Five directions and no distance among them

An image of an asteroid records a direction and throws the range away. It fixes which way the object lay at a stated instant, and says nothing whatever about how far along that line it was.

So the orbit has to be assembled out of directions, and since an orbit is six numbers while one look supplies fewer, how many looks is an arithmetic question with a definite answer. It depends on how many unknowns are carried, and the two counts that matter here are different.

The drawn geometry is deliberately unkind: its arc is centred on the stationary point in geocentric longitude, where the apparent motion reverses and the sight lines fan out least, which is where Piazzi’s arc of 1801 happened to lie. Even so the five lines span 1.37° of sky, and an orbit comes out of them.

Four numbers in a plane, six in space, and the count is the argument #

A planar orbit is four numbers: the semi-major axis, the eccentricity, the direction of the apse line, and an epoch saying where on the curve the body was at a named moment. A sighting in the plane is one number, an angle. Four unknowns, one equation per look. Three sightings therefore give three equations for four unknowns and leave a one-parameter family, which is not a determination but a curve of them. The generator behind these figures refuses that case outright, reporting that three sightings give three angles for the four numbers of a planar orbit, which leaves a 1-parameter family rather than one. Asked for four, where the count is even, it reports that four leaves two orbits rather than one — an even count of nonlinear equations does not guarantee a single root. Five over-determine the problem by one, and five is what the hero figure draws.

The planar reduction has already spent one angle per image: setting the inclination and the node to zero removes two unknowns and discards the observed ecliptic latitudes at the same stroke, because it has asserted they are zero. Six unknowns become four, two angles per look become one, and a count that closed at three now closes at four. So no figure here illustrates the three-observation method, and reading the drawn five as a lax version of the classical three gets the arithmetic backwards.

Two unknown ranges, and three angles to answer for them #

The determination runs best in variables that are not elements at all: the ranges $\rho_1$ and $\rho_5$ at the first and last sighting. Each places a point,

$$\mathbf{P}_k = \mathbf{E}_k + \rho_k\,\hat{\boldsymbol{\lambda}}_k,$$

with $\mathbf{E}_k$ the Earth’s known position and $\hat{\boldsymbol{\lambda}}_k$ the observed direction. Two positions and the time between them are Lambert’s problem, which returns the velocity at the first — and a position with a velocity is an orbit. So two numbers generate all four elements, the end sight lines and the two dates having supplied the missing pair. The orbit is then checked against the sightings it was not built from. Propagating to each interior date gives a predicted direction, and the three residuals are

$$F_k = \arg\!\left(\mathbf{P}(t_k) - \mathbf{E}_k\right) - \lambda_k, \qquad k = 2,3,4 .$$

Three residuals for two unknowns, which is the same over-determination as five angles for four elements arrived at from the other side. A Gauss–Newton step with a halving line search drives them to $10^{-16}$ radians, and the two ranges are the answer. Every residual evaluation contains a propagation and every propagation contains Kepler’s equation solved by iteration, so the determination is a transcendental solve nested inside a nonlinear one.

The law that closes it with no propagation in the argument #

A second route to the same orbit shares not one step with the first, and it is the more instructive. Take three sightings — first, middle, last — and slide all three crossings out along their sight lines by a common factor. Every member of that family satisfies every observation, since a direction with no range permits it. But three points determine a conic with a focus at the Sun, and do it by one linear solve: writing the conic as

$$\frac{1}{r} = A + B\cos\theta + C\sin\theta,$$

three points give three linear equations for $A$, $B$ and $C$, and then $p = 1/A$ and $e = \sqrt{B^2 + C^2}/A$. Each candidate distance therefore predicts two swept areas, $S = \int \tfrac12 r^2\,\mathrm{d}\theta$ between consecutive crossings, and the area law says what they must be: in the ratio of the elapsed times. One scalar equation, one unknown, and the root is the distance. That route contains no propagation whatever — no Lambert solve, no Kepler step, no iteration on elements, one linear conic fit and two quadratures — and it returns $a = 2.7658$ AU again. Two determinations of one orbit sharing no machinery is the strongest internal check available, and it is why the area law is not a consequence to be verified afterwards but one of the equations.

The two rejected candidates make the point from outside: a 15% error in the common distance turns $a$ into two-thirds of itself and an 8% error doubles it, which is how badly conditioned this problem is.

Determined rather than fitted #

An exactly determined orbit and a least-squares orbit are different epistemic objects, and the difference shows when the data are poor rather than when they are good. A determination consumes as many observations as it has unknowns, plus a few to be safe, and inverts. There is no averaging, no down-weighting of a bad night, no error model: what comes out is the orbit those particular angles imply, and an error in one of them reaches the elements undiluted.

So the failure mode peculiar to a determination is not noise but multiplicity. Which root a single starting guess converges to is an accident of the guess, and one root of this system puts the object inside Earth’s orbit on a high-eccentricity path that fits four of the five angles. The machinery here sweeps sixteen starting pairs of distances, from 0.8 to 4.4 AU, and refuses to draw anything if a second solution fits as well as the best. A rival root is not a choice to be made but a statement that the sightings do not single out one orbit.

Which is why initial-orbit determination is a first stage and never the last: it supplies the starting point that the differential correction — where the perturbations enter — refines into a covariance.

What was actually measured, in Palermo, in 1801 #

Piazzi’s instrument was a Ramsden vertical circle, and it measured two things: the moment a body crossed the meridian, against a clock, and its altitude then, off a graduated circle. Those give right ascension and declination — two angles and one epoch per observation, and no distance of any kind. Twenty-four of them between 1 January and 11 February 1801 covered about three degrees of sky, and then the object was lost in the Sun’s glare.

Getting from those readings to a heliocentric direction took a chain of assumptions, each a place the answer can be wrong: the clock rated against sidereal time, refraction removed, aberration and precession and nutation reduced to a common equinox. And the Earth’s own position at each date had to come from solar theory, because the baseline here is the Earth’s orbit — an error in the assumed position of the observer enters the answer exactly as an error in the observed angle does.

One consequence separates a good number from a known one. The semi-major axis comes out in astronomical units, because the ruler was Earth’s own orbit; converting it to kilometres needed the solar parallax, which in 1801 rested on the Venus transits of 1761 and 1769 and was good to about a per cent. The orbit was known far better in astronomical units than the astronomical unit was known in kilometres — an ordering that held until radar supplied a range directly, the one observable this whole construction lacks.

The historical count is a count too. The methods available before 1801 were built for comets and assumed a parabola, which pins the eccentricity at 1 and leaves five unknowns for three observations to over-determine. Gauss carried all six free, so the conic class came out of the data instead of being asserted before the arithmetic began — and which class it is is often the most interesting thing a determination reports.

The baseline is the Earth’s, and its direction matters more than its length #

The hero figure gives two numbers for one baseline — 0.691 AU of travel, 0.403 AU of it across the sight lines — and the second is the one that does the work. Baseline length is not the criterion, because the component of Earth’s motion along a sight line is invisible: it changes the range and not the angle. A test on distance travelled would pass a two-day arc observed straight down Earth’s track, where there is no triangle at all, and fail a forty-day arc observed square to it, which is a fine geometry. So the criterion is a fraction — at least 2% of the travel across the lines — and the drawn arc clears it with 58%.

The stellar parallax measurement then fails in the same reciprocal way. There the fractional error in a distance is the angular error divided by the parallax, so a star ten times beyond the limit has an error bar including infinity. Here the reciprocal appears as a sensitivity: shorten the arc, shrink the angles, and the recovered semi-major axis becomes an arbitrarily violent function of the measurement. The only structural difference is whether the target is allowed to move while the baseline is traversed.

Why forty-one days was barely enough #

The obvious thing to say about a 41-day arc is that it was very little to work with. The interesting thing is how sharply little is defined.

What makes the panel an argument rather than a caution is the product. 0.66% of 2.7658 AU is 0.018 AU, which is the difference between a lost object and a found one — it decides whether a prediction lands in a telescope’s field or outside it. Forty-one days was not too few. It was barely enough, and the exponent is what “barely” means: at a quarter of that arc there is nothing left to be precise about.

The scale of the arithmetic has changed and the shape of the problem has not, which is the next section’s subject.

What the picture cannot show #

The reduction is planar, and the real problem is not. Every figure here solves four unknowns from one angle per look. A real determination carries six and takes two angles per look, which is why the classical method needs three observations and not five. The drawn count is honest for the drawn problem, and it is not the count in Gauss’s memoir.

The sightings are exact. Each direction is generated from the true elements and handed to the solve without error, which makes the round trip to three parts in $10^{12}$ a test of the machinery rather than a claim about observing. Error enters only in the conditioning panel, as a linearised response to one arcsecond rather than as a data set with correlated systematics in it.

The Earth’s orbit is a circle here, and its ephemeris is assumed known. The baseline is the observer’s own motion, so everything wrong with the assumed motion is wrong with the answer, and no care with the target’s angles can detect it.

The propagation is two-body. Between sightings the object moves on a conic — excellent over 41 days, and over the ten months between losing an object and recovering it the perturbations are what decide whether the prediction is half a degree out or five.

None of those four is a limitation of the method as practised; each is a simplification made so that the count at the top of this essay could be drawn.

The same count, in a transit light curve #

The structure generalises anywhere observables are counted against unknowns, and the clearest parallel here is in another field. A transiting planet’s light curve offers four contact points — the instants at which the disc touches and clears the stellar limb — and four unknowns follow from them by geometry alone, with no fitting and no stellar model. Count what the data give, count what the model needs, and the difference is the number of assumptions imported from elsewhere.

Each assumption buys observations that were never made. Kepler’s reconstruction of the Martian orbit was that trade run honestly: he bought the distances he lacked by pairing observations one Martian year apart, which is an assumption about the period rather than a measurement of a range.

What the arithmetic cost, then and now #

The determination described here is a nonlinear solve with a transcendental equation nested inside it, and it is worth remembering that in 1801 every evaluation was done by hand.

Gauss’s own account is that the reduction took him several weeks. What he produced was not merely an answer but a method designed for the arithmetic available: the sequence of approximations is arranged so that each step is a small correction to the last, and the quantities that have to be computed to many figures are kept as few as possible. A method that converges in three iterations rather than ten is a different proposition when each iteration is an afternoon.

The same problem produced the least-squares method itself. Gauss used it on this data and published it in 1809, remarking that he had been using it since 1795; Legendre had published it in 1805 and objected. The priority dispute is well known and the more interesting point is what the problem demanded: an orbit fitted to more observations than it has parameters needs a rule for what “best” means, and the rule was invented because the observations existed.

The modern position is the same problem at an entirely different scale. A survey telescope detects tens of thousands of moving objects a night, most of them already known, and every one has to be given a preliminary orbit before it can be matched against the catalogue. The determination that took Gauss weeks is now run millions of times a night, and it is run with the same structure: a fast method producing a rough orbit, then a differential correction, then an identification against everything already known.

What has not changed is the conditioning. The arithmetic is free and the geometry is not, so the binding constraint on a modern survey is still the length of the arc it obtains before an object moves out of the observed fields — which is a scheduling problem rather than a computational one, and is why survey cadences are designed around re-visiting rather than around covering the most sky.

Two named methods do the determination and the difference between them is worth stating, because it is the same distinction the two routes above already draw. Gauss’s method works with the positions: it takes three sight lines, writes the middle position as a linear combination of the outer two — which is exact for coplanar motion — and closes the system with a ratio of sector to triangle areas that is refined by iteration. Laplace’s method works with the derivatives instead: it fits the observed direction and its first two time derivatives at a single epoch, and closes the system with the equation of motion evaluated there. Gauss’s is better conditioned on real data because differencing three well-separated observations is stabler than differentiating a short arc twice, which is why every modern pipeline is a descendant of it and Laplace’s survives mainly in textbooks.

What replaced both for the shortest arcs is worth naming too. An arc of a few hours constrains the direction and its rate but not the range or the range rate, so the solution is not a point but a two-dimensional set — the admissible region, the range and range-rate pairs consistent with a bound heliocentric orbit. Modern surveys carry that region forward as a probability distribution rather than choosing a root from it, and a second night’s observation cuts it rather than correcting a guess. It is the honest generalisation of everything above: when the count of observations does not close the system, the answer is a set, and pretending otherwise is what produces the spurious orbits of the previous section.

One consequence of all of this is worth stating in the terms a survey works in. A modern survey does not determine orbits for its detections one at a time; it links them. A night produces tracklets — pairs or triples of detections minutes apart, each giving a direction and a rate and nothing more — and the problem is to decide which tracklets from different nights belong to the same object before any orbit is attempted. That is a combinatorial problem whose size grows as the square of the catalogue, and the admissible region above is what makes it tractable: two tracklets can be joined only if their regions intersect, which rejects almost every pair without solving anything. Orbit determination in the sense of this essay then happens once, on the survivors, and the six numbers it returns are the end of a process whose hard part was deciding what data to give it. The linking step is also where the false orbits of the previous section are created and killed: a spurious link produces a determination that fits, and what refutes it is a third night rather than a better solver.

Where the ladder goes next #

Later rungs on this anchor: Gauss’s method proper, in three dimensions, with the sector-to-triangle ratio that does the work. Laplace’s method, which needs the second derivative of an angle. Charlier’s theory of the double solution — when two orbits fit three observations exactly, which the four-sighting refusal above is a shadow of. Differential correction and the covariance it produces. The linkage problem: whether two tracklets are the same object. And the impact probability, which is that covariance propagated forward.

Ceres was last seen on 11 February 1801 and recovered on 7 December, near where Gauss’s elements put it. The remarkable part is not that a young mathematician was clever. It is that 41 days of directions, an arc of a few degrees, and no distance of any kind contained a semi-major axis at all — at 0.66% of itself for each arcsecond of error in the angles — and that the same arithmetic says how much shorter the arc would have had to be to contain nothing.