NavList:
A Community Devoted to the Preservation and Practice of Celestial Navigation and Other Methods of Traditional Wayfinding
Re: How good is St. Hilaire?
From: George Huxtable
Date: 2010 Mar 1, 15:22 -0000
From: George Huxtable
Date: 2010 Mar 1, 15:22 -0000
I had referred to the section in AstroNavPC, the UK Nautical Amanac office publication, page 68 in the 2000-2005 edition, which took the least-squares procedure a bit further and showed how the error ellipse was deduced for their software. In passing, I referred to "...a minor error in its text..." To which Herbert Prinz has responded- "This rings a bell. Checking my copy of AstroNavPC, from 2001, I find that I had penciled "A = C and B = 0" into the margin of p. 68. Is that it? " That's it, exactly. The same is scribbled in mine. We agree. These are the conditions that produce a confidence-circle rather than a confidence-ellipse, and not A=B and C=0, as was stated. There's another error, actually. The text tells us that two quantities, sigmaL and sigmaB, are the standard deviations of the estimated position in longitude and latitude, respectively. Indeed, one is the standard deviation of the latitude, in arc-minutes or miles. But the other is the standard deviation of the East-West position, in miles, which isn't the same (except at the equator) as the SD of longitude. It's what old-salts termed the "departure". Herbert continues- "But why do you think "the whole concept of deducing an error-ellipse from a single cocked-hat is fundamentally flawed" ? The ellipse serves to measure the quality of the data set which is at the basis of that single cocked hat. Specifically, it does a better job of it than the cocked-hat itself." Well, it does a very similar job to the cocked hat. The problem is that that "measure of the quality of the data", cocked-hat or confidence-ellipse, is based on only a single sample. It's relying on the statistics of a single event. This is a problem that affects not just the computer predictions of the program supplied with the AstronavPC booklet, but presumably applies to other software that uses similar algorithms to reduce multple observations. To simplify the picture, assume that just three stars are observed, equally spaced around the horizon 120º apart, in which case any cocked-hat will be an equilateral triangle, and any confidence-ellipse will become a simple circle. Assume, also, that any systematic errors have been corrected for and nulled out, and only random scatter remains. When you plot the position lines that result from a round of three star sights, any pair of such lines will cross at a point that's somewhere in the general vicinity of the true position. When you draw in the third position line, it will, in general, miss that point, but will go past somewhere close, to one side or the other. How close? It depends on chance, to a large extent. Sometimes, just by chance, it will happen to pass indistinguishably close to it. In which case, the cocked-hat, and the deduced error-circle, will be minutely small. But that doesn't allow you to draw any firm conclusions about that particular round of sights, that it was somehow "better" than the round of sights you took a few hours before, which provided a more reasonable-looking cocked-hat, a few miles across, or a confidence-circle a few miles in diameter. It was just a matter of chance. So, let's say you were attempting a narrow unmarked passage through a reef, which relied on that round of star-sights giving a precise position. Would you feel emboldened to try it, by that tiny cocked-hat, and confidence-circle? Only a gambler would do so. The mariner should be guided, not by the confidence-circle resulting from that one round, but by the general run of such confidence-circles that his rounds of star-sight have been producing recently. And that leads to what I regard as another flaw in the text on that page of the AstronavPC manual. The confidence-ellipse is suggested to be drawn with a scale-factor k, of 2.4, to give a confidence level P of 95%. The text then says- "statistical theory shows that the estimated position has a probability P of lying within" such a confidence ellipse. Whatever that means. As I see it, the confidence ellipse is intended to be plotted around that estimated position, so the estimated position will therefore ALWAYS be placed at the exact centre of such a confidence ellipse! That is the only place in the book I have been able to find any explanation of the meaning of the confidence ellipse. The associated computer program simply displays the ellipse on the screen, gives its dimensions and states the 95% confidence level, without giving any further guidance about what it's supposed to mean. What would be nice to have, if it was possible, is the probability that such a confidence-ellipse will contain the TRUE position, not the estimated position. But say a star-round has, just by chance, produced a tiny cocked-hat and ellipse, and we have nothing else to go on? Can we then claim that this tiny ellipse has a 95% probability of containing the true position? We can not! However, we have nothing better, to tell us otherwise, so it's the best that can be done with that data, a single star-round. But the apparently "scientific" nature of the confidence ellipse lends an entirely spurious respectability to its predictions, to which it is unentitled. I am not suggesting that such a confidence-ellipse is valueless - far from it! All I'm saying is that it has to be grouped with others before a savvy navigator is able to assess its value. Those warnings apply particularly to a cocked-hat triangle, when only three bodies are involved. The more star-sights that are included in the assessment, the less likely it will be that they will accidentally combine to produce a spuriously-small confidence ellipse. Several years ago, before the next edition became due, I drew some of these defects in the 2000-2005 edition to the attention of the UK Nautical Almanac Office, so it's possible that changes have been made in later editions. I would be interested to learn if they have. George. contact George Huxtable, at george@hux.me.uk or at +44 1865 820222 (from UK, 01865 820222) or at 1 Sandy Lane, Southmoor, Abingdon, Oxon OX13 5HX, UK.