I've had too much going on at work to really contribute here much lately, but this seems a topic where I can without too much brainpower expended.
@QuietTexan is right about sample size and I can see where
@Culpeper was going, but I don't think the details are quite right.
If we assume that bullet speed per a fixed charge is normally distributed (and I have no idea whether it is or not), then the relationship between extreme spread (which is called the range in statistics), standard deviation, and the mean is as follows. The mean is a measure of central tendency: the value around which the data (our bullet speed) is centered. The range (or ES) is the set of values that define the lower and upper bounds of possible value for bullet speed (0 and infinity, we'll get to why later), and the standard deviation tells us how the data are distributed around our measure of central tendency.
Let's talk about standard deviation. A small SD indicates most values are close to the mean and a larger one means they become more disperse around the mean. If we assume a normal distribution, roughly 68% of the data will be within +/- 1 SD of the mean, 95% 2SD, and 99% 3SD. Once we get to around 5 SD we have pretty much captured all of the possible values with high certainty (but still not the range).
In statistics there are two versions of most summary statistics (mean, SD, range, median, etc.): the population version and the sample version. The population version is the universe of all data for that measurement. If you could record all bullet speeds for that charge then you would know the population. The sample is what you actually observe. If you are doing three shots and then calculating your statistics, then your sample is of size three. As you can imagine, if your sample size increases, your certainty of the veracity of the statistics increases and eventually becomes indistinguishable from the population versions. Think of creating a histogram (a bar chart with the y axis being the number of times a value is observed and the x axis being the values in increasing order from left to right). If you only take three measurements, then that histogram is not likely to be very accurate. In the same vein, taking 1000 measurements is likely more than you need to characterize the shape of the histogram. What is the right number depends on the range and the standard deviation (which we do not know ahead of time, unfortunately). So taking a reasonable guess for the sample size seems like an appropriate course of action.
Let's say we shoot 30 rounds and measure the speed of each one. We could create a histogram and inspect its shape. Assuming speed is normally distributed, does your histogram look like that normal bell shape curve? If not, maybe shoot 20-30 more times and then check. I would guess that you would need 50-100 rounds to really capture something that looks like a bell shaped curve. Take your mean and standard deviation of those 50-100 rounds and I think you could be confident of the performance of that particular recipe. 3x the SD will tell you what speeds you should expect 99% of the time (actually 99.7% of the time). I would not worry about range (or ES as it's know here) as long as 3x the SD gives you upper and lower bound numbers you can live with.
The important thing is that we have an appropriate sample size to have values we can trust. I would not trust a SD calculated from 3 values. Who knows what adding a fourth would do to the calculated SD. As you get into larger numbers, that adding a new number matters less and less. Hence trusting values from larger sample sizes more.
I don't know if this helps anyone or not, but I feel better now.
