Hal Murray writes: > Looks like we are getting into the old precision vs accuracy tangle. > Do we have a glossary with useful descriptions? > That slot should represent the number of useful bits.
It really doesn't help that precision, accuracy and resolution have a number of different meanings depending on the context. It is doubly unfortunate that NTP necessarily uses several of these contexts. In measurement statistics, precision describes the variability of the distribution of many repeated measurements of the same quantity and accuracy the deviation of the location parameter on that statistic to the "ground truth" (or more commonly to another measurement that ius known to have higher precision and accuracy). It is generally assumed that the measurement process has much higher resolution than the precision of the measurement and the measurement distribution can be assumed to be continous. Resolution in this context is the smallest change in the measured quantity that will yield a discernible change of the measured value. In numerics and software in particular, precision and resolution are often confused to interchangeably mean the number of digits in the computation or presentation. Precision should really be reserved for the significant number of digits going into and coming out of a computation. Unfortunately accuracy is often used to describe the same (how many digits can I trust). Accuracy should really mean "how far from the true result is the number I've got". Accuracy can easily be greater than the precision (e.g. if I start a calculation with a low-precision number and have high numerical resolution and an algorithm that doesn't amplify errors). In the case of NTP, we have the measurements themselves, the quantized representations of those measurements and algorithmically processed quantities that were derived from measurements. Asuuming that the resolution is always large enough to not conceal otherwise available precision, you can't get any more accurate or precise than the measurement if you have only a single one. Once you start having multiple measurements, you can get both more precise (using data from the same measurement process) as well as accurate (using data from multiple measurement processes) than the individual measurements themselves, provided the statistics are well-behaved and all calculations have suitable error propagation characteristics. Another hidden assumption is that the local clock must be stabilized over the time-scale used for data fusing to somewhat better than that final error since NTP uses it both as the target of the control loop and the reference of the measurements. Regards, Achim. -- +<[Q+ Matrix-12 WAVE#46+305 Neuron microQkb Andromeda XTk Blofeld]>+ Waldorf MIDI Implementation & additional documentation: http://Synth.Stromeko.net/Downloads.html#WaldorfDocs _______________________________________________ devel mailing list devel@ntpsec.org http://lists.ntpsec.org/mailman/listinfo/devel
