Scott,

>I know what you mean. In case anyone's interested, this is how you
measure
>linearity error.
>

>2) Find the least square's curve fit of the data. In other words,
find the
>line that best fits the data.


I "cheat" and use Excel's linest function to achieve the same result.
Linest does a least squares staright line curve fit identical to what
you described.

Format is

    INDEX(LINEST(Ax:AY,Bx:By),Z)

Where Ax:Ay is ther range of the first variable (millivolts here)
Bx:By is range of second variable - sigmadelta count here
And Z is 1 for slope and 2 for intercept of the best fit (least
squares) line.
You then just multiply each sigma-delta count by the slope and add
the intrecept to get a "predicted" voltage output from the converter.
Plotting this against the actual millivolts input

i) As two lines - gives you a good visual check on any strange
values.

ii) As an X/Y scatter graph - SHOULD give a rluler flast line if its
working properly. It's an excelent way of spotting discontinuities
and non linearities . If the capacitor is too small you get
"plateaus" at certain points - the first to show as the capacitor is
reduced is around  the half maximum count point eg 1024  in a 2048
count decimal system.


>1) Collect a whole bunch of data and record a) the voltage applied
and b)
>and the converted digital value. You also should collect several
samples
>at each voltage value and use the mean of the digital values to
represent
>converted digital value. There's another test for measuring the
spread.


I haven't done any formal "spread" measurements yet but I let the
converter run at a single input value and watch the output count - if
it varies by more than +/- 1 count over a short trm I want to know
why. Sometimes you get a few counts variation towards one or other
end of the range. If the comparator / gate switching point is not
well defined the count will fluctuate more widely for a fixed input
and you can be sure that the results are going to be poor.

So far I get MUCH worse results than you do using gate inputs but
much the same using a comparator.



>2) Find the least square's curve fit of the data. In other words,
find the
>line that best fits the data.
>
>3) Find the sample that most deviates from this line.



Plotting the graph in Excel does this visually and lets me see all
the data at a glance. For runs with good results I also plot the
limit lines for 1 and 2 count errors to see how tightly the data
matches theoretical. I often get many of the actual data values well
under the 1 count error limit indicating that they are the best
possible result available for that particular input and number of
samples.


OTHER

I also do a "real" estimate based on what I'm going to have in the
final system - a zero point which can be calibrated any tine on the
fly and a max value point available only during setup.

Russell estimated voltage = (COUNT - COUNTmin) x
(mVmax-mVmin)/(COUNTmax-COUNTmin) + mVmin

This gives a curve with zero error at each end and a positive error
curve between the two points. It is generally more inaccurate than
the least squares curve but represents something like what I'm going
to be able to actually do in practice. In the final version there
will (probably) be one adjustment = max value set, and the zero value
will be read as required.



RM