Fig. 1. The curved starting line of a racecourse in Corinth, ca. 500 B.C.. (Courtesy American School of Classical Studies, Corinth Excavations.)
Fig. 2. The archaeological team from the University of Pennsylvania with the instruments used to measure the data points. One of the authors (Romano) is in the center behind four graduate assistants (left to right: Chris Campbell, Mimi Woods, Elizabeth Johnston, and Ben Schoenbrun). The cliffs of Akrocorinth loom in the background and a portion of the starting line is barely visible in the foreground. (Courtesy Irene Bald Romano.)
Fig. 3. The curved starting line and the 21 data points along its front edge.
|
Data Point |
x-coordinate (meters) |
y-coordinate (meters) |
|
01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 |
19.880 20.159 20.676 20.919 21.171 21.498 21.735 22.810 23.125 23.375 23.744 24.076 24.361 24.597 24.888 25.375 25.166 25.601 25.979 26.180 26.142 |
68.874 68.564 67.954 67.676 67.379 66.978 66.692 65.226 64.758 64.385 63.860 63.359 62.908 62.562 62.074 61.292 61.639 60.923 60.277 59.926 59.524 |
Table 1. Coordinates of 21 data points along the front edge of the starting line.
Fig. 4. The two circles determined by the two methods. Also plotted are the centers of the 680 triplet circles used to determine the three-points circle.
Fig. 5. The surface determined by the least-squares error in Eq. (3).
Fig. 6. Points A and B are known to lie on the circle. The location of midpoint C is uncertain.
Fig. 7. Probability distribution of the radius of the triplet circle in the simplified example.
Fig. 8. Histogram of the radii of the 680 triplet circles using 5-meter cells.
|
Starting Position |
x-coord (meters) |
y-coord (meters) |
Angle to Next Position |
|
01 02 03 04 05 06 07 08 09 10 11 12 |
20.050 20.617 21.168 21.706 ... 22.756 23.354 23.975 24.573 25.122 25.612 26.214 |
69.034 68.426 67.825 67.149 ... 65.723 64.885 63.990 63.042 62.215 61.407 60.421 |
0.878 ° 0.860 ° 0.913 ° 1.871 ° ... 1.088 ° 1.151 ° 1.184 ° 1.049 ° 0.999 ° 1.220 ° |
Table 2. Coordinates of the center points of the front toe grooves of 11 of the 12 starting positions on the starting line. The last column is the angle subtended by two successive toe grooves with respect to the center of the least-squares circle (-20.940, 33.618).
Also visible in Fig. 1 are the front and rear toe grooves of twelve starting positions for the runners cut into the starting line. We wish to determine the angles between the center of the arc and the twelve starting positions to see what angular units might have been used at that time.
To determine the center and radius of the starting line, measurements were taken along its front edge at 21 points where the edge was fairly well defined [2]. These measurements were taken using an Electronic Total Station (Fig. 2), which includes an electronic theodolite and an electronic distance meter. Figure 3 shows the locations of the 21 points on the starting line and Table 1 lists their 
-coordinates with respect to an arbitrary coordinate system.
Of course, it is not possible to fit a circle exactly through these 21 points due to many cumulative errors. These include errors made by the original surveyors in laying out the starting line; errors made by the stonecutters and builders of the starting line; shifting of the starting line over two-and-a-half millennia; errors in locating points to measure due to erosion and other damage along the edge; and errors in the measuring instruments. Below we discuss two methods for determining the "best" circle that fits the data.
2. Three-Points Circle. The approach taken by the archaeologists who measured the data points is based on the fact that if only three points were measured along the starting line, then by any (reasonable) criterion the best circle would be the exact circle passing through those three points. Consequently, they first determined the exact circles passing through various triplets of the 21 data points. The -coordinates and -coordinates of these "triplet" circles were then averaged to determine the - and -coordinates of the center of a final circle, and the radii of the triplet circles were averaged to compute the radius of the final circle. The triplets were chosen to be all triplets that have at least two other data points between any two of them using the ordering in Table 1--a criterion that resulted in 680 triplets. The average center of the resulting 680 triplet circles was found to be at (-22.943, 32.506) and their average radius was 56.242 meters. The range of their 680 radii was [27.643, 160.668] and their sample standard deviation was 12.058 meters. The resulting "three-points" circle is plotted in Fig. 4, together with the centers of the 680 triplet circles.
3. Least-Squares Circle. We next discuss a method of finding the best circle that fits the 21 data points using a least-squares criterion. This method determines the maximum-likelihood circle assuming the coordinate errors of the measured data points have identical normal distributions with mean zero. The resulting circle is the one for which the sum of the squared orthogonal distances of the data points to the circle is a minimum [3].
In general, let 
be the coordinates of 
points imperfectly measured along a circle whose center and radius are to be determined. Let 
be the center of an arbitrary circle in the plane and let 
be its radius. The least-squares error associated with the maximum-likelihood circle is then
| (1) |
|
We need to find values of , and that minimize 
. By setting the derivative of with respect to equal to zero, we quickly find that
| (2) |
|
That is, for any given center the corresponding optimal radius is the average of the distances of the points to that center. By substituting (2) into (1), the problem reduces to minimizing the following function of two variables
| (3) |
|
This problem does not appear to have a closed-form solution. Using numerical techniques (the "fmin" function of MatLabĒ [4], which implements a Nelder-Mead simplex search [5]) the center of the circle determined by (3) was found to be (-20.940, 33.618) and the corresponding radius determined by (2) was 53.960 meters (Fig. 4). The 21 data points encompass an arc of 12.134 ° with respect to its center. The range of their distances from the center is [53.938, 53.991] and their sample standard deviation is 0.0133 meters.
In Fig. 5 we show the surface determined by Eq. (3). This least-squares error surface has a long narrow valley that points toward the center of the arc determined by the 21 data points. The valley reveals the small sensitivity of the location of the center point along a line perpendicular to the arc subtended by the 21 data points. At the center of the least-squares circle (the global minimum of the surface) the value of 
is 0.0036 and at the center of the three-points circle, slightly down the valley away from the data points, the value of is 0.0084.
4. Comparison of the Two Methods. From Figures 4 and 5 we notice that the centers of the 680 triplet circles lie along the valley of the least-squares surface. Thus both methods agree as to the location of a line along which the center of a best circle should lie. But the radius of the three-points circle is 2.282 meters longer than that of the least-squares circle--a significant difference if one were searching for physical evidence of the center in the stadium.
To suggest the reason for this difference, let us look at a considerably simpler problem. Suppose we wish to determine the radius 
of a circle on which two points 
and 
lie whose positions we know (Fig. 6). Let the distance between these two points be 2
. We imperfectly locate the midpoint 
of the arc between and by measuring the distance 
along the perpendicular bisector of the chord . Specifically, we take measurements 
of . Suppose that is a normal random variable with standard deviation 
whose mean 
is the exact distance to . A little geometry then shows that the exact radius of the circle is
| (4) |
|
Taking the average of the measurements of as an estimator for , we then obtain the following estimate for the radius of the circle:
| (5) |
|
This procedure is equivalent to finding the least-squares circle that passes through and determined by the data points.
The analog of the three-point method would be to first compute the radii of the circles through all triplets consisting of the points and and one of the measured points. Then these radii would be averaged to estimate the radius of the final circle. Now, if .gif)
is the normal distribution function of the random variable , then the distribution function .gif)
of the radius of the triplet circle is
| (6) |
|
Actually, this distribution function is correct only if the support of lies in the interval [0, ]. We assume that the mean and standard deviation of are such that this is approximately the case.
Figure 7 is a graph of when = 0.0740 meters, = 0.0133, and = 2.8240. These values where chosen so that: (1) the radius of the exact circle is 53.960 meters, the same as the least-squares radius of the starting-line circle, (2) the points and subtend an arc of 6 °, about half of the 12.134 ° arc subtended by the original 21 data points, and (3) the standard deviation is equal to the sample standard deviation of the distances of the 21 data points to the least-squares center of the starting-line circle. The mean of using these values is 55.910 meters (computed numerically). This compares favorably with the radius, 56.242 meters, of the three-points circle. The distribution function in Fig. 7 likewise compares favorably with a histogram of the 680 triplet circles shown in Fig. 8.
In summary, the nature of the sources of the errors in this problem suggests that the measured coordinates of the data points should be regarded as normal random variables. This, in turn, implies that the least-squares circle would best locate the center employed by the ancient surveyors. The three-points method overestimates the true radius because when the errors in the three points used for a triplet cause them to "line up" they yield a circle that can have a significantly larger radius than the true radius. This is evident in the way that the distribution in Fig. 7 and the histogram in Fig. 8 are skewed to the right.
5. Angles of the starting positions. As mentioned in the introduction, a primary interest for determining the center of the starting-line circle was to determine the angles subtended by the twelve starting positions of the runners. The measurements of the coordinates of eleven of the front toe grooves are given in Table 2. The toe groove for position 05 was in too poor a condition to be measured accurately. The last column gives the angles between the starting positions with respect to the least-squares center of the starting-line circle.
The average angle between the twelve starting positions is 1.019 °. Although the range of the angles is rather large, this average angle is sufficiently close to one degree to warrant attention. Among the Greeks, the earliest known use of the degree as a unit of angular measurement is found in the writings of Hypsicles in the second half of the second century B.C. [6]. He adopted the Babylonian practice of dividing the twelve signs of the zodiac into 30 equal parts [7]. The choice of 30 is probably because the sun takes about 30 days to pass through each sign of the zodiac. That is to say, the division of a circle into 360 parts can probably be traced to the fact that there are roughly 360 days in a year. It may be that in laying out the starting positions, the Corinthians used the angle traveled by the sun along the zodiac in one day as the angle allotted each runner. If so, the Corinthian starting line is evidence of an early use of the degree as a unit of angular measurement among the Greeks.
6. Acknowledgements. The archaeological field survey on which this study was based is the Corinth Computer Project of the University of Pennsylvania Museum of Archaeology and Anthropology, David Gilman Romano, Director. The project is run under the auspices of the Corinth Excavations of the American School of Classical Studies at Athens, Charles K. Williams, Director. During the years 1989-91, Mr. Benjamin Schoenbrun was Research Intern for the Corinth Computer Project and assisted with the analysis of the staring line data. I (D. G. Romano) am grateful to all of the University of Pennsylvania students who assisted me both in the field and in the laboratory.
The Corinth Computer Project is supported by the Corinth Computer Project Fund of the University of Pennsylvania Museum of Archaeology and Anthropology.