Least-significant digit accuracy at low end of range

[grenert]

My understanding from the EEVblog on accuracy/resolution is that accuracy specs are made at full scale.  Does the +/- LSD spec change at the low end of the scale?  For example, if a 2000-count meter is specified as 0.1% accuracy +/- 5 counts, does that mean on the 200 V range a measurement of 1 V could appear on the meter as anywhere from 1.5 to 2.5 V?  That's an effective error of 50%! 

[Bored@Work]

You talk about digital multimeters, right?

Typically, specs are given as +/- (percentage of reading + Digits). Where Digits is another way to express percentage of nominal full scale value, or LSDs. The Digits rating doesn't change at low measurement values.

Therefore, one of the golden rules of using a meter is to try to use a range where your value is in the upper 2/3 of the full scale of that meter.

[saturation]

With that much drift, zero volt input to the DMM can cause it to fluctuate between 200.5 and 199.5 making resolving 1V on the 200V scale unhelpful.

But, using the same specs on a 20,000 count DMM you'd have:

200.05 to 199.95 at zero input,

and at 0.1% accuracy 1V would be 0.999 to 1.001 and it would now be viable, with readings somewhere between 200.95 - 201.05, for about 5% total error.

My understanding from the EEVblog on accuracy/resolution is that accuracy specs are made at full scale.  Does the +/- LSD spec change at the low end of the scale?  For example, if a 2000-count meter is specified as 0.1% accuracy +/- 5 counts, does that mean on the 200 V range a measurement of 1 V could appear on the meter as anywhere from 1.5 to 2.5 V?  That's an effective error of 50%!  

[alm]

My understanding from the EEVblog on accuracy/resolution is that accuracy specs are made at full scale.  Does the +/- LSD spec change at the low end of the scale?

No.

For example, if a 2000-count meter is specified as 0.1% accuracy +/- 5 counts, does that mean on the 200 V range a measurement of 1 V could appear on the meter as anywhere from 1.5 to 2.5 V?  That's an effective error of 50%! 

Correct. It's x% of value + y% of full scale (often expressed in digits). At the bottom of the range, the second term will become the most significant. You should use the lowest range possible, like BoredAtWork wrote. This will give you both better accuracy and better resolution.