn't fit perfectly,
then providing more points can increase the certainty with which
the randomness is averaged, or the profile mis-fit is confirmed.
If you put very few points in, then it's easy for the model to
exactly fit these points, giving a very small self-fit error.
(Argyll does try and compensate for these effects by increasing
 the smoothing factors for smaller numbers of points, and decreasing
 them for larger number of points).


Or is that a sign of the device ill-behavior?  This is from an Epson
printer, by the way. Linearized by a CMYK RIP.


The self fit error differences between devices measured with a comparable
number of points will be an indication both of the degree of fit the model
has to the native device behaviour, and the level of randomness in
the measured values. So low numbers generally indicate a repeatable device
that has smooth, consistent behaviour. (The smallest self fit errors I ever
saw were for an old fashioned chemical proof. The physical separation
of the colorant layers makes the optical mixing simpler, and it was a
very repeatable consistent print process)

Graeme Gill.

Other related posts: