Distribution Functions Change Behavior
AFter the extensive input from folks here last week, and my examination of alternatives, I accepted the visual appearance of Gaussian curves in our model. As I check the plots for data errors I find a behavior change when the x-axis length is 14 rather than 100, and I do not understand why. I would appreciate the ideas of the mathematically adroit here to help resolve the problem. However, if this is not the appropriate forum for such a discussion, please point me to the more appropriate one. Most of the plots cover a range of independent values of 0-100. Some are greater (e.g., 0-1,000 or 0-10,000). However, for pH the range is 0-14. The plots for pH do not consist of sigmoid curves (a 'Z-curve' from 0-7 and an 'S-curve' from 7-14); they are straight lines. The Gaussian curve does not approach zero on either end (0 or 14). I can provide the functions for these curves. I need to understand why the results change when the range of the independent variable is short and learn how to produce correct results. Rich -- Richard B. Shepard, Ph.D. | Integrity Credibility Applied Ecosystem Services, Inc. | Innovation http://www.appl-ecosys.com Voice: 503-667-4517 Fax: 503-667-8863
Hi Rich, If your data is truncated at zero, it is not Gaussian (drawn from a normal). You will notice this when you shrink the range of values (unless the variance is tiny). Cheers, Alan Isaac
I may not understand what you are asking, Rich, but I'm not sure I
agree with Alan. A Gaussian fit to data x should fit exactly as well
as data fit to ax, a > 0, just with a variance a^2 times the original.
The only way this would not be true is if:
1. You are not fitting the variance, but only the mean
2. There's some numerical issue (like some of the data are represented
as integers, etc.)
Don't know if it could be one of those issues...
--Hoyt
On Mon, Apr 28, 2008 at 7:18 AM, Alan G Isaac
Hi Rich,
If your data is truncated at zero, it is not Gaussian (drawn from a normal). You will notice this when you shrink the range of values (unless the variance is tiny).
Cheers, Alan Isaac
_______________________________________________ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
-- +++++++++++++++++++++++++++++++++++ Hoyt Koepke UBC Department of Computer Science http://www.cs.ubc.ca/~hoytak/ hoytak@gmail.com +++++++++++++++++++++++++++++++++++
Wait, I think I see what Alan is saying. When you use a gaussian
approximation on truncated data, the accuracy of the truncation is
very dependent on where in the interval the mean is. If it's near the
edges, the results will be worse. The width of the interval, though,
is a separate factor.
--Hoyt
On Mon, Apr 28, 2008 at 8:29 AM, Hoyt Koepke
I may not understand what you are asking, Rich, but I'm not sure I agree with Alan. A Gaussian fit to data x should fit exactly as well as data fit to ax, a > 0, just with a variance a^2 times the original. The only way this would not be true is if:
1. You are not fitting the variance, but only the mean 2. There's some numerical issue (like some of the data are represented as integers, etc.)
Don't know if it could be one of those issues...
--Hoyt
On Mon, Apr 28, 2008 at 7:18 AM, Alan G Isaac
wrote: Hi Rich,
If your data is truncated at zero, it is not Gaussian (drawn from a normal). You will notice this when you shrink the range of values (unless the variance is tiny).
Cheers, Alan Isaac
_______________________________________________ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
-- +++++++++++++++++++++++++++++++++++ Hoyt Koepke UBC Department of Computer Science http://www.cs.ubc.ca/~hoytak/ hoytak@gmail.com +++++++++++++++++++++++++++++++++++
-- +++++++++++++++++++++++++++++++++++ Hoyt Koepke UBC Department of Computer Science http://www.cs.ubc.ca/~hoytak/ hoytak@gmail.com +++++++++++++++++++++++++++++++++++
On Mon, 28 Apr 2008, Hoyt Koepke wrote:
A Gaussian fit to data x should fit exactly as well as data fit to ax, a > 0, just with a variance a^2 times the original. The only way this would not be true is if:
Hoyt, This is what I expected, too.
1. You are not fitting the variance, but only the mean
That's probably it; I'm using the midpoint and the width as a surrogate for FWHM. However, by adjusting the FWHM for the Gaussian curve, and tau for the sigmoid curves when pH is to be plotted, I can achieve the results I need. Just why I don't really understand, but since this is a very small part of a large, complex, approximate reasoning model I'll happily remain ignorant as I move on to testing the rest of the code. Thanks to both you and Alan, Rich -- Richard B. Shepard, Ph.D. | Integrity Credibility Applied Ecosystem Services, Inc. | Innovation http://www.appl-ecosys.com Voice: 503-667-4517 Fax: 503-667-8863
On Mon, 28 Apr 2008, Hoyt Koepke wrote:
I may not understand what you are asking, Rich, but I'm not sure I agree with Alan. A Gaussian fit to data x should fit exactly as well as data fit to ax, a > 0, just with a variance a^2 times the original
My point was different. If you truncate at zero but your mean is at 50 and your standard deviation is 10, you will hardly notice the truncation. If your mean is at 14 and the standard deviation is 10, you will definitely see the truncation. That's what I understood Rich to be seeing. Cheers, Alan Isaac
participants (4)
-
Alan G Isaac
-
Alan Isaac
-
Hoyt Koepke
-
Rich Shepard