Am 14.01.2011 20:03, schrieb Nikolay Pokhilchenko: > Graeme, I see an analogy with audio analog-to-digital conversion There is indeed such an alalogy, but... > (2D quantization - time and amplitude). May be the profiler must > automatically limit the harmonic spectrum of device response bumps when the > bumps spectrum is wider than can "digitize" the A2B. There should be the > Nyquist-Shannon sampling theorem taken in account. I suggest to perform low > pass filtration (LPF) of device response data before computing of A2B table. ...when the A2B table is created, it is already too late to apply an anti-alias filter. When we reach this stage, the device response _has already_ been sampled (at the points corresponding to the the patches on the target). The sampling of the device response basically happens when you print (and then measure) the target, and not when the A2B table is afterwards created from the samples. But a Nyquist filter must be applied _BEFORE_ sampling. This means you would need to install the anti-aliasing low-pass filter _in the device_ when you print the target, so that the printed target already reflects the sampling of the low-pass-filtered device behavior. But if you had any opportunity to make the device response smoother, then you would not have to care any more about high-frequency bumps, because they would disappear then anyway. So we rather have to consider the case where the device is as is, and you can't change its behavior. In this case you can't apply any anti-aliasing filter and the consequence of the sampling theorem is that you can only sample at a high enough frequency, greater than twice the highest frequency of the bumps in the device response, which simply means that you need to use enough patches. Sampling at higher and higher frequencies may become impracticable though, as CMYK space is 4-dimensional, i.e. doubling the sampling frequency requires to print and measure 16 times as many patches! Creating the A2B table from the measured patches does not involve any sampling, but spline fitting. The sampling theorem does not play a role here. What here happens is is actually the opposite of sampling, i.e. given a set of discrete samples we try to find a smooth, continuous function which approximates the original device behavior. Fitting smoothing splines to scattered data does an implied smoothing anyway, so I don't see any need for a separate smoothing pass before fitting the splines to the data. Regards, Gerhard