(with a bit on sigfigs thrown in for good measure)
Consider the dartboards shown below, in which the 'grouping' of thrown darts is a proxy for our laboratory measurements. A 'precise' measurement means the darts are close together. An 'accurate' measurement means the darts hit close to the bullseye. Notice the combinations:
|Measurements are precise,
just not very accurate
|Measurements are accurate,
but not precise
precise nor accurate
precise and accurate
The remedy for this situation is to find the average diameter by taking a number of measurements at a number of different places. Then the scatter within your measurements gives an estimate of the reliability of the average diameter you report. Note that we usually assume that our measured values lie on both sides of the 'true' value, so that averaging our measurements gets us closer to the 'truth'.
Another approach, especially suited to the measurement of small quantities, is sometimes called 'stacking.' Measure the mass of a feather by massing a lot of feathers and dividing the total mass by their number.
Suppose your sensor reports values that are consistently shifted from the expected value; averaging a large number of readings is no help for this problem. To eliminate (or at least reduce) such errors, we calibrate the measuring instrument by comparing its measurement against the value of a known standard.
It is sometimes quite difficult to identify a systematic error. Get in the habit of checking your equipment carefully. Make a preliminary analysis of your data early in the experiment; if you gather all the data without checking for systematic error, you might have to do it all over again!
The reading of a vernier caliper may vary within the members of a lab group because each person reads it slightly differently. Or one observer's estimate of the fraction of the smallest caliper division may vary from trial to trial. The mean value computed from multiple trials averages out some of the random error; repeated measurements are required. Some people even say "one measurement is no measurement."
Another subtlety is the recognition of 'outlying' or 'low probability' data points. If justifiable (and that often takes some thought), excluding 'bad data' will reduce your error.
Average deviation from the mean is a measure of the precision of the measurement: the smaller, the better.
In general, report a measurement as an average value "plus or minus" the average deviation from the mean. Another totally acceptable format is
% deviation = 100 * average deviation / mean value.
For example, the chart below shows data from an experiment to measure the life of two popular brands of batteries.
(Data from Kung, Am. J. Phys., Vol. 73, No. 8, p.774).
|Trial||Duracel (hours)||Energizer (hours)|
|Averages||Duracell: 9.4 hours||Energizer: 10.0 hours|
The question here is: which brand of battery lasts longer? How can we tell?
Using the 'general procedure' above, calculate the deviation from the mean of each data point. Then calculate the average deviation. Thus we would report battery life for Duracell as '9.4 +/- 2.3 hours'.
Another important consideration when comparing two sets of data: the spread of the data points, sometimes called the range of the measured values. What does it suggest if the range of measurements for the two brands of batteries has a high degree of overlap?
But in case you are curious, standard deviation is computed as follows:
If M is the mean of N measurements xi, then the standard deviation is
This algebraic expression gives rise to the term 'root mean square error,' (abbreviated RMSE). A related quantity is the variance, which is just the square of the standard deviation. Note that a low RMSE value does not equate to a 'right' answer!
Suppose we want to measure a known quantity that has the value A, but our experiment gives us the mean value m deviation from the mean r. By the average deviation procedure, we report that the measured value is m +/- r. How do we decide if we can live with the size of r?
This last expression will be used frequently! Do not confuse experimental uncertainty with average deviation. If you are still uncertain of the distinction between these two, go back and look at the dartboards again.
Once you have your uncertainty of measurement, you can show this quantity on your graphs. Using Graphical Analysis, right click on the data table and select Column Options. Under the Options tab, check Error Bar Calculations, then enter either a percentage, fixed value or put your error numbers into a column of their own and select Use Column.
If the actual quantity you want is calculated from your measurements, in some cases the calculation itself causes the uncertainties to increase dramatically. Here are the common situations:
Then z +/- dz = ( x +/- dx) (y +/- dy) = xy +/- xdy +/- ydx + dx dy.
Since dx and dy are both small (we hope) the dx dy term should be small enough to neglect.
Let's drop the +/- for the sake of clarity. Thus the relative uncertainty in z is dz / z = (x dy + y dx) / (x y) = dx / x + dy /y which means the relative uncertainty in the calculated product of two measured quantities is the sum of the two measured uncertainties (dx / x plus dy / y).
In this case, it can be shown that dz / z = n dx / x (it has to do with logarithms).
Therefore the relative error in the calculated quantity z is the power n multiplied by the relative error in the measured quantity x. Squaring the measured quantity doubles the relative error!
Shamelessly assimilated (resistance is futile!) from: University of Wisconsin Physics Lab Manual