Owl2623 wrote:The effect of the gain would just be in the amount of variation in the measured signal.

And a change in the size of the range you can measure. In the analogy above, a 2.5kg weight would be outside the 0kg to 1kg range's limits, but within the 0kg to 5kg range.

In general, you get the best resolution from the smallest range that covers the largest value you need to measure.

Owl2623 wrote:The Gain = 1 would be noisy since it only takes .161 MicroTesla to change the measured value versus 0.805 microTesla at gain = 5. Is this correct?

Vaguely.. noise is a different subject, whose relationship to a given signal, sensor, and resolution is complicated. The signal itself will always contain some noise that sets the minimum resolution where it's possible to get meaningful information. Any sensor will have internal noise sources, some of which are proportional to the range or resolution.

The most useful thing you can say about resolution is its definition: the smallest change a sensor can measure repeatably. The resolution you need is overwhelmingly dictated by the way you plan to use the measurement.

Owl2623 wrote:I'm I interrupting this table correctly?

Yes, that sounds like the correct reading of the table.

That approach to measurement truncation is related to the vaguely-tolerable concept of 'significant figures'.. the idea that you can't divide 100 by 3 and expect an answer of 33.333... to be accurate to an infinite number of decimal places.

With sig-figs, the value '100' implies an error range less than +.-0.5. The value '100.0' implies an error range less than +/-0.05, and so on. The core idea of significant figure arithmetic is that you can't expect the result of a calculation to have more significant figures than the original values. 100/3 means (100 +/-0.5)/(3 +/-0.5), which can land anywhere between 28.4 (99.5/3.5) and 40.2 (100.5/2.5). The shortest original value has one significant figure, so the result can only be reliable to one significant figure: 3x10e1.

(Sig-figs are a disaster when applied without informed judgement BTW, and 'informed judgement' is harder to explain than proper statistical error propagation).

In the MLX90393, you can select 16 significant binary digits from a 19-digit raw value. If you select the most significant set, you give up the three lowest digits of resolution. If you select the least significant set, you give up the three highest digits of range.