Re: Miller's errorbars

Surfer wrote:
On Tue, 22 Sep 2009 22:11:00 +0100 (BST), Tom Roberts
<tjroberts137@xxxxxxxxxxxxx> wrote:
Surfer wrote:
We have a key difference in that:
1) You calculated errorbars from Miller's data BEFORE removal of drift and noise, whereas,
Not true. I calculated the errorbars on his reduced data for each run, INCLUDING the "removal of drift and noise", USING THE METHOD MILLER USED.

I am not sure why you say "Not true".

I said "not true" because what you said is indeed not true.


You may have calculated the
errorbars "on his reduced data....", but to do so you used his raw
data from which drift and noise had NOT BEEN removed.

You need to READ THE PAPER. In it you will find a description of the errorbar computation. It is indeed, AS I SAID, the errorbars on his reduced data, the last line of his Fig. 8 (my Fig. 1), which he plotted to show the "signal".


But your errorbars are never the less calculated from the raw data.

That has been my descriptions on USENET, including this and other newsgroups, because the difference is negligible and it's much easier to describe. But the calculation in the paper is on his reduced data, appropriate for the algorithm he used.

For example, the data that were averaged to compute the
value in the first column of his last line consists of:
a) the 20 values at Marker 1
b) the 20 values at Marker 9 plus 1.5
The histogram in Fig. 4 has that 0.15 fringe added to the
Marker 9 values, and the sigma of that histogram is the
errorbar on this point because it is almost completely
systematic (see text). Even if one ignores the systematic
pattern of the variations, closes his eyes and holds his
nose (needed because this is incredibly ugly and stinks),
and divides the errorbar by sqrt(40), it is STILL larger
than the "signal" that Miller found.

Now look at Fig 4 and ask yourself: how important is
0.15 added to the values of marker 9 compared to the
variance of the histogram? -- NOT VERY (it's smaller
than the bin width). That's why I have not bothered to
explain this detail around here.


2) I believe its more reasonable to calculate errorbars from Miller's data AFTER removal of drift and noise.
But you must calculate them CORRECTLY, and you do not. Removing the drift introduces ENORMOUS correlations in the resulting data points, and your computation fails to take that into account.

I have been looking into that.

But it turns out that since linear drift is subtracted from full
turns, the effects of the correlations have a period of a full turn.
As a result they cancel out when half turns are averaged by the last
step of Miller's algorithm.

Not true. That would be true ONLY if the drift had a perfect period of 1 turn. It doesn't -- not even close.

You REALLY need to learn about error analysis; essentially random values like Miller's "drift" never cancel out. Just LOOK at the data to see how ludicrous your claim is.


You have argued that 2) is inappropriate because Miller
".....ASSUMED the background is linear; it isn't...."
You don't even get my argument right. I have repeatedly argued that (2) is inappropriate because IT IS NOT WHAT MILLER DID.

I don't think you wrote what you intended here. Miller calculated his
probable error AFTER using his data reduction algorithm to remove
drift and noise. That is exactly what (2) says is reasonable :-)

<sigh> Miller did not include the errorbar associated with his averaging when he claimed a "probable error" -- that errorbar completely dwarfs what he called "probable error". TODAY we know about this, and it is a standard and necessary component of an error analysis; in Miller's day it was not very common at all.


This is so very basic I don't understand why it is so hard for you to understand: in order to accept Miller's result, you must accept WHAT HE DID.

His data reduction algorithm differs from a modern DSP algorithm, but
I am satisfied that its overall EFFECT is sufficiently similar that
the difference doesn't matter.

They are not similar at all, and your claim is just plain wrong.

And I repeat: YOU MUST COMPUTE ERRORBARS CORRECTLY. If you apply a DSP filter, you must include the correlations that it produces among its output data, and that is problematical.


When the averaging is done last, the errorbars due to the averaging
are smaller than the signal.

Only when you compute the errorbars incorrectly.

And I repeat for the zillionth time: IF YOU WANT TO ACCEPT MILLER'S RESULT, YOU MUST ACCEPT WHAT HE DID. WHAT HE DID has errorbars much greater than the "signal" he found, which completely negates his results.


You are just repeating yourself and displaying your unwillingness to actually read my paper and understand it. Don't expect me to continue.


Tom Roberts

.

Relevant Pages

  • Re: Millers errorbars
    ... You calculated errorbars from Miller's data BEFORE ... INCLUDING the "removal of drift and noise", ... probable error AFTER using his data reduction algorithm to remove ...
    (sci.physics.research)
  • Re: Sagnac laser speedometer
    ... to Millers data to remove drift and noise before calculating errorbars. ... One must compute errorbars for WHAT HE DID, not for some other analysis technique. ...
    (sci.physics.relativity)
  • Re: Sagnac laser speedometer
    ... calculating errorbars. ... One must compute errorbars for WHAT HE DID, not for some other analysis technique. ... calculated AFTER noise and drift have been removed from the raw data. ...
    (sci.physics.relativity)
  • Re: Millers errorbars
    ... for his measurements this contribution completely dominates the total ... You calculated errorbars from Miller's data BEFORE ... INCLUDING the "removal of drift and noise", ...
    (sci.physics.research)
  • Re: Millers errorbars
    ... You calculated errorbars from Miller's data BEFORE ... from Miller's data AFTER removal of drift and noise. ...
    (sci.physics.research)
Quantcast