Accuracy and precision
The data sheet says 2.5 metres, and in the field the device is six metres off. Both can be true. Accuracy figures are probability statements, and without the abbreviation behind them the number is worthless.
Trueness and precision
In everyday speech accurate simply means close. In measurement the term splits into two independent properties.
Precision describes how closely repeated measurements agree with each other. A precise device returns nearly the same result every time and says nothing about whether that result is correct.
Trueness describes how far the mean of many measurements sits from the true value. A true device is right on average, even if individual fixes scatter.
Accuracy in the strict sense is both together, and the distinction matters because precision can be improved by averaging and trueness cannot. Averaging for ten minutes over a point whose coordinate is in the wrong map datum yields a beautifully stable value a hundred metres away from where you want to be.
| Type of error | Examples | Reduced by averaging? |
|---|---|---|
| random, degrades precision | receiver noise, rapidly changing reflections | yes |
| slowly varying | ionosphere, orbit errors, multipath at a fixed site | only over hours |
| systematic, degrades trueness | wrong datum, wrong antenna height, geoid model | no |
The measures
Because position errors scatter, there is no radius containing every fix. Each figure therefore names a radius and the probability of landing inside it, and the common abbreviations differ in exactly that probability.
| Symbol | Meaning | Share of fixes inside | Relative to CEP |
|---|---|---|---|
| CEP | circular error probable, the median radial error | 50 % | 1.00 |
| DRMS, RMS | root mean square radial error | ≈ 63–68 % | 1.20 |
| R95 | radius containing 95 per cent | 95 % | 2.08 |
| 2DRMS | twice the DRMS value | ≈ 95–98 % | 2.40 |
Factors for a circular normal distribution, meaning equal scatter north-south and east-west. With an elongated distribution the shares for DRMS and 2DRMS move within the ranges given.
The arithmetic matters more than the definitions. A manufacturer quoting 2.5 m CEP is saying that every second fix falls outside 2.5 metres. For 95 per cent of fixes expect a little over 5 metres, for 2DRMS about 6. The same receiver quality can therefore be advertised as 2.5, 3, 5.2 or 6 metres depending on the measure chosen. Where no abbreviation is given, CEP is almost always meant, because it yields the smallest number.
Height and three dimensions
All of the above refer to horizontal position. Height has its own figures, and the three dimensional equivalent of CEP is SEP, the radius of a sphere containing half the fixes. Height with GPS is fundamentally one and a half to two times worse than horizontal position, because every satellite is above the receiver and none below. A horizontal specification therefore says nothing about height.
Under what conditions
An accuracy figure without conditions is incomplete. Data sheets normally assume an open sky, often average over 24 hours, and sometimes assume SBAS correction. Under forest canopy or between buildings the real values are several times worse. More importantly, these measures capture random errors only. An outlier caused by a reflection off a glass façade, or a wrong setting, appears in no manufacturer statistic.
Checking your own device is easy: log a fixed point with a clear sky for an hour, look at the point cloud, and find the radius containing half the points. That is the CEP under the best conditions. Trueness is then checked against a surveyed benchmark whose official coordinates are published.
What follows from this
- Distrust single fixes. One position is a sample from a distribution. For an important waypoint, average over several minutes.
- Stable is not the same as correct. Points tightly clustered but systematically displaced almost always indicate a setting, not reception.
- Compare like with like. 3 metres CEP and 5 metres 2DRMS describe practically the same device.
- The accuracy shown on screen is an estimate derived from signal quality and geometry, and it can be far too optimistic in difficult terrain.