Sei sulla pagina 1di 10

Why Does Higher Q Make Better Clocks?

John Haine1

Introduction
Figure 1: Accuracy vs. Q (after Bateman)
1.00E-03 1.00E-04 1.00E-05 1.00E-06
Accuracy

1.00E-07 1.00E-08 1.00E-09 1.00E-10 1.00E-11 1.00E-12

High Mid Low Quartz Pendulum

Figure 1 is my version of a famous graphic produced originally by Bateman, which plots clock accuracy against resonator Q over 4 decades. Matthys reproduced Batemans diagram in his article in HSN 2010-5 [1]. The original includes many more measured data points (over many more decades), and as Matthys points out shows excellent empirical evidence that higher Q resonators make for more accurate clocks. No reasons for this high degree of correlation between Q and accuracy have been given in the horological literature, and some regard it as spurious.

The purpose of this paper is to propose a simple mechanism whereby higher Q makes for better Q accuracy, which places horological oscillators in this continuum alongside electronic and atomic clocks. In essence, I suggest that clock accuracy is governed by the effect of noise, largely generated within the escapement and the train, which is filtered by the pendulums frequency response, and produces timekeeping errors by causing small phase errors in the impulse timing. Higher Q means less noise and lower sensitivity of the escapement according to Airys theory, and hence better timekeeping. Its worth making a few observations on the diagram. The solid line is plotted through a couple of points (triangles) that correspond to good pendulum clocks, in particular the Shortt clock. The slope of this line on a log/log scale is -1, so accuracy seems to be inversely proportional to Q. The dotted lines are at plus and minus one order of magnitude, and they seem to encompass most of the points in the original graph (with some outliers). So whatever phenomenon is affecting accuracy seems to be pretty universal. One assumes that all the observations have been corrected for systematic errors such as temperature drift, barometric pressure etc. What is left is an apparently random effect, probably essentially due to some noise mechanism. Of course, it is also true that higher-Q clocks tend to be better built and regulated and kept in better controlled conditions. The original graph added an alternate vertical scale essentially the reciprocal of the accuracy. So for example 1 s/day corresponds to a little over 10-5; 1 s in 1000 years is slightly better than 10-10. Whilst its easy to assess accuracy to a second a day, obviously a second in a thousand years has to be extrapolated from shorter-term measurements. This means that any theoretical approach has to include an assessment of how errors add with time. Finally, Batemans full graph included a wide range of resonator technologies, from balance wheels to atomic nuclei, all fitting broadly within the same band. This surely indicates some common factor at work, and it ought to be possible to infer something about possible error mechanisms in pendulum clocks from what is known about other types. In particular, noise in electronic oscillators and how it depends on resonator Q is well understood.

john.haine@ieee.org

Response of a pendulum to noise


From my previous article in [2]2, the response X() of a pendulum in the frequency domain to an applied force F() is given by:

Thus the frequency response of the pendulum is:

Thus, if the resonant frequency of the pendulum is

In practice clock pendulums have rather high Q which equates to a small bandwidth. We can then substitute for the quantity 0+ where is the deviation of frequency from the centre frequency to get:

using the fact that the high Q ensures that the expression is very small except when is small, and the identity Since the noise variance is important we need to know not the amplitude of H but its squared modulus, which is so

(Equation 1). In the second expression the normal resonant frequency has been substituted for the radian frequency and the normal frequency difference of Hz for , since noise density is expressed per hertz not per radian per second. The value of H reduces by when

giving the well-known -3dB bandwidth of

. Equation 1 converts the bandpass response of the resonant pendulum

into an equivalent lowpass response which is a very good model as long as Q is high (which is the case for interesting pendulums), and is easier to work with.

In this article I generally use the same symbols as in the previous article.

Now suppose that the forcing function X of the pendulum is white noise with spectral density of variance N (newton2/Hz). The spectral density function of the noise variance of the pendulum is found by multiplying the noise density N by the expression in Equation 1 above:

(Equation 2) To calculate the total noise variance of the pendulum motion this has to be integrated over the full frequency range:

A standard derivative [3] is

Thus

Using this:

Finally we have to recognise that the narrow-band approximation that is used above works only round the positive resonant frequency strictly we must include the noise which is effectively at negative frequencies, so this expression has to be doubled, to get the overall noise variance:

(Equation 3). The expression has been written in this form since it contains four significant components. The pi is just a constant that comes out of the integration. The fraction is the square of the scaling factor which relates the equilibrium amplitude of a pendulum with loss factor k to an applied force at its resonant frequency. N of course is the noise density; and is the 3 dB bandwidth of the pendulum, so their product is the power passed by the pendulums frequency response. Later on well need the r.m.s. variation rather than the variance, which is just the square root of this expression.

Effect of noise on timekeeping


Woodward [4] has presented an analysis of the effect of support noise on pendulum timekeeping. As far as I can see his analysis concentrated on how the noise influenced the oscillation of the pendulum itself, ignoring any possible effects on the 3

whole oscillator (which has to include the escapement). In what follows I am not going to repeat Woodwards work but focus on the effect on the complete oscillator. The noise in the pendulums motion is concentrated in a very small band either side of the resonant frequency, so-called narrow band noise. The pendulum also has its own natural motion, assumed to be sinusoidal, at the resonant frequency. Lets assume that this motion has amplitude a so it can be described by:

The narrow band noise signal can be represented by the sum of two uncorrelated components with equal variance, one of which is in phase with this carrier and the other in quadrature. This narrow-band noise model is commonly used in electronics [5]. From [5], the two components have equal variance, each of which is equal to the total variance, so the noise can be written as:

Where i and q are the inphase and quadrature components of the noise respectively. This models the noise, which is filtered by the pendulum into a narrow band of width and q, each of bandwidth centred around the resonant frequency, as two lowpass noise signals i

, and each modulated by the cosine or sine carriers. The i and q vary very slowly for

example for a seconds pendulum with a Q of 1,000 they have bandwidth 2.5 x 10-4 Hz. Figure 2: Oscillator model
Pivot

Bob

In order to assess the effect on timekeeping we need a model of the whole oscillator as was presented in [2] and shown in Figure 2 (left). Here the pendulum position is sensed by a zero-crossing detector (for example a photodiode), which drives an impulse generator that kicks the pendulum by a small amount in the direction of movement at each zero-crossing, adding an increment of momentum sufficient to overcome the energy loss in the previous half-cycle. Though this is a very electronics oriented model, in my belief it quite accurately represents an idealised mechanical escapement as well, and if necessary it is easy to add for example hysteresis to the detector to better model the escapement. The zero-crossing sensor actually sees a signal which is the sum of the pendulum position and the filtered noise. Clearly the in-phase noise i must have a zero crossing at the same instant as the signal and therefore cannot influence the impulse timing; but at this moment of time the quadrature noise q must be at its peak and will affect the timing. Figure 3 illustrates the effect.

Impulse generator

Zero-crossing sensor

Figure 3: Signals entering zerocrossing sensor Normally the zero-crossing would be detected when the normal swing amplitude was zero however because of the slowly varying noise component (in this example shown grossly exaggerated) the actual sum of the two signals is not zero until the phase is about 7 degrees and so the impulse phase to the pendulum will be shifted by 7 degrees. Because the noise is very small compared to the wanted signal, the induced phase shift is also small, and we can assume that around the zero crossing the noise component is constant. To calculate the phase error we note that around the zero crossing the signal time dependence is sinusoidal and for small angles:

The impulse time shift

produced by the noise can then be found from the equation:

Thus

To find the impulse phase shift this produces we need to divide by the period and multiply by 2, which gives a phase shift

In practice we will be interested in the values of the phase shift at specific instants of time which are a multiple of the pendulum period t0, for which we could write:

Next we have to determine the time error this produces. The peak velocity of a pendulum swinging with amplitude a at frequency is just . Its stored energy is given by:

From the definition of Q the energy lost per cycle is:

Suppose the impulse produces a velocity increment v, then the corresponding energy increment can be equated to the energy lost in equilibrium, giving the equation:

Figure 4: Phase plane plot of pendulum impulse.

a/Q
a/Q D A a

The simplest way to find the resulting pendulum phase shift is to use a phase plane plot of the pendulum motion, on which the displacement is plotted on the x-axis and the velocity divided by on the y-axis. Figure 4 shows the normalised increment of velocity AB which occurs at an angle before the zero crossing (which without phase error would occur when the point A passed through D). AB can be resolved into a radial component AC (which is the amplitude increase) and a phase step BC. The angle which BC subtends at the origin O is the amount by which the actual pendulum phase is changed by the impulse phase error. As long as the Q is large, so that AB is very small compared to the equilibrium amplitude a, it is evident that the pendulum phase change is given by:

However, we have from above:

So the phase error inserted into the nth period of the pendulum is:

The time variation of this period will be (dividing the phase variation by 2 and multiplying by the period):

In practice of course we dont know the actual value of the noise (its random after all) but only its statistical variance (or equivalently its r.m.s. value). The last equation shows the amount by which the noise is scaled to produce the time variation thus it also predicts the amount by which the r.m.s. value of the noise is scaled to produce the r.m.s. time variation of a single cycle. Substituting for the r.m.s. noise from equation (3) above, the r.m.s. variation (Greek tau) of the period is:

To find the relative time error we divide this by the period, that is multiply by 6

, to get:

The factor

under the square-root sign is effectively the noise to signal ratio. This equation shows that the r.m.s. time . In itself this

error of a single cycle due to escapement error induced by noise is essentially inversely proportional to demonstrates why it is that higher Q makes a better clock.

Addition of time error over successive cycles


The time indicated by a clock regulated by our noisy pendulum will be the sum of a succession of time periods, each one slightly different, with an expected time variation as calculated above. To compute how the time error grows with time we need therefore to add up the variances of the individual periods. If the error from one cycle to the next is uncorrelated, then it is well known that the variance of the sum of n successive cycle periods grows as . However, errors generated by the mechanism analysed here are not independent - there is a high degree of correlation between the period of successive cycles and the cumulative time error may grow significantly faster than . The variance of the sum of n random variables is computed as follows [6]. We need to know the variance of each variable and also the covariance between each pair of variables. The variance is the mean of the square of the deviation of a variable from its mean. The covariance between a pair of variables is the mean of the product of their deviations from their individual means. Then we build the covariance matrix, or table, like this (shown for 5 variables).

Then the variance of the sum of the 5 variables is just the sum of all these 25 covariances. The terms on the leading diagonal are the covariance of the variables with themselves in other words just the individual variances. Terms placed symmetrically either side of the leading diagonal such as and are equal because it doesnt matter in which order two numbers are multiplied. If the variables are completely uncorrelated, all the covariances are zero, so the variance of the sum is just the sum of the variances. If the variables are completely correlated, all the terms would be equal, and the variance of the sum would be 25 times the variance of only one. So we need to find the covariance matrix for the errors in successive cycles of the pendulum. This is made easier by the fact that all the periods have the same mean and variance, so the terms on the leading diagonal are equal. In this case, if all the covariances are normalised by dividing throughout by the common variance (the leading diagonal), the leading diagonal terms becomes unity and all the other terms become the correlation coefficients between the variables. The correlation coefficient measures the similarity between two variables if it is unity they are essentially the same; if zero they are totally uncorrelated; if -1 they are negatives of each-other. So we need to find the correlation function that describes how the variation in the first period of a sequence of n periods is correlated with the second, the third, the fourth, and so on up to the nth. This is described by the autocorrelation function of the distribution of the periods and this has the property that it forms a Fourier Transform pair with its spectral density. We derived above the spectral density of the noisy part of the pendulums motion in terms of the offset frequency : from the resonant

This signal causes the phase (time) variation cycle-to-cycle through its interaction with the impulsing function; and the effect of the impulse function is to sample this noise at to produce a noise signal which is the time variation of the period, but which is centred on zero frequency rather than . So the form of the spectral density of the time variation is:

K just being a constant which mops up Q, Nh etc. A standard Fourier Transform is [7]:

The frequency-domain function here is very close to what we want, and with a bit of tedious algebra the autocorrelation function we need is just:

R(t) has the same form as the decay of the amplitude of a pendulum with quality factor Q, which indicates the time scale on which we would expect to see similarities in the noise oscillations such as shown in the simulation plotted above. Using this function we can construct the covariance matrix for n successive periods of the pendulum as follows. Obviously for t=0, R(t) = 1, so as expected the leading diagonal will be unity. Second, for two successive periods (such as the first and second), t is just the pendulum period, which is correlation becomes: . Then the

In general,

So if we represent

as say y, then the correlation matrix for n successive periods can be written: 1 y y2 y3 . . . . . yn-1 y 1 y y2 . . . . . . y2 y 1 y . . . . . . y3 . y 1 y . . . . . . . . y 1 y . . . . . . . . y 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 y yn-1 . . . . . . . y 1

It immediately follows that if Q was infinite, so y was 1, the variance of the sum of n successive periods would grow as n2, so the r.m.s. value would grow as n. On the other hand if Q was very low so y was essentially zero, the variance would grow as n and the r.m.s. value as root-n. Working with this matrix gets very unwieldy for large n, but its regular form makes it possible to calculate a closed-form expression for the sum of its terms:

where

Figure 5: Error growth function, Q=10,000


1E+20 1E+17

1E+14 1E+11 1E+08


1E+05 1E+02 1E-01 1E+00 n squared S(n) n

This interesting function is plotted against n in Figure 5 for a Q of 10000, along with plots of n and n2, on a log-log scale. For small n it grows as n2 but flattens off to grow as n, the break-point visually occurring at about n = 2Q/. In fact, taking the limit as n tends to infinity, it can be shown that:

for large n.

Thus, for periods of up to about oscillations of the pendulum, the r.m.s. n error due to the noise effects analysed here can be expected to grow approximately proportionally to time 2 (that is, the variance growing with n ). Over much longer periods however, the error growth tends towards the square root of
1E+02 1E+04 1E+06 1E+08 1E+10

time (that is, the variance growing with n) but with a constant multiplying factor the error of a single cycle produced above we then get:

. Applying this to the equation for

which is inversely proportional to Q.

Discussion and conclusions


Previous work on noise effects has focused on support vibration. Philip Woodwards articles [4] analysed the effect of seismic noise causing movements of the support point. Such movements manifest themselves as a force on the pendulum which is proportional to its mass, and in effect the movement produced is multiplied by Q. This multiplies the above equation by Q (times a constant), making the long term error independent of Q though the short term error grows as 1/root-Q. This may have helped to create a general view that Q wasnt a factor in long-term stability. However of course there are many other potential sources of a noise force that may affect the pendulum. I believe that the main source of noise in many clocks may be the impulsing mechanism and the train. Considering for example a dead-beat escapement, for most of the pendulums cycle a point on the scape wheel is scraping on a rest face of the pallet, and surface irregularities in the latter will directly lead to a noise force applied to the pendulum via the crutch. Once unlocked, there is a different force applied but again by a point scraping on a face which can create noise. Another factor is noise generated by the train. Prior to the point where the train is unlocked, the wheels are obviously stationary. Moon and Stiefel [8] present experimental and theoretical research on chaotic motion in clocks: essentially once the escapement is unlocked, rather than the train silently and smoothly starting to move, it does so rather irregularly and jerkily, 9

again creating a noisy force which is transferred to the pendulum. Electromagnetic methods of impulsing based on for example photoelectric sensing should be much better. Using the methods described by Moon it would be possible to measure noise effects in escapements to test the theory presented here. Poised for its next swing at an instant of time, the entire past history of a pendulums motion is encoded into its current deflection. The period taken to reach the equivalent point in the next cycle will depend on whatever random factors influence its swing and the exact timing of any impulse it receives. There is no inner flywheel telling the pendulum what time it ought to keep and applying gentle corrections. Once the clockmaker has removed all systematic factors the only pendulum characteristic which can reduce the impact of random disturbances is the pendulums ability to filter out noise components outside a very small band around the resonant frequency. The higher the Q of the pendulum, the narrower this band is and the smaller will be the long-term error. The theory above predicts that for a given noise level, the long-term r.m.s. time error will be in inverse proportion to Q, which corresponds to one of the main features of Batemans graph with which we started.

References
1. 2. 3. 4. 5. 6. 7. 8. R. Matthys; Q, Again; Horological Science News, No. 5, 2010 J. Haine:Some thoughts on disturbances to pendulums; Horological Science News, No. 2, 2011 See http://www.wolframalpha.com/input/?i=differential+of+arctan%28ax%29%2Fa Woodward on Time: p146 et seq; BHI 2006 Mischa Schwartz; Information Transmission, Modulation, and Noise; McGraw-Hill, 1970; chapter 6. http://en.wikipedia.org/wiki/Variance#Sum_of_correlated_variables See for example http://www.mas.ecp.fr/Personnel/lilla/classes/image_processing/pdf/FourierTransformPairs.pdf http://rsta.royalsocietypublishing.org/content/364/1846/2539.full.pdf

10

Potrebbero piacerti anche