On 9/6/15 9/6/15 7:59 PM, Anton Shepelev wrote:
> Dayton Miller consistent-
> ly observed a small effect, unconnected with any
> earthly factors but well-correlated with the orbital
> motion of the Earth[...]
> If you will carefully read just these two documents
> you shall see that the "null result" is a falsifica-
> tion and that Shankland's analysis fails to explain
> Miller's positive result. They also tell why numer-
> ous other repetitions failed.
Miller's "signal" was statistically insignificant [#]. It is essentially his
instrument's (thermal ?) drift aliased into the same bin where any signal would
be, because he used a FLAWED analysis algorithm. It is sinusoidal and LOOKED
like a signal would look, because of the properties of his noise and the comb
filter he (inadvertently) applied.
In his day, before Shannon, he could not have known this.
But today, with modern DSP techniques it is obvious.
Moreover, back then they did not routinely perform an
error analysis. Today we do, and the basic error analysis
of his data shows that the errorbars are MUCH larger than
the variations in his data -- his "signal" is NOT
statistically significant [#]. And since he averaged his data
down to just a few frequency bins (in DSP-speak this is a
comb filter), the aliased ~ 1/f noise looks like a sinusoid
with frequency corresponding to the lowest bin (where any
signal would be).
[#] this is bulletproof, because we know how averaging affects
the errorbars on the mean.
> [quoting Einstein]
> I believe that I have really found the relation-
> ship between gravitation and electricity, assuming
> that the Miller experiments are based on a funda-
> mental error. Otherwise, the whole relativity the-
> ory collapses like a house of cards.
Miller's experiments WERE based on a fundamental error (actually several of
them). Einstein and relativity are vindicated.
See my paper about this:
http://arxiv.org/abs/physics/0608238
It discusses the above criticisms. It also applies a model of his instrument's
drift plus an orientation-dependent signal. The signal is consistent with zero
in all runs in my dataset, and exactly zero in all reliable runs. The drift is
HIGHLY nonlinear, negating the basic assumption of Miller's analysis algorithm.
I used some of Miller's original data sheets (amazingly, copies are still
available from the CWRU archives).
In a following article you said:
> Thomas J. Roberts must be famous, for others have
> used his article in arguments with me about the
> Miller experiments.
Probably because it is the only MODERN analysis of his data (I used copies of
his original data sheets from the CWRU archives).
> I
> wonder why he states that "because of the 180 degree
> symmetry of the interferometer, any real signal can
> depend only on orientation modulo 180 degrees," when
> a full-period, first-order effect was detected by
> Miller and explained by Hicks, whose calculations I
> have at hand should you need them.
The 180-degree symmetry is OBVIOUS in his drawings. And since his basic signal
was statistically not significant by a VERY large margin, I have no doubt that
the "full-period first-order effect" is likewise not statistically significant.
For instance, if a linear drift was assumed the whole thing is useless, because
the drift of that instrument is HIGHLY nonlinear.
Similarly, if data were averaged then the implicit comb filter and aliasing
destroy any usefulness in the result.
> To be concise, the interferometer can't be perfectly
> symmetrical
Sure. But this is PHYSICS, not any "perfect" system. The question is whether the
asymmetries are SIGNIFICANT. They surely are not.
Look at figure 3 of my paper -- I count 10 points above the line, 5 below, and 5
on it. That's hardly decisive.
Look at Figure 5 -- the errorbars are ~ 0.8 fringe, which is MUCH larger than
the difference point-to-line of any turn in Fig 3. So that IN-decisive
"half-wave" you think you see is NOT significant.
Amateurs look for patterns, professionals look at errorbars.
You see a pattern, I see errorbars that show that it is in your imagination, not
the data.
Humans are extremely good pattern-finding machines, and often
find patterns where none are present. Some people see "the
Virgin Mary" in water stains and burnt toast, while you (and
Miller) see a "signal" here.
> His other careless statement that Miller arbitrarily
> chose the linear model of the systematic drift is
> also wrong,
It was not "careless" at all -- LOOK at figure 10 of my paper -- the drift is
HIGHLY nonlinear, for every orientation.
> for in that same 1933 paper Miller ex-
> plains that from the many tests he found temperature
> effects to maintain a constant influence though each
> 50-second turn due to the high thermal capacity of
> the interferometer, and a constant derivative calls
> for a linear function, which may not affect the
> clarity of any harmonic signal.
Nevertheless an actual model of the drift shows it to be HIGHLY nonlinear (fig
10, and also figs 2 and 3). Look at figure 4 of my paper -- the variation in the
data at this particular orientation is CLEARLY dominated by the systematic drift.
I challenge ANYBODY to look at figure 2 (which is just Miller's raw data with
his adjustments restored), and claim that the drift is "linear"!
(at the level of ~0.5 fringe one could claim that for turns
1-3 and 14-18, but most turns are nowhere close to linear.)
> Therefore Roberts
> is wrong in calculating the error bars before sub-
> tracting the linear model. The error would have
> been several times lower if had done so after the
> subtraction, which may be thought an essential part
> of the measurement procedure. One should calculate
> the dispersion of the measured signal, which is the
> difference of the fringe shift from the linear tem-
> prerature drift.
I did it both ways while preparing the article. There's no significant
difference in that the errorbars are MUCH larger than the variation in the
means. IOW the drift is HIGHLY NONLINEAR, and subtracting a linear drift does
not affect much.
> Quoth Miller:
>
> The observation is a differential one and can be
> made with considerable certainty under all condi-
> tions. Disturbances, due to temperature or other
> causes lasting for a few seconds or for a few
> minutes, might affect the actual amount of the
> observed displacement and give less certain val-
> ues for the velocity of the ether-drift while, at
> the same time, the position of maximum displace-
> ment is not altered.
>
> I would add that rapid disturbances will be reduced
> during the averaging, as well as any additive noise
> originating outside the interferometer and therefore
> unperiodic in its turn.
Like Miller, you CLEARLY do not understand the TERRIBLE effects of averaging. It
completely hid from him what was really going on. Had he just plotted out my
Figure 2, he would KNOW that his assumptions were simply not valid. All you need
to do is LOOK AT FIGURE 2 to see your claims are nonsense.
To claim to find a "signal" in Fig 2, with amplitude ~ 0.1
fringe is just LUDICROUS.
[Had he been able to make the instrument perform like turns 1-3
all the time, then he might have had something. But out of the
~1,000 data runs that I have looked at, whenever he did so it
CLEARLY had no signal at all, just a smooth drift (see my final
two paragraphs below).]
> Neither does Roberts explain the way that his hypo-
> thetical instrumental error could produce a sinu-
> soidal pseudosignal with a constant galactic orien-
> tantion:
OF COURSE NOT! There is no significant signal. Yes, there are variations, and if
one INSISTS on interpreting them as "signal" then it must have some orientation.
> How can an instumental error track the time of day
> and the season of the year in precisely the way to
> neutralize them, maintaining a constant galactic di-
> rection in *all* the experiments over the years?
Experimenter's bias. He knew what he wanted to find. As do you.
Re-plot with errorbars, and you will find that there is a point in one
direction, with errorbars that span the sky!
> He is also wrong about Miller's averaging of half-
> turns as part of data analysis. Miller did no such
> thing and employed harmonic analysis to his full-pe-
> riod curves.
Just LOOK at my Fig 1 (Miller's Fig 8). You are wrong.
> Why does Roberts misinform us, presenting Miller's
> visualisation procedure intended for "a preliminary
> study of the observations" for the one he used in
> "the definitive study of the ether-drift effect?"
> The 1933 article describes both procedures and re-
> veals Roberts's confusion.
No "misinform" or "confusion". Miller's "signal" was not statistically
significant. Nothing can change that. Including your attempts to dismiss it.
> One of Miller's errors in Roberts's opinion is the
> averaging of data, so I wonder why he has retained
> the supposed averaging of the half-turns in his re-
> analysis, while Miller himself did not employ it.
READ THE PAPER. I did no averaging at all in part IV, my new analysis of 67 of
his original data sheets.
> I would he had
> practised as he preaches and had analysed the uncom-
> pressed, full-period measurements,
READ THE PAPER. I analyzed all 20 turns of each run, with no averaging at all.
> Might Roberts have intentionally analyzed the wrong
> algorithm to hide the uncomfortably small first har-
> monic?
I analyzed Miller's ORIGINAL DATA SHEETS. Not his "algorithm" (which I did
discuss in parts II and III of the paper, but did not use in any way in part IV).
The first harmonic, and any other harmonic, is IRRELEVANT in my analysis.
> Roberts claims to have found a model of the system-
> atic drift in Miller's experiment, but he has not
> revealed this model even for a single twenty-turn
> set so that readers might study it by themselves.
READ THE PAPER. This is my Fig. 10. The text describes in detail how it was
obtained. The algorithm has to be applied to each data run individually, because
the drift is HIGHLY idiosyncratic to each run.
Basically the signal must be the same at each orientation,
while the drift happens sequentially. The technique I
used cleanly and rigorously separates out the orientation-
dependent part from the data, presenting just the drift in
Fig. 10.
> I
> for one should subtract it from the raw data and
> look at the result,
OBVIOUSLY you did not read the paper. That is NOT the correct thing to do, and
not what I did.
I fit the sum of
an arbitrary orientation-dependent function
plus
the systematic drift (determined as described in the text)
to the data for each of my 67 runs. Ignoring 11 clearly unstable runs, the
result for the orientation-dependent signal is ZERO (Fig 11).
> because Roberts did it in fre-
> quency domain which is rather unituitive and con-
> tains quantization and other errors of the discrete
> Fourier transformation.
OBVIOUSLY you did not read the paper. Part IV is all in the "time domain"
(really orientation over 20 turns).
My discussion of "frequency domain" was in part III, where I used DSP techniques
to explain where Miller's "signal" came from (his noise), and why it looks
sinusoidal (it is 1/f noise aliased into the lowest frequency bin, where any
real signal would be; when one Fourier bin dominates, the time-domain plot looks
sinusoidal with that frequency).
> Roberts's only precaution against unintentional in-
> clusion of the real signal into the systematic drift
> is the subtraction of the first turn from all the
> other turns. It does not seem a statistically right
> operation when the noise is many times more powerful
> than the signal, as it is in our case.
You OBVIOUSLY do not understand. That enables me to separate the
orientation-dependent part of the data from the drift. And it does so
INDEPENDENT of whatever noise was present in the first turn.
> The noise
> may have conspired completely to mask any singal in
> this turn.
Doesn't matter, because:
A) the subtraction removes any orientation dependence REGARDLESS of
whatever noise was present in the first turn,
and
B) the fit will find any orientation-dependent, because it uses all
20 turns.
You REALLY need to READ THE PAPER and understand what I actually did. Your
GUESSES are wrong.
> Leaving the nature of the instrumental drift unex-
> plained and uncommented upon, Roberts's "explana-
> tion" seems but an abstract mathematical excercize
> with little physical value because it offers no in-
> sight into the physical cause of this drift.
You REALLY need to READ THE PAPER and understand what I actually did. Your
GUESSES are wrong.
Yes, my analysis does not identify the source of the drift. But it DOES extract
whatever orientation-dependent signal is present in the data, and that signal is
ZERO.
THINK ABOUT IT: no analysis of his data sheets can possibly
identify sources of drift -- that requires examining and
operating the instrument and its environment.
>> "Michelson and Morley (1887) analyzed their data
>> using a data reduction algorithm quite similar to
>> Miller's, and therefore their result suffers from
>> the same serious flaws discussed above. They did,
>> however, have a smaller systematic error, and they
>> contented themselves with putting an upper bound
>> on the earth's speed relative to the ether of 7.5
>> km/s."
>
> Which is close to Miller's result but somewhat
> smaller because the experiment was performed closer
> to the ground.
LOOK at figure 12 of my paper. M&M's "signal" is not statistically significant,
either. Again, these errorbars are an UNDERESTIMATE, as they include ONLY the
portion due to the averaging; they are therefore unassailable.
Your last sentence is just factually wrong: much of Miller's data was taken IN
THE BASEMENT at CWRU (but in a different basement than Michelson and Morley used
30 years earlier). Later a special "interferometer house" was built with glass
windows and the interferometer at ground level (in Cleveland). Later still he
made the heroic trek to carry it up to a lab near the top of Mt. Wilson (which
is probably what you are thinking of). Then he brought it back to CWRU and took
still more data. The basement of the physics building at CWRU has special
pillars for the interferometer -- they are isolated from the building and go
down to bedrock. Miller was, of course, the chairman of the physics department
when the building was built.
After writing the paper, I was of course aware that one of the major problems
with Miller's technique was his "adjustments" to the interferometer to keep the
central fringe visible [#]. Basically the drift was too large. When I was at
CWRU (presenting a colloquium on this), Prof. Fickinger and I spent a few hours
in the archives selecting data runs which had no adjustments. We had time to
look at more than half of his >2,000 data sheets. That criteria of course
preferentially selects runs with small drift. To no great surprise, most of
these runs clearly had no "signal" at all. Many of them were simple progressions
like
+1 +1 +1 ... +2 +2 +2 ... +3 +3 +3 ... +4 +4 +4 ...
for all 20 turns. Many had a total drift less than one fringe, and all had
drifts less than 3 fringes.
Miller never mentioned these runs with no signal (probably because they were
only a few percent of his runs [@]). Nor do the people like you who so
desperately want his "signal" to be real. But it isn't.
[#] He would place a small weight on one arm to distort the
instrument and move the central fringe toward the center of his
field of view.
[@] This is more a statement about the (lack of) stability of
his instrument than anything else.
Tom Roberts