200 Years of Least Squares Method

A. Abdulle and G. Wanner Geneve November 6, 2001

Exactly 200 years ago, the exciting discovery of a new planet established Gauss' celebrity and marked the beginning of the least squares method. This method is today an essential ingredient of computation in a great number of scienti c areas. At the same time, Hegel `proved' by philosophical arguments, that such a planet could not exist. In this paper, we want to retrace the original ideas and computations of these events. The paper terminates with some nice examples of today's applications.

Abstract

1 The discovery of Ceres

The discovery of Ceres , the rst of the asteroids between Mars and Jupiter, was the great scienti c event at the beginning of the 19th century with important consequences for the later development of science, despite the fact that this feeble light-point has hardly been seen ever since by non-specialists, and thus had absolutely no immediate `practical' importance.

1.1 The Rule of Titius-Bode

:::

Das Daseyn dieses Planeten scheint insbesondere aus einem merkwurdigen Verhaltniss zu folgen Sollte der Urheber der Welt diesen Raum leer gelassen haben? (J.E. Bode, Anleitung zur Kenntniss des gestirnten Himmels, 6. Au ., Berlin 1792, quoted in Hegels Werke 5, Anmerkungen p. 810)

The astronomer priests of Babylon discovered seven distinguished celestial bodies: rst the Sun and the Moon, then Venus (1600 B.C.), nally Mars, Mercury, Jupiter, and Saturn. Soon, days were alternatively consecrated to these divinities (Sunday, Monday, Mardi, Mercredi, Jeudi, Vendredi, Saturday) and since thousands of years all human activity on the earth pulsates in this 7-days rhythm. All this time, nothing was added to these Babylonian Gods, until Sir William Herschel, a German organist and amateur astronomer living in England, discovered the 13th of march 1781 a new planet, through a huge telescope of his own construction. Herschel wanted to name `his' new planet Georgium sidus, in devotion to the British King, but Bode's proposition Uranus (in Greek mythology the father of Saturnus), was felt less patriotic and became generally accepted. 1

The discovery of Uranus also revived the discussions about the formula of Johann Daniel Tietz (Titius) and Johann Elert Bode, which stated that the semi-major axis of the orbits of the planets were given by the rule 0:4 0:4 + 0:3 = 0:7 0:4 + 2 0:3 = 1 (the earth)

: : : 0:4 + 2n;2 0:3 : : : :

This, for n = 1 2 3 4 6 7, determines quite precisely the orbits of the known planets, and still worked rather well, with n = 8, for the orbit of Uranus. However, the number n = 5 was missing, which suggested that the `Creator of this world' has certainly not left empty this gap (see citation). To prove or disprove this conjecture became then a major scienti c challenge of the time.

1.2 The Thesis of Hegel

Sehen Sie sich doch nur bei den heutigen Philosophen um, bei Schelling, Hegel, Nees von Esenbeck und Consorten, stehen Ihnen nicht die Haare bei ihren De nitionen zu Berge? (Brief von Gauss an Schumacher, 1. 11. 1844, Werke 12, p. 62)

Also one of the most in uential philosophers, Georg Wilhelm Friedrich Hegel, took part in these discussions and submitted in 1801 his thesis (Dissertatio philosophica de orbitis planetarum, Ienae MDCCCI, Werke 5, p. 221-253) at the University of Jena. He starts by `proving' the laws of Kepler without any need of mathematics or physics, and, in the last part, turns his attention to Bode's rule. This latter had of course no philosophical contents. No, we have to look up Plato's Timei and nd the magic numbers 1 2 3 4 9 8 27, where we are allowed to replace the 8 by a 16 (!)1. Then we take the third roots of the fourth powers p of these numbers, still replacing without hesitation (`ponamus') the number 1 by 3 3, and we obtain the sequence2 1:4 2:56 4:37 6:34 18:75 40:34 81 in which in fact `between the fourth and fth position is a lot of space' 3 hence there is no planet missing in this gap. It was of course bad luck, that this conclusion was ridiculised precisely the same year by the discovery of Ceres. Needless to say, that all this was not favourable to the mutual esteem between scientists and philosophers (see quotation). The rst of January 1801, the Italian astronomer Giuseppe Piazzi discovered in the Taurus constellation a tiny little spot, and was able to follow its orbit until the 11th of February, when it was disappearing behind the Sun. He named it Ceres Ferdinandea, so that later astronomers just had to drop the name of Piazzi's king and the new celestial body had a generally acceptable name. The data of Piazzi's observations of Ceres were published in the
`16 enim pro 8 quem legimus ponere liceat'. most of these roots are wrong in the last digit. 3 `inter quartum et quintum locum magnum esse spatium'
1 2

1.3 The Discovery of Piazzi

September issue of the Monatliche Correspondenz (see 19]). We present these values in Table 1.1, where the latitudes i are taken southward. The great challenge was now to rediscover 1801 Jan. 1 2 3 4 10 13 14 19 21 22 Table 1.1: The observations of Piazzi Longitude Latitude Longitude 530 23 06.38 30 06 45.16 23 530 44 12.46 530 19 38.18 30 02 26.46 28 540 15 18.52 0 0 53 16 37.70 2 58 08.04 30 540 30 10.52 530 14 21.44 20 53 51.98 31 540 38 05.58 530 07 57.64 20 28 53.64 Feb. 1 540 46 27.14 530 10 05.60 20 16 46.08 2 540 55 01.52 530 11 54.20 20 12 54.02 5 550 22 45.20 530 26 01.98 10 53 37.82 8 550 53 04.52 0 0 53 34 22.68 1 46 13.06 11 560 26 28.20 530 39 11.58 10 42 28.80
0 0 0 0 0 0 0 0 0 0 00 00 00 00 00 00 00 00 00 00 0 0 0 0 0 0 0 0 0 0 00 00 00 00 00 00 00 00 00 00 0 0 0 0 0 0 0 0 0

00 00 00 00 00 00 00 00 00

Latitude 10 38 46.78 10 21 04.92 10 14 14.24 10 10 51.02 10 07 34.18 10 04 18.10 00 54 34.54 00 45 08.28 00 35 55.02
0 0 0 0 0 0 0 0 0

00 00 00 00 00 00 00 00 00

this lost body towards the end of the year, and many astronomers tried to extrapolate as good as possible its orbit (Burckhardt, Olbers, Piazzi). But a certain \Dr. Gauss in Braunschweig" computed a totally di erent solution \nach einem eigenthumlichen Verfahren" and published it the 29th of September 1801. Not satis ed with that, with an enormous computational e ort, he recomputed and readjusted the parameters continuously, and nally arrived in December 1801 at the values presented in Table 1.2 4 . Table 1.2: The elements of Ceres (Gauss Dec. 1801) Sonnenferne ...................................................... 3260 53 50 ....................................................................... 810 1 44 Neigung der Bahn ............................................. 100 36 21 Logarithmus der halben grossen Axe ................ 0:4414902 Excentricitat ..................................................... 0:0819603 Epoche: 31 Dec. 1800 mittl. helioc. Lange ...... 770 54 29
0 0 00 0 0 00 00

00

W.W.R. Ball, History of Mathematics, London 1901, p. 458 quoted from R.E. Moritz, Memorabilia Mathematica, p. 157.
5

The 7th of December 1801, Freiherr von Zach re-discovered Ceres precisely at the position predicted by Gauss. With this event, Gauss' \analysis proved him to be the rst of theoretical astronomers no less than the greatest of arithmeticians" 5. 4 the argument of the perihelion is given by w = Sonnenferne ; 1800 ; .

2 The First Computations of Gauss

Die von Kreis- und Parabel-Hypothesen unabhangige Bestimmung der Bahn eines Himmelskorpers aus einer kurzen Reihe von Beobachtungen beruht auf zwei Forderungen: I. Muss man Mittel haben, die Bahn zu nden, die drei gegebenen vollstandigen Beobachtungen Genuge thut. II. Muss man die so gefundene Bahn so verbessern konnen, dass die Di erenzen der Rechnung von dem ganzen Vorrath der Beobachtungen so gering als moglich werden. (Gauss, Summarische Ubersicht see 8], p. 148) Gewiss, jeder der die Rechnungen kennt, die die Bestimmung der Elemente eines Planeten und dann jeder daraus herzuleitende Ort erfordert, muss es bewundern, wie ein einzelner Mann in so kurzen Zeitraumen so vielfache muhsame Rechnungen zu vollenden vermogend war. (von Zach, Marz 1805, see Gauss Werke 6, p. 262)

Le ciel est simple is the leitmotiv of the amateur observatory in St. Luc, Switzerland, at least as long one does not try to understand the computations of Gauss. The great advantage of Gauss' ideas over his rivals was, that he assumed nothing else than Kepler's laws for his planet and no other hypotheses. But Gauss never reveiled details of his calculations. Urged by Olbers, he nally sent in August 1802 a manuscript Summarische Ubersicht without any desire to see it published. This text was nally printed in 1809 by von Lindenau (see 8]) with all the excuses of the editor for the many `imperfections'. An excellent English description of Gauss' calculations has appeared recently (see 18]). All the di culty stems from the great number of variables involved. Indeed, we have to work with Elements of orbit Geocentric Heliocentric spherical w arg. of perihelion coordinates long. of ascend. node (A) (B) coordinates (2.1) 0 1 i inclination of orbit 0 1 ( ) ( ) xC a semi-major axis ByA B C @ @ A e eccentricity z l0 mean heliocent. long. The quantities measured are the angles and (the distance is unknown, of course) for several time values, the quantities to be computed are the elements of the orbit. So we need formulas for the connecting passages (A) and (B). Passage (A). For a given time t, we have rst to nd the position of the planet on the ellipse, i.e., to nd the eccentric anomaly u (see Fig. 2.1, left). We rst assume the point t = 0 at the perihelion. Then Kepler's second law (`same times, same areas') tells us that t is proportional to the area A. The period P of the orbit thereby corresponds to the total area ab of the ellipse, so we have A = t:

We now stretch the ellipse to a circle (Fig. 2.1, right), so that B = a A, but also B = a22 (u ; b e sin u) (di erence of the areas of a sector and the triangle T ). The three equations lead to nt = u ; e sin u (Kepler's equation) (2.2) 4

ab

a b u Pe r A v f a
f e

Pc B v

Figure 2.1: Kepler orbit

a

focus (the sun), e the planet, anomaly, semi-major axis (perihelion), eccentricity.
P

the true anomaly,

a sin u T u ae 0

a
v

the eccentric

where the constant n = 2P is called the mean angular rate. If we nally transfer the origin of time to the correct place, we have to add the epoch l0 and to subtract the argument of the perihelion and the longitude of the ascending node. Thus

l0 (w + ) + nt = u e sin u
; ;

(2.3)

is, nally, the transcendental equation to be solved for u. We still have Kepler's third law to exploit, which states that a3 is proportional to P 2. This means that

n2a3 is a known constant:

(2.4)

We nally incline this orbit by an angle i with respect to the earth's orbital plane, and arrive with the help of elementary spherical geometry at formulas for the heliocentric ecliptic coordinates

x(w

i a e l0 t) y(w

i a e l0 t) z(w

i a e l0 t):

(2.5)

Passage (B). For this, we have to know the solar geocentric coordinates (X Y Z ) (for the
= x+X = = y+Y = = z+Z = cos cos cos sin sin

same date and time) and we obtain the geocentric ecliptic coordinates of the planet by adding these and taking spherical coordinates (2.6)

(more details about the above computations can be found in celestial mechanics texts such as 4, pp. 182-186] or 3, pp. 84-90]). Gauss' Procedure. At the time of the discovery of Ceres, it was well known how to compute the six elements of the orbit of a planet from two sets of heliocentric coordinates x y z. This consists in solving 2 3 nonlinear equations in six unknowns. The great di culty was, that there were only two geocentric observed values i i per data point. After long formula manipulations of the above expressions, Gauss was able to reduce the computation of one set of heliocentric coordinates x y z to the knowledge of two sets of observed values i i via the solution of a complicated transcendental equation. Therefore, from three sets of 5

observations, he obtained two sets of heliocentric coordinates and thus the desired elements. These developments, too long to be given here, are excellently presented in 18] see also 11]. Thereby, it was advantageous to have the third point exactly in the middle of the two others. So Gauss started with the data Jan. 2, Jan. 22, and Feb. 11. The obtained values of the elements were then recomputed repeatedly by changing the dates, and by checking them for the remaining data. All these results of calculations, and Gauss' later results for the subsequent discoveries of Pallas Olbersiana, Juno and Vesta, are impressively documented in Gauss' Werke, vol. 6, p. 199{402. As a conclusion, we see that these computations were not performed with the method of least squares.

3 The Method of Least Squares

Der Verfasser gegenwartiger Abhandlung, welcher im Jahre 1797 diese Aufgabe nach den Grundsatzen der Wahrscheinlichkeitsrechnung zuerst untersuchte, fand bald, dass die Ausmittelung der wahrscheinlichsten Werthe der unbekannten Grosse unmoglich sei, wenn nicht die Function, die die Wahrscheinlichkeit der Fehler darstellt, bekannt ist. (Gauss, Gott. gelehrte Anz. 33 (1821), p. 321-327.)

Things changed, however, after Ceres had been rediscovered in December 1801 and when more observations became available. Now the task was to improve the orbital elements to still higher accuracy with the help of all these new data. Here Gauss started to apply the method of least squares, again without ever reveiling details to anybody. The only evidence is the last sentence of the Summarische Ubersicht: \hat man schon Beobachtungen von 1 oder mehrern Jahren : : :, so halte ich der Gebrauch der Di erential-Anderung, wobei man eine beliebige Zahl von Beobachtungen zum Grunde legen kann, fur das beste Mittel" | and the precision of the results (see also 17]). Legendre contra Gauss. In 1805 appeared the work Nouvelles methodes pour la determination des orbites des cometes by A.-M. Legendre, containing in an appendix an extremely beautiful presentation of the methode des moindres quarres. The clarity of this work together with numerical examples made the least squares method immediately known in all scienti c communities. Despite of Legendre's work, Gauss called, in his famous treatise 9] Theoria motus corporum celestum, published 1809, obstinately the least squares idea \my principle, which I have made use of since 1795". Legendre then protested in a long letter to Gauss, which is worth reading (Gauss Werke X/1, p. 380 complete text in English translation in 15], p. 242{243 \je n'ai jamais appelle principium nostrum un principe qu'un autre avait publie avant moi"). Gauss never answered to Legendre, but mentioned to others the existence of a cryptic entry in his diary from June 17, 1798, simply saying \# Calculus probabilitatis contra La Place defensus." (see facsimile in Fig. 3.1). Legendre never forgot this, also because the young Jacobi (concerning elliptic functions, see 15], p. 246) and the young Bolyai (non-Euclidean geometry, see 2], p. 99) made similar experiences with Gauss. Gauss' Probabilistic Justi cation of the Least Squares Principle. Going much further than Legendre, Gauss gave an answer to the question: \Why least squares and not, for 6

Figure 3.1: Entry in Gauss' diary, June 17, 1798 example, least fourth powers or least sixth powers?". To explain the idea, we treat a simple problem, i.e., the approximation of three `observations' xi yi (i = 1 2 3) by an `orbit' which is a straight line y = a + bx (3.1) (see Fig. 3.2). If now the three points don't lie on one line, there are three di erent lines,
2

y2 y1 x1

y3 x3

y2
2 1

y3

Figure 3.2: Motivation for the least squares method.

x2

y1

x1

x2

x3

none of which is satisfactory (see the left picture). Now suppose that there exist values i on one line (which we do not know, however), and that the measurements yi contain errors with a certain probability. The most common probability law is (see the right picture the probability is symbolized by varying grey tones)6

above probabilities, i.e.,

e; i h2 i : P (0 i yi ) = h Now, the probability for having measures the three values y1 y2 y3 is the product of the three
(
;

y )2

h
3 X

3 3

i=1

( e;

i ;yi ) h2

e;

P3

i=1 ( i ;yi ) h2

We have then maximum likelihood of our result, when this probability is maximal, i.e., when the exponent
i=1
6

yi)2 = min!

and with (3.1)

3 X

i=1

(a + bxi ; yi)2 = min!

Laplace arrived at this law from the binomial coe cients and a passage n ! 1 for Gauss it was simply the law which reproduced the well-proved arithmetic mean.

which is, precisely, the principium nostrum. Di erentiating the last expression with respect to a and b we obtain 0 1 1 x1 1 xi a = yi or AT A = AT y A = B 1 x2 C : (3.2) @ A 2 xi xi b xiyi 1 x3 These are called the normal equations. Good luck, that the principium with best probabilistic justi cation also leads to the easiest possible problem to solve, a linear system of equations. Further Developments. Of the many important consequences which followed the least squares idea, we mention the following. Gaussian elimination. In order to prove the solvability of the normal equations, Gauss made in 9] the rst clear description of the elimination algorithm for linear equations. Gauss-Newton method. In the same paper, Gauss also explained how nonlinear least squares problems are linearized in the neighbourhood of a rst approximate solution, which is then iteratively re ned. Laplace's central limit theorem. In 1809, Laplace published his central limit theorem, showing that any probalility function, after taking arithmetic means, tends to the normal distribution for n ! 1. Soon after, he extended this to justify the principle of least squares for arbitrary probability functions and n ! 1. A great publication of all these results was Theorie analytique des probabilites from 1812. In 1823, Gauss publishes a second fundamental treatise on least squares, 10] Theoria combinationis observationum erronibus minimis obnoxiae in two parts, which contains a new justi cation of the least squares principle, independent of the probability function, which is today called the Gauss-Markov Theorem. In 1828, Gauss publishes a Supplementum, which contains impressive calculations for the geodesic triangulations of the Netherlands and the country of Hannover. Also in 1828, Bessel discovers, originally for the discrete case, the relation between the least squares idea, the orthogonality relations, and the Euler-Fourier formulas for the trigonometric approximation this discovery, extended by Gram (1883) to the continuous case, is the basis of the L2 Hilbert space theory of Fourier series. In 1845, Jacobi publishes his method for solving the normal equations with the help of successive rotations in R2 . These rotations lead in the 1950ies to Givens' method for triangularization and the rst stable eigenvalue algorithm. In 1900 appears the classical paper of Karl Pearson 14], which combines the least squares method with the 2 distribution and led to the famous 2 -test for the reliability of hypotheses. 8

In 1958 appears Householder's re ection algorithm, which, by replacing Givens' rotations, leads to the QR decomposition, and, by Golub (1965), became the nowadays standard algorithm for least squares problems. All examples which follow, have been computed with this method, using a code written by E. Hairer for his lecture `Analyse Numerique' (http://www.unige.ch/math/folks/hairer/polycop.html). A complete modern treatment of numerics for the least squares method, which contains nearly 1000 bibliographical references, is the book of Bjork 1]. Of valuable help for readers interested in Gauss' contributions is the bilingual edition of Theoria combinationis observationum with Supplementum and Anzeigen, and, most important, a carefully written Afterword, due to G.W. Stewart 16]. Many original texts translated into English can also be found in 12], Sect. 4.9 and 4.10. The orbit of Ceres with the least square method. In possession of a modern algorithm, we now want to compute the elements of the orbit of Ceres using the data of Piazzi with the least squares method. This we did as follows: for given orbital elements w i a e l0, we designed a subroutine, computing with the aid of formulas (2.1) through (2.6), for the times ti of Piazzi's observations, the geocentric longitudes i and latitudes i . The necessary expressions for the solar heliocentric coordinates of the earth (X Y Z ) were obtained from the server of l'Institut de mecanique celeste et de calcul des ephemerides (http://www.bdl.fr/ephemeride.html). These values compared to the actual observations b i bi de ne a function

F : IR6 (w i a e l0)
and we have to nd w
k

; ! 7!

IR38
t1

(: : :) ; b 1 : : : b t1 (: : :) ; 1 : : :

(: : :) ; b 19 b t19 (: : :) ; 19
t19

(3.3)

i a e l0 such that 19 X F (x) 2 = ( ti b i)2 + ( 2

k ;

i=1

ti

bi)2

; !

min!

(3.4)

As initial values we choose values close to Gauss' values and after 5 Gauss-Newton iterations the least squares solution was precise to 7 digits (values displayed in Table 3.1). Table 3.1: The elements of Ceres (least square) Sonnenferne ...................................................... 3180 12 27 ....................................................................... 800 55 9 Neigung der Bahn ............................................. 100 35 38 Logarithmus der halben grossen Axe ................ 0:4448506 Excentricitat ..................................................... 0:0694885 Epoche: 31 Dec. 1800 mittl. helioc. Lange ...... 750 47 31
0 0 00 0 0 00 00

00

t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15 t16 t17 t18 t19

= 53.38 = 53.33 = 53.28 = 53.24 = 53.13 = 53.17 = 53.20 = 53.44 = 53.57 = 53.65 = 53.74 = 54.26 = 54.50 = 54.64 = 54.77 = 54.92 = 55.38 = 55.89 = 56.44

= = = = = = = = = = = = = = = = = = =

3.11 3.04 2.97 2.90 2.48 2.28 2.21 1.89 1.77 1.71 1.65 1.35 1.24 1.18 1.13 1.07 0.91 0.75 0.60

4.0

Latitude

3.0

2.0

Least square Piazzi Longitude

1.0

53.0

54.0

55.0

56.0

57.0

Figure 3.3: The computed and observed positions of Ceres In Fig. 3.3 we compare Piazzi's observations and our computations. These latter computations t better Piazzi's observations than Gauss' values, but Gauss' elements are closer to the true orbital elements. The reason is, that some of Piazzi's measurements contain slight errors (this was already observed by Gauss), and that these errors in uence enormously the solutions (this was also observed by Gauss). This phenomenon is today called a badly conditioned problem.

4 Three Today's Examples

The method of least squares is the automobile of modern statistical analysis . . . (The rst sentence of Stigler 17].)

Nearly everywhere, where data have to be analysed or models adjusted, one applies today the method of least squares, very often to problems of enourmous dimensions. For particularly impressive examples, we refer the reader to a forthcoming book by Deu hard 5]. Here, in this paper, we explain in some detail three nice examples from everydays life.
Problem. We have a photograph (see Fig. 4.1), on which we distinguish a couple of points b b with measured local coordinates (uk vk ). Of the same points, we determine the corresponding space coordinates (xk yk zk ) from a map, where the origin for x y is placed arbitrarily and z are the altitudes. The question is to nd out the position of the camera, its focus and its angles of inclination. A copy of the map, the Swiss national map 1:50000 folio 5003, can be found under http://www.unige.ch/math/folks/hairer/polycop.html. In Table 4.1 are given the values used in our calculations. eee For the solution of our problem, we denote by (x y z ) the position in space of the camera's objective, and by ~ = (a b c) the vector between the objective and the projection plane. a Finally we allow the camera to be rotated around ~ by an angle . There are thus seven a

4.1 The Position of a Camera

10

Figure 4.1: A photograph (from the Montblanc region Photo: G. Wanner) Table 4.1: The data for the camera problem (in meters). 1. 2. 3. 4. 5. 6. Col des Grandes Jorasses Aiguille du Geant Aig. Blanche de Peuterey Aiguille du Tacul Petit Rognon Aiguille du Moine

0:0480 0:0100 0:0490 ;0:0190 0:0600 0:0125

; ;

b uk

0:0290 0:0305 0:0285 0:0115 ;0:0005 ;0:0270

b vk

9855 5680 8170 5020 2885 730 8900 7530 5700 7025 8980 11120

xk

yk

3825 4013 4107 3444 3008 3412

zk

unknowns to determine. Very similar to the calculations of Gauss, but much easier, we have, once these 7 variables xed, to nd out the relations between the coordinates x y z in space and the corresponding projection points u v on the photograph. For this, we x two orthogonal vectors in the projection plane

b ~ = p 1 B ;a C h @ A a 2 + b2 0

;ac ~ = q 2 2 1 2 2 2 B ;bc C : g @ A (a + b )(a + b + c ) a2 + b2

(4.1)

Then, for a given point (x y z) (see Fig. 4.2) we compute a vector w by ~

w= ~

0 e1 Bx xC e @y yA e z z
; ; ;

(4.2)

11

(x y z) (

eee (x y z)

w ~ ~ a

~ g ~ h
:::

(u v)

Figure 4.2: Perspective representation: today (left)

and 500 years ago (Durer 1525, right)

where the factor is determined by hw ; ~ ~ i = 0. Then = hw ~ i and = hw ~ i are the ~ aa ~h ~g coordinates of the projection point, which are nally rotated by : u = cos ; sin : (4.3) v sin cos We have then the best solution, when these projected points (uk vk ), for the data (xk yk zk ), b b correspond in the best possible way to the measured data points (uk vk ) of the photograph. Thus, according to `principium nostrum', if
6 X

k=1

b b (uk ; uk )2 + (vk ; vk )2 = min!

(4.4)

For this problem, the Gauss-Newton algorithm leads, with very rough initial values, after a couple of iterations, to the solution

e x = 9679

e y = 13139

e z = 4131:

The photograph has thus been taken from the summit of the Aiguille Verte, whose altitude is known to be 4122 meters. The precision of these amateur calculations is not that of professional Swiss topographers. We now apply the same algorithm as above to an example from the history of art. The `mountains' are now the exact vertices of a regular icosidodecahedron somewhere placed in space, and the `photograph' is a drawing by Leonardo da Vinci (see Fig. 4.3, left) which was performed for the book De divina proportione by Luca Pacioli, Venice 1509. Pacioli says in his preface that they were done `by the divine left hand of my friend Lionardo of Florence' (quoted from 6], p. 253). After having placed the `camera' in the best possible position, we can project the exact vertices back to the picture and thus nd out, nearly 500 years later with the aid of modern computing tools, the actual precision of Leonardo's drawing. 12

4.2 Leonardo's Polyhedron

16 9 17 1

15

14

2 18 10 3 19 6 4

5 13

7 12

20

11

Figure 4.3: Left: Drawing of Leonardo da Vinci (1510, Codex Atlanticus fol. 707r Bibliotheca
Ambrosiana, Milano), from 13], p. 100 right: Leonardo's vertices and, in grey, the `corrected' drawing (Assyrus Abdullus & Gerhardus Wannerus, lingu programmatori Fortranus & Postscriptus, Calculatores SunBlade 100, Universitas Genav )

Table 4.2: Measured vertices in Leonardo's drawing (in mm)

1 2 3 4 5 6 7

5.409 -26.388 -13.259 26.517 37.265 2.734 55.650

uk

30.691 6.720 -30.369 -28.782 8.054 -52.888 -18.639

vk

8 9 10 11 12 13 14

36.865 -25.283 -45.244 18.814 48.271 56.767 46.037

uk

34.219 36.394 -16.728 -55.828 -34.749 0.764 33.043

vk

15 16 17 18 19 20

17.609 -17.522 -45.244 -56.768 -45.433 -18.198

uk

52.536 52.122 31.161 -2.147 -35.867 -56.563

vk

For this, the coordinates of the 20 visible vertices haven been measured on the picture (see table 4.2). These points are re-drawn in the right picture of Fig. 4.3 in black, together with the `corrected' polyhedron (in grey). We see that the drawing is very precise in the centre, but some vertices towards the periphery and in the background are less `divine'. Inspired by Leonardo's polyhedral skeleton, the authors have not resisted the pleasure to produce a stereographic view of this beautiful object in Fig. 4.4. In summer 1999 a hanging glacier high up in the mountains above Grindelwald (Switzerland) started to advance and threatened the region below by an enormous ice fall. In order to avoid a serious accident, a precise breaking o forecast was of great importance. Scientists from the ETH Zurich (M. Funk) therefore implanted a surveying stake on the ice masses (see 13

4.3 The Hanging Glacier above Grindelwald

Figure 4.4: Stereographic view of Leonardo's polyhedron (hold the picture rather close to the eyes
and blend the two drawings by staring through the paper)

Fig. 4.5, left) and observed carefully the advancing positions of the stake. The measured data are reproduced in Fig. 4.5 to the right. The time t = 0 corresponds to the 18th of July, 1999, at 7 a.m.
14 12 10 8 6 4 2 0 0 5 10 15 20

meters
t 1 = 0.0 t 2 = 2.0 t 3 = 7.0 t 4 =12.0 t 5 =13.0 t 6 =15.0 t 7 =16.0 t 8 =17.0 t 9 =19.0 t10 =20.0 t11 =22.0 s 1 =0.000 s 2 =0.529 s 3 =1.936 s 4 =3.598 s 5 =3.977 s 6 =4.802 s 7 =5.219 s 8 =5.667 s 9 =6.669 s10 =7.208 s11 =8.464

t =27.25 s0 = 0.01 v0 =0.09 a0 = 1.79

25

days

Figure 4.5: Left: the Grindelwald hanging glacier with the surveying stake (Photo: M. Funk, from
7]). Right: the data and the breaking o forecast.

Earlier experiences with ice falls (in particular one at the Weisshorn) have shown, that the increasing speed of such ice masses satis es a formula
0 v(t) = v0 + (t at )n 1
;

where n

1 2

. By integrating this, we obtain for the positions

; ; ;

1;n 1;n (4.5) s(t) = s0 + v0 t + a0 (t1 t) 1 t1 : n The problem is now, to determine the unknown constants s0, v0, a0 and t1 in such a way,

that this function approaches the measured data points with minimal least squares error. The solutions obtained in this way are given in Fig. 4.5, and predicted the ice fall for t1 = 27:25, which corresponds to the 14th of August at 1 p.m. Actually, the glacier fell the 14th of August 14

at 2 a.m. The forecast, 5 days before the event, was thus wrong by less than half a day. For more details, see 7]. Conclusion. After having seen, how the observations of a couple of stars have helped to develop modern science in such an extraordinary way, we must say, really, that stars in uence our lifes, just not in such a way, as readers of horoscopes are believing. Acknowledgements. The authors are greatful to S. Cirilli, P. Wittwer and E. Hairer for many stimulating discussions and suggestions in particular, Stephane has rendered an invaluable help by scanning all the photographs and preparing their ps- les.

References
1] A. Bjork, Numerical methods for least squares problems, 408pp., SIAM 1996. 2] W.K. Buhler, GAUSS, eine biographische Studie, Springer Verlag, 1987. 3] G.W. Collins II, The Foundation of celestial mechanics, Astronomy and Astrophysics Series, vol.16, Pachart Publishing House, Tuscon, 1989. 4] A. Danjon, Astronomie generale, Albert Blanchard, Paris, Seconde ed., 1986. 5] P. Deu hard, Newton methods for nonlinear problems. A ne invariance and adaptive algorithms, to appear. 6] J.V. Field, Rediscovering the Archimedean polyhedra: Piero della Francesca, Luca Pacioli, Leonardo da Vinci, Albrecht Durer, Daniele Barbaro, and Johannes Kepler, Arch. History Exact Sc., 50 (1996), 241{289. 7] M. Funk & H.-E. Minor, Eislawinen in der Alpen: Erfahrungen mit Schutzmassnahmen und Fruherkennungsmethoden, Wasserwirtschaft vol. 91 (2001), 362-368. 8] C.F. Gauss, Summarische Ubersicht der zur Bestimmung der Bahnen der beiden neuen Hauptplaneten angewandten Methoden, Monatliche Correspondenz, herausgeg. Freiherr von Zach, Sept. 1809, Werke, vol. 6, pp. 148-165. 9] C.F. Gauss, Theoria motus corporum coelestium, Perthes et Besser, Hamburgi (1809), Werke vol. 7, 1{288. 10] C.F. Gauss, Theoria combinationis observationum erronibus minimis obnoxiae, Pars Prior et Pars Post., Comm. Soc. Reg. Scient. Gott. 5, (1823), Werke vol. 4, 1{26 27{53 Supplementum Comm. Soc. Reg. Scient. Gott. 6, (1828), Werke vol. 4, 55-93. 11] C.F. Gauss, Elliptische Bahnbestimmung, aus Gauss' Nachlass, Werke, vol. 11, pp.221252. 12] H.H. Goldstine, A history of numerical analysis from the 16th through the 19th Century, Springer Verlag 1977. 15

13] O. Letze & T. Buchsteiner, Leonardo da Vinci, scientist inventor artist, Exhibition Catalog 1999, Verlag Gerd Hatje, Ost ldern-Ruit, Germany. 14] K. Pearson, On a criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonably supposed to have arisen from random sampling, Phil. Mag. (5) 50 (1900), 157-175 corr. Phil. Mag. (6) 1 (1901), 670-671. 15] R.L. Plackett, Studies in the history of probability and statistics. XXIX, The discovery of the method of least squares, Biometrika 59 (1972), 239{251. 16] G.W. Stewart, Gauss, Theory of the combination of observations lesat subject to errors, bilingual edition of 10] with an Afterword, Classics in Appl. Math., SIAM 1995. 17] S.M. Stigler, Gauss and the invention of least squares, The Annals of Stat., 9 (1981), 465-474. 18] D. Teets & K. Whitehead, The discovery of Ceres: How Gauss became famous, Math. Magazine 72, (1999), 83{93. 19] Freiherr von Zach, Fortgesetzte Nachrichten uber den langst vermutheten neuen HauptPlaneten unseres Sonnen-Systems, in Gauss Werke, vol.6, pp. 199-204.

16