steven
|
Hello Yawei,
I thought a little more and came to understand the statement. When doing code density test, the histogram for the codes is plotted. Then the DNL would be regarded as the y-coordinate variance(the height of each bin) on the histogram while INL as the x-coordinate variance(the width of each bin). In this case, since the width is sampled in statistical way or large number theorem so if using the same confidence interval of the height, then a much larger number of samples has to be used to vision the true width of each bin. On the other hand, the height variance of each bin or DNL is also coming the statistical measure but only relates to the input sinusoidal signal amplitude. The algorithm on calculating the DNL shows this. So the later part of the paper, it was suggested to use FFT test on INL (and other dynamic parameters), which requires shorter samples.
I think the sum of DNL results in INL is just for derivation. In testing, INL would be tested separately because errors would corrupt the histogram (see Razavi's book "data conversion system design", page 248).
But there is still something I am not sure I have understood completely.
Thanks, Steven
|