David Llewellyn wrote:
> BUT! Please tell me this, if you know, as it's something I have long
> wondered about. Is any inaccuracy with a digital thermometer a 'constant'
> inaccuracy, which you can calibrate for, and therefore correct for when
> taking readings? Or is it a 'variable' inaccuracy e.g. you measure boiling
> water one day and get 100.5 C, and another day and you get 99.4 C (assuming
> constant atmospheric pressure)?
In all instrumental measurements, the total uncertainty has many parts
- a constant bias from the instrument - this is the part that would be
cancelled when taking a difference of measurements as is the case
- a linearity error - a bias such as described above may have a
constant value at a certain point of the scale, and a totally
different value at another point of the scale. Important linearity
errors are often seen in hydrometers for example. The better part of
this would also be compensated for in a differential measurement
because the 2 temperatures are quite close. It could become important
if you took the difference in temperature between boiling and freeing
temperatures for example.
- a random error, which comes from the intrinsic resolution of the
sensor and related electronics, and on how this may be affected by
external conditions - this part is unpredictable...
These are what is called the instrumental error. To this you should
add another element:
- the manipulation error, which is caused by the person who may use
the instrument in an inappropriate way. For example in this case, the
sensor could be inserted too low or too high in the ebuillometer, or
there could be a leak of steam, or some other factor caused by the
user could lead to an additional error.
The problem with low cost instruments is that it is pretty difficult
to evaluate these different elements. Only when you buy
instrumentation for laboratory and high precision applications will
you get this sort of information - but you pay quite a bit for this!
So the question really is - what is the random error of such a
thermometer? You might get a feeling for this as you use it, but you
may probably assume in first approximation that the quoted uncertainty
(+/- 1C) may be divided in 3 equal parts for the constant bias, the
linearity, and the random error, which would give you approximately
+/- 0.3C of possible random error. Does this make sense to you?