Kevin T. Kilty
Copyright © 1998, All Rights Reserved
For those readers who do not know what inverse theory is, perhaps a quick summary is in order. The term inversion in the world of geophysics means finding parameter values for a model from observed or experimental data. Meteorlogists use the term retrieval because inversion has a separate meaning. The parameters that a person finds through inversion are optimal in some sense. Because we ordinarily use a model with specified parameters to simulate a process through a forward calculation, people often refer to finding the parameters as inverting the data. In essence finding a mean value and standard deviation from sample data is an inverse problem. More commonly, though, when people use the term inverse problem, they mean something more complex than mere statistical calculation; for example, everyone would agree that finding the parameters of a differential equation or a function or functional in an integral equation are inverse problems. In fact, inverse theory most likely refers to the general methodology of doing the estimation, rather than the specific application itself.
Inverse theory is a discipline all its own, and its practitioners like to focus on the esoteric aspects of their subject. They love the complex and beautiful; they shun the simple and crude. Thus, a person who views finding a mean and standard deviation as a inverse problem is ridiculed as an inverse theory troglodyte. Exhaustive search often provokes ridicule as well. Anyone using it seems compelled to apologize, refer to their analysis as barely adequate, and conclude that a more complex method would lead to better results (See for example ref. 1).
Exhaustive search has advantages, though. When a problem involves only a few parameters, exhaustive search provides a complete view of parameter space; and provides an immediate idea of the parameter confidence intervals, and how parameters are correlated with one another. In 1978 I took a class in inverse theory from the late, redoubtable, Gerald Hohmann. As a class project I decided to examine the application of inverse theory to a problem in astronomy; and, because the problem was non-linear, I decided to use exhaustive search 2. Hohmann chided me at the time by saying that he learned more about astronomy from my paper than he did about inverse theory; and, of course he didn't learn any astronomy from my paper. However, this particular application makes a more interesting story now. It is a rare example of a problem in which the true parameter values were unknown at the time I wrote the paper, but are now known exactly.
Having accurate diameters of satellites of planets in our solar system is important for a variety of reasons, not the least of which is to test theories of their formation. However, diameters were difficult to measure accurately at one time simply because the satellites are far away and appear so small. Prior to space missions that could fly-by these satellites for a close look, astronomers used a variety of other means to make diameter measurements. One method put to use just prior to the Voyager missions making their way to Saturn, was to analyze light curves from lunar occultation of Saturn's satellites.
An occultation occurs when a nearby astronomical body eclipses another one that is farther away. For example, earth's moon often eclipses a distant star or planet. As the eclipse proceeds the amount of light coming to earth from the distant body diminishes steadily until the eclipse is total. On March 30, 1974 there occurred an occultation of Saturn by the earth's moon. Astronomers using the Mauna Kea observatory in Hawaii, observed the occultation of Saturn and its satellites through two large telescopes 3,4. They obtained light curves by counting the number of photons that arrived at a photomultiplier at the focus of the telescope in 10 millisecond intervals. The telescopes were 0.61m and 2.24m in diameter, which are so large that diffraction of the satellite images from the limited aperture was negligible. The moon was near first-quarter which provided an extensive dark limb for the occultation. The occultation occurred shortly after sunset with the moon quite high in the sky. Thus, in all respects this occultation was ideal for finding satellite diameters through analysis of the resulting light curves.
A light curve in this context means the time series of photon counts in 10 millisecond intervals beginning at a time when a satellite is fully visible above the limb of the moon, until some time after it has disappeared completely behind the moon. Thus, the light curve exhibits a decline in intensity or photon count throughout the event. The time taken to decline from full intensity to minimal intensity is related to the diameter of the satellite. However, there are other factors to consider. Figure 1 shows the geometry of an occultation, and explains the meaning of a few factors.
Satellites reflect light according to their size, the amount
of their image that is covered by the occulting body, and
according to the distribution of apparent brightness across their
surface. This last factor has a great effect on the diameter that
a person infers from the light curve. Most planets and satellites
are not uniformly bright from center to limb. One model to
describe this varying brightness is Minneart's law where the
brightness is proportional to
The parameter e is the angle between the observer on earth and a vector normal to the satellite's surface (see Fig. 1). Dead center on the satellite e=0, and the scattered light is a maximum. At the limb of the satellite e=PI/4, and the scattered light is zero. The angle a is that between the earth and Sun as seen from Saturn. Obviously if a is not zero, then the satellite is illuminated asymmetrically. Saturn is so far from earth that a is very close to zero, and its satellites appear almost fully lit at any time. At the time of this particular occultation the satellites each had a small crescent zone near one limb that was not lit; but, the resulting distortion amounted to less than 0.3% of any satellite's full diameter.
The parameter k in the Minneart law is called a limb darkening coefficient. If k<1/2 the limb appears brighter than the center of the satellite. At k=1/2 the satellite is uniformly bright like a flat disk. Lambert scattering corresponds to the case of k=1; and, a highly limb-darkened body is one for which k>1. There are other models of limb darkening, but this one describes ideal bodies well, especially for limb darkening coefficients between k=1/2 and k=1.0.
It is not easy to identify the duration of an occultation from the initial decline of the light curve to a point of complete eclipse; reason being that the event begins and ends gradually. It is more easy to identify a mid point for the event and produce a light curve from specified values of duration and k that best explains the observed light curve. Complicating this situation is the fact that the clock drives for telescopes run at a sidereal rate-the rate of the background stars. Thus, once the telescope is oriented to observe a particular satellite the dark limb of the moon slowly covers the field of view. This causes background light to increase with time.
Elliot considered the background brightness to increase linearly with time, and his photon counting plots for all satellites except Titan show exactly this. Titan, however, shows quite clearly a decreasing count before occultation and an increasing count afterward (see fig. 2 in Elliot I). Thus, two parameters may not be sufficient to completely describe background effects-a factor I will elaborate upon later.
Finally, to convert the duration of occultation to diameter of the satellite we need a few astrometric quantities, such as the distance from the earth to the satellite, the angular rate of the moon against the background sky, and so forth. Luckily these are known to great accuracy by other means.
In summary, occultation photon counts or light curves should depend on six parameters.
Elliot described his method as a least squares inverse procedure, seeking 6 parameter values by fitting data to a theoretical curve. However, the fitting wasn't done simultaneously for all parameters. Instead, Elliot found background light parameters from light curves that obviously precede and follow the occultation. The exact inversion method is unnamed but is a non-linear least squares program available at the time through the Cornell Computing Center (NLLS). The method was unstable if both k and duration were free to vary; so the strategy was to fix k arbitrarily and allow the program to find the best estimate of duration of the event by fit to the data. Elliot did this for a series of assumed values of k. Finally, Elliot considered photon counting to be his main source of random error, so he weighted photon counts by 1/count. Table 1, below, summarizes Elliot's results for satellite diameter and limb darkening coefficient.
|Satellite||Diameter (km)||Limb Darkening|
|Titan||5832 ± (53)||1.25 (+0.75,-0.25)|
|Iapetus||1595 ± (139)||0.75 ± (0.25)|
|Rhea||1576 ± (86)2||1.0 ± (0.5)|
|Tethys||1042 ± (115)||-NA-|
|Dione||825 ± (148)||-NA-|
1Elliot considered the 5832 km diameter of Titan as a minimum, despite the (() figure, because larger diameters actually provided a better fit.
2A lower limit of 1325km for the diameter of Rhea was available from its occultation of a background star in August 1974.
Since 1978 two Voyager missions have flown past Saturn, Voyager I in Autumn 1980 5 and Voyager II in Autumn 1981 6; and the Hubble space telescope has made additional observations of Saturn's moons. We have very accurate values for the diameter of the satellites. Table 2 shows actual diameters as obtained from the Los Alamos National Laboratory web page on Saturn.
|Dione||1120||Large, 2.5 perhaps|
Note:The value of k in this table is the one that would produce the requisite diameter using Elliot's model and analysis.
Keep in mind that Elliot regarded the small satellites as dusty or frost covered bodies that would scatter light according to Lambert's law, and would each have a limb darkening coefficient close to 1.0. He also regarded Titan as highly limb darkened, and would have rejected a value for k of 1/2 as too small. A cursory inspection of the occultation light curve in Elliot's paper shows that k=1/2 does not describe the situation at all.
I analyzed only the data for Titan because it is the most detailed light curve of the set. Rather than use a non-linear inversion method, I chose to calculate the misfit for the entirety of parameter space by brute force. Figure 2 shows two of the parameters, k and duration, with contours of the sum of squared residuals. The residuals are related to confidence intervals; in particular the 90% and 99% confidence intervals are equivalent to the contours labelled 580 and 600, respectively.
Note that contours are elliptical in a slightly oblique direction. This shows that k and diameter are correlated, but only slightly so. Therefore, a person should be able to find values for k and duration independently of one another. By choosing to call the point at the minimum residual my best fit, I found that a diameter of Titan of 5686 (114km and the limb darkening coefficient of 1.75 (.15. This is a substantially smaller diameter for Titan than Elliot found, and lies below his estimated uncertainty interval. My estimate of the mid-time of the occultation is 0.07seconds later than Elliot's, which is to say about 3.5% later when compared to the 1.94 second duration of the event.
My estimate of k is tighter than Elliot's because I used the entire light curve for fitting to the data. The toe and shoulder of an occultation curve are much more sensitive to the value of k than is the central portion of the curve. More important, if a person uses the entire light curve to estimate k, then k and duration are independent of one another. By using only the central portion of each light curve to estimate k, Elliot caused k and duration to become highly anti-correlated with one another. This contributed to the instability that prevented Elliot from using k and duration simultaneously in his fit; and shows an advantage of exhaustive search.
Figure 3 shows ideal light curves for varying values of k. An inspection of these curves show that the light curve has more shape at its extremes, which makes the extremes more sensitive to the value of k. Note, too, that in the central half of the light curve there is almost no difference between increasing k and shortening the duration. This explains the anti-correlation of these two parameters that plagued Elliot.
Why do Elliot's results fare so well for three satellites (Iapetus, Rhea, and Tethys) and fare poorly for the other two? The first issue is Titan, which is so large, and has so detailed an occultation light curve that the method should have worked well. It seems that there are three reasons for failure, here.
First, the model of a constantly increasing background light curve for Titan is inadequate. The data do not support it. In a letter that I sent to Elliot in the summer of 1979, I pointed out that the optimum value for k found from the data depends very much on the model of background light. That, in fact, reasonable variations in the background light model could push the derived diameter for Titan to values much greater than 5800km, which Elliot in fact preferred at the time 2,3, or to values much less than 5800km, which it turns out is correct. Geophysicists refer to inadequacy of a model as representational error. A person has to avoid it in the analysis a priori, because when there are many parameters in the model, correlated with one another to some extent, a person cannot detected it from an objective analysis of the data. Representational error is completely unlike random error.
Second, Elliot used only the central portion of the light curve to find k, which makes k and diameter highly correlated. Figure 3 shows there is almost perfect anticorrelation between k and duration using only the central half of the light curve; but, using the entire curve decorrelates the two almost completely. Thus, an improvement was possible if Elliot had used a method which considered the entire occulation light curve. Exhaustive search had something to offer in this instance.
Third, Titan has a visible atmosphere that rises well above its limb. Images from Voyager missions showed that the visible image of Titan is 400km in diameter larger than its solid surface 6. Thus, an estimate made from an occultation light curve ought to be about 5550km, which is very close to the size I determined. Titan's atmosphere actually extends 300 km above its surface. It is tenuous at this height but loaded with enough aerosol so that its light scattering may affect the occultation light curve. Add 600km to the true solid diameter of Titan, and the result is practically within the uncertainties of both Elliot's estimate and mine. The occultation light curve provides the diameter of the visible image, not necessarily that of the solid surface.
The smaller satellites are another story. They have no atmospheres to complicate the issue of their diameters, but they are very complex bodies. They have highly cratered surfaces that vary in brightness. Some of the brightness variation results from eruptive products being distributed unevenly on the surface. Additional brightness variation results from the way the satellites orbit Saturn. Some of these satellites present the same face to Saturn as they orbit. The side constantly facing into the orbit picks up debris, which gives it different optical properties than the trailing face. Iapetus is so extreme in this regard that the albedo of the forward face is about one-tenth that of the trailing face. Cassini claimed that he could see Iapetus only when it was to one side of Saturn 7,8.
Despite this complexity, Elliot's estimates of the diameters of each satellite except Dione were extremely good. Dione presents a very strange example. The leading hemisphere of Dione is heavily cratered and uniformly bright. The trailing hemisphere is very dark except for wispy material that brightens the center of the body. The image on the frontispiece of this paper shows the complexity of the trailing hemisphere. This bright wispiness is too subtle to resolve through the Moana Kea telescopes, but it was apparent in the occultation light curve. A bright center and dark periphery is precisely a highly limb darkened object. Thus, if Elliot had been willing to accept a value for k of 2.5, his estimate of Dione's diameter would have been very accurate. I have not performed the analysis myself, but exhaustive search may have had something to offer here, as well.