Celestial Globes Exposed, Answered

1 · The claim, stated fairly

The argument runs from p. 143 to p. 145 of the printed draft. p. 143 sets up the objection accurately: measure Polaris's elevation from several latitudes, compute how high it must be above a flat plane, and the answers disagree. p. 144 offers the diagram — visual geometry against tangible geometry, with Angle A, what you see, always smaller than Angle B, what is physically there. p. 145 concludes: “The answer is linear perspective!” Its figure shows straight sight-lines converging on Polaris, captioned “This is undeniably wrong!”

Several things here are correct.

• — Linear perspective is real, and the hallway analogy is a fair description of it: parallel lines do converge in angle with distance.
• — Refraction is real, and it genuinely corrupts low-elevation measurements. Navigators correct for it on every sight.
• — The objection is often stated badly. “Do the trig, it doesn't work” is not an argument to someone who thinks the angles are unreliable; the reply usually skips a step.
• — And one calculation in the chapter is exactly right. p. 136 puts Polaris's angular size at 0.0000009° against an eye-resolution limit of 0.017°, and calls it 18,860× smaller. Checking it: Polaris A at roughly 46 solar radii and 433 light years subtends 3.2 milliarcseconds (an arcsecond is 1/3,600 of a degree; a milliarcsecond a thousandth of that), or 8.95×10⁻⁷ degrees, and the ratio is 18,990. The difference from his 18,860 is the distance he used, 430 light years against 433 here; Polaris's distance is genuinely disputed, and nothing here turns on it.

So the disagreement is not about whether perspective exists. It is about how much of it there is.

2 · The multi-observer problem

A single observer can never settle this: whatever one person measures, some arrangement produces it. The constraint bites only when one arrangement has to satisfy every observer at once — and that test needs no claim about perspective or refraction, so it goes first. Start where the chapter starts, with tangible geometry: take the measured angles exactly as they come off the instrument and ask what shape of world they describe. The flat model gets one free parameter, how high the lamp sits, so let it choose that height from one measurement and check the result everywhere else.

The ruler that never changes length

Polaris drops one degree for every 69 miles you travel south, everywhere, to within one percent.

It is 68.7 miles per degree at the equator and 69.4 at the pole, and that one percent is the Earth's oblateness — its flattening at the poles — found by the French Geodesic Missions of 1735–44 when Maupertuis measured an arc in Lapland, La Condamine measured one in Peru, and the two came back different. That expedition was this experiment, run with quadrants and chains three centuries ago, and precise enough to detect a one-percent departure from a sphere; the flat model needs a hundred-fold departure. The distance half of the ruler owes nothing to satellites, for a reader who distrusts GPS: the 1730s arcs were chained on the ground, and a road odometer, a surveyor's traverse or a flight time gives the same 69 miles today. Now anchor a lamp above the pole at the height the model's own 45° measurement implies, 5,009 km, and walk outward.

The same thing as a rate, degree by degree:

Measured answer: 69 miles, at every one of those elevations. A lamp at any finite height has to make the rate climb as you walk away from it — twenty-five-fold across the usable range, a hundred-fold by 5° — and never reaches the equator at all, since it only sinks to the horizon at infinite distance. Stop half a degree short, at 0.5° elevation, and the flat model puts the “equator” 356,680 miles from the pole, half again as far as the Moon. The measured distance is 6,225.

The size that never changes either

A lamp a few thousand miles up is not only at a different angle from different latitudes — it is at a different distance.

From the high Arctic to the tropics you would be almost exactly twice as far from the lamp. That has three consequences.

First, it would get dimmer. Polaris is a point source to the eye, so a distance change shows up as brightness. Double the distance and the light falls to a quarter. (Magnitude is the astronomer's brightness scale: larger numbers are fainter, and one step is a factor of about 2.5 in light.)

A four-fold drop, a star and a half of magnitude, between Alaska and Ecuador. Polaris is a familiar second-magnitude star; at 3.5 it would be an inconspicuous smudge in the tropics.

Second, the constellation would shrink — and this needs no equipment, because it compares two stars. Polaris and Kochab, at the end of Ursa Minor, sit 16.6° apart, and Kochab circles Polaris once a night at that radius, so the prediction is the size and shape of that circle:

Two effects compound: the circle shrinks as 1/distance, and it is foreshortened north–south because you look at a horizontal circle increasingly edge-on as you travel away. So Ursa Minor must halve between Alaska and Ecuador and the Kochab–Polaris circle must flatten into an ellipse. Neither happens: constellations keep their shape and angular dimensions everywhere they can be seen, to within the fraction of a percent that differential refraction costs near the horizon. From 10°N Kochab is not circumpolar — it sets for part of the night — so the comparison is between the visible arcs, which is enough. Giving Kochab its own separate height buys one latitude and breaks all the others.

Third, and most precisely, the disc itself. Interferometry — combining the light of separated telescopes — resolves Polaris's disc at 3.2 milliarcseconds — one measurement from one instrument, the Navy Prototype Optical Interferometer in Arizona (Nordgren et al. 1999, in the sources), not repeated from the tropics, so this is the least accessible of the three checks. The flat model's prediction for it is definite, though: a lamp twice as far away shows a disc half the size. The Sun is the same check run in daylight — its disc, about 0.53° across, is the same width from every latitude — and The Sun Does Not Shrink prices that.

3 · What the visual geometry would have to do

The chapter's answer is that the angles were never tangible in the first place: what a person measures is Angle A, visual, smaller than the physical Angle B. Before that can be tested it has to be given physical content. A fact about light known since Fermat: no medium can know how far away the source is. A ray arriving at a point carries no label saying where it started; what it does next depends only on its direction there and the refractive index there. So a bending field cannot be one thing for a star 433 light years away and another for a source eight light-minutes away — which is what makes the Sun and Moon usable as test objects later.

And perspective is a property of a projection, not of a light ray. It is what happens to rays that have already arrived when they are flattened onto a retina, a sensor or a canvas: two rails converge in a photograph because the projection squeezes them, not because the rails moved. But the angle at which the ray comes in is precisely what the measurement is. A theodolite is levelled against gravity and reads the angle between that level and the incoming ray; it has no brain, no prior experience and no perspective processing, and it returns the same angle a person does.

The figure on p. 145 shows the problem without naming it: what replaces its converging straight lines is a set of curves, one per observer, each bent by a different amount, all produced by the same air. Air does not know who is looking through it. So how much bending is needed? Anchoring at 45°N again:

Note the direction as well as the size. Light bends towards denser air, and above the first few metres the denser air is underneath, so refraction makes objects appear higher. The rescue needs starlight pushed down, by two to three hundred times the real effect that pushes it up.

Pick whichever anchor helps most — it still fails

Those rows are samples; the measurement is 69 miles per degree continuously — at 71°N, at 43½°N, at 12°N — so the correction has to come out right along the whole walk. Written out in full, still anchored at 45°N, it does something the three-row version hides:

The same air, the same star — lifting the ray for an observer in Norway, lowering it for one in Mexico, and passing through exactly zero in between.

Real refraction is zero at the zenith, largest at the horizon, and clear of the surface layer bends one way only, so the friendliest requirement is shaped like that. Two facts fix its ends for every lamp height: at the pole the correction is zero, because Polaris is overhead; at the equator it is a downward push, because a lamp at any finite height above a flat plane stands above the horizon from anywhere, and Polaris is observed on it. Between them the height decides the shape. Below a certain height the sign flips on the way; at or above it the correction is downward everywhere and grows smoothly from pole to equator — monotonic, single-signed, refraction-shaped. That threshold is 3,963 miles, or 6,378 km.

So take it. Raise the lamp to one Earth radius, drop the 45° anchor, and grant the chapter the most physically respectable field its own geometry can produce:

The high-latitude agreement is genuinely good, and it is why the flat picture feels workable to anyone who has only looked at Polaris from the far north. At 80°N this lamp is out by six arcminutes (an arcminute is a sixtieth of a degree; the Moon is about thirty across). That is not luck: a lamp one Earth radius above a flat plane tangent to the sphere at the pole reproduces that sphere, to first order, near the pole. The flat model works precisely where a flat approximation to a sphere works.

Does any of this assume the globe?

Refraction theory was built on a spherical Earth, and the almanac tables were fitted to observations reduced with spherical astronomy. If “real refraction is 0.09°” only ever means “observed altitude minus globe-predicted altitude,” quoting it against a flat model is circular. The site's standing answer — sightline geometry alone cannot separate curvature from refraction, so both models are handed the same inputs and the flat branch is then priced against what air can physically do — is stated once on the Method page; here the pricing is done with a star instead of a shoreline. Set the almanac aside and rebuild the number from two things that have nothing to do with the shape of the Earth.

• 1. The refractive index of air, measured on a bench. n − 1 = 0.000293 at 0 °C and one atmosphere. No sky, no stars, no assumed planet.
• 2. Snell's law across horizontal layers. On a flat plane the air stratifies horizontally — that is what a fluid does in a downward acceleration field, whichever account of the field you prefer.

Together they give an exact conserved quantity along any ray:

n × sin z = the same number at every point on the path

where z is the ray's angle from vertical. It is exact at every angle short of the horizontal, has no Earth radius in it, and the whole profile cancels — inversions, ducts, exotic lapse rates — leaving only the two endpoints:

sin z_observed ÷ sin z_true = n at the source ÷ n at the observer

Three things follow, and none of them needed a globe.

The direction is fixed. Polaris sits above the air, where n = 1; the observer sits at the bottom of it, where n is larger. So the apparent zenith distance shrinks and the star appears higher — always, for every profile there is or could be. Getting it to appear lower needs n at the observer below one: air less optically dense than a vacuum.

The magnitude comes out right. At 45° elevation the formula with bench-measured air returns 1.008 arcminutes; the almanac says 1.0. At 10° the plane-parallel form returns 5.7 arcminutes where the almanac tabulates 5.3 and Bennett's standard formula gives 5.4, the excess being the curvature term a flat Earth does not have — so the flat model is handed the larger of the two. Every “real refraction” column on this page uses the standard 5.4′ (0.090°) rather than the flat-friendly 5.7′. The tables are a check on the argument, not an input, and they pass. Refraction is also measured on the ground, with no astronomy in it: surveyors take reciprocal vertical angles between two towers whose heights come off a levelling staff and whose separation off a tape, and the discrepancy is refraction — roughly 4 arcseconds per kilometre of sight, so a target a kilometre away appears about 2 arcseconds too high, always in that direction. In two centuries of geodesy nobody has had to apply the correction the other way, and a visual-geometry compression of the size the chapter needs would show up on that same target, whose height can be checked with a tape measure.

And it has to be one number for everybody. The ratio in that formula is a property of the air and the two altitudes, so it is one physical constant for every observer looking at the same lamp. Here is what the flat model needs it to be, with the lamp at one Earth radius:

One atmosphere, one lamp altitude, one ratio — and the flat model needs it to hold five values at once, spread across twenty per cent. Even one of them is absurd: a ratio of 1.198 puts the refractive index at Polaris's altitude at 1.198 against 1.000293 at the ground, and since n − 1 tracks density that is air at roughly 870 kg/m³, seven-tenths the density of water, sitting on sea-level air without falling — while the observer at 80°N needs the same air at 45 kg/m³.

Two caveats. The identity assumes n varies with height only; a dome-shaped atmosphere with its optical layers tilted relative to the ground would break it, at the cost of horizontal pressure gradients large enough to read on a barometer and to drive permanent hemisphere-scale winds. And it permits a turning point when a ray reaches horizontal, which is what ducting is; that is a real loophole within a degree or so of the horizon, which is why the equator row is the soft one. It is not available at 45° or 30° of elevation, and those rows fail on their own.

And if it is a lens rather than a layer?

One more mechanism genuinely escapes the identity above, which assumed the air is stratified, its optical layers parallel to the ground. Suppose instead the atmosphere works as a lens, with curved optical surfaces. Either the surfaces lie parallel to the ground everywhere, and n is a function of height alone and the case is closed, or they do not, and the index gradient has a horizontal component. What a lens buys with is the one thing layers can never do: it deflects light sideways. And to move Polaris the lens has to be centred on the pole, so its deflection is radial about the pole, not toward the horizon. For Polaris those are the same thing, which is why the idea is tempting. Nothing else in the sky is due north.

• — Anything off the meridian — the north–south line through the point overhead — gets shoved sideways. The angle between “up” and “toward the pole” is the parallactic angle, and at 45°N it reaches 45° for an object on the eastern or western horizon. A field strong enough to move Polaris several degrees moves that object comparably far horizontally.
• — Which is a clock error, and then a navigation error. Displace azimuth by a degree and, for the equinox Sun at 45°N, its transit — the moment it crosses the meridian — moves by 2.8 minutes. Three degrees is 8.5 minutes, and at sea four minutes of time is a degree of longitude — up to sixty nautical miles, less as you go north. Meridian instruments were the definition of time for two centuries and resolve arcseconds; this proposal introduces degrees.
• — Constellations would be sheared, not merely squashed. Layered air makes circles into ellipses with upright axes everywhere in the sky; a pole-centred lens compresses along the line to the pole, so the axes come out tilted, by a different angle in every direction you look.

The field, written out as a magnification

If the angles are compressed by the amounts above, the compression has a shape, gentle overhead and severe near the horizon. Expressed as a vertical magnification — how much a small vertical gap in the true sky is stretched by the time it reaches the eye, at each apparent elevation:

That is the field the rescue requires. It is not an add-on to the argument; it is the argument, written out with numbers in it.

The field pushes every object down, yet stretches the gaps between objects near the horizon, because it pushes the higher of two objects down by less than the lower one. On the 45°N fit the magnification falls back below 1 above about 62° of apparent elevation — the sign change seen from the other side; the rows quoted are the range the Sun and Moon occupy in §4. They come from the 45°N fit, the weaker requirement: the Earth-radius version needs 25.6° of correction at 10° elevation rather than 19.4°, so everything §4 does with this table understates what the steel-manned version demands.

4 · Grant it anyway — and walk the Sun and the Moon through it

Set every one of those objections aside and suppose the field exists, in exactly the shape §3 computed, however it manages it. A distortion field is not selective, and two of the things that cross it are looked at by everybody, every day.

A static field can always be tuned to a static object. Polaris never moves. The Sun and Moon cross that same sky twice a day, at every elevation from the horizon to overhead.

The field cannot let them through untouched, because the stretching is the whole point of it. The Sun and Moon are each about half a degree across; their width is unaffected, since the distortion is vertical, and their height is multiplied by the table above.

What is observed is the opposite, and far smaller. Refraction squashes the Sun vertically — wider than tall — by 0.8% at 10° elevation and 0.2% at 20°. The only distortion anybody sees without instruments is the oval Sun in the last degree before it sets, a squash of about a fifth: standard refraction lifts the lower limb some 6 arcminutes more than the upper, off a disc 32 arcminutes tall. That is the entire observable catalogue.

The same test also settles the lens question from §3, because the Sun walks the whole sky and so samples the field's orientation as well as its strength:

5 · Two smaller things

The chapter's own arithmetic refutes its parallel-light claim

p. 140 argues that because Polaris is a tiny dot, its light must spread radially and cannot arrive as parallel rays — that textbook diagrams “visually blow Polaris up to the size of Earth” to make parallelism look plausible. But the correct calculation four pages earlier settles this the other way. Angular size is precisely the quantity that bounds how non-parallel a source's light can be. All light from Polaris arrives inside a cone 3.2 milliarcseconds wide — the chapter's own figure — and rays reaching opposite sides of the Earth differ in direction by 6.4×10⁻⁷ arcseconds, roughly 1,500 times finer than the best optical interferometer can resolve. The smaller Polaris is correctly shown to be, the more parallel its light must be. The textbook diagram is not a trick, it is undrawable: with Earth as a 1 cm dot, Polaris belongs 3.2 million kilometres off the page, about eight times the distance to the Moon.

Perspective cannot manufacture a second pole

“Polaris is invisible from the southern hemisphere” is not by itself decisive: on a flat map the far south is simply far out, and a star sinking to the horizon with distance is what the model already claims. What perspective cannot do is produce a second centre of rotation. A vanishing point makes things disappear; it does not create a new pivot for everything else to turn around. From Ecuador, a camera pointed north records the sky turning anticlockwise about Polaris just above the horizon while a camera pointed south records it turning clockwise about σ Octantis just above the opposite horizon — simultaneously, from the same spot. A flat plane has one centre, or none. The sky has two.

6 · What would settle it

• — Photograph the Sun at 10° and at 45° on the same day, same lens. Measure the width and height in pixels. The rescue in §3 requires a 3.3× vertical stretch at the lower elevation; refraction predicts a 0.8% squash. These are not close, and the equipment is a phone and a tripod.
• — Drive 69 miles south and measure Polaris. Then do it again from a different latitude. If the drop per 69 miles changes with latitude, the flat model is in business. If it stays at one degree, it isn't.
• — Shoot a rising Moon and a high Moon. Compare the maria. Any vertical distortion of the required size is unmissable in the pattern, not just the outline.
• — Point a theodolite at a target of known height a kilometre away. Tape-measure the height independently. If visual angles are compressed, this is where it shows up at a scale anyone can check.

The caveat that matters is the one in §3: the sign change in the 45°N fit is an artefact of an anchor we chose, so the verdict rests on the Earth-radius version and on the Sun's shape, not on the sign flip.

The invitation this page is answering

“I may be wrong about the stars, and I am open to changing my mind… Could someone more knowledgeable than I take my angular resolution analysis to the next level?”
— Levi Miller, Globe Deconstruction, p. 131 (review draft)

That sentence is taken at face value. Everything above is offered as the answer to it: the arithmetic run out to where it stops working, the strongest version of the rescue built, and the places where the chapter is right marked as such. The Sun's shape is not a measurement anyone has to be trusted about, which is the best kind of answer to have.

Sources & further reading

• Refraction figures: G. G. Bennett, “The calculation of astronomical refraction in marine navigation,” Journal of Navigation 35 (1982) — the standard approximation used in navigation almanacs, scaled here by pressure and temperature — and A. T. Young's refraction pages for the flattened Sun and the sinking/looming glossary.
• The oblateness measurement: Maupertuis, La Figure de la Terre (1738), and La Condamine's Peru arc — the French Geodesic Missions of 1735–44, which are this experiment run three centuries early.
• Latitude by Polaris at sea: Nautical Almanac Pole Star tables and Bowditch, American Practical Navigator — three centuries of navigators using altitude = latitude with a fixed correction.
• Polaris's angular diameter: Nordgren et al., ApJ 118, 3032 (1999), NPOI interferometry, 3.28 mas. The distance remains disputed — Turner et al., ApJ 762, L8 (2013) against the Hipparcos value.
• The Moon illusion: Kaufman & Rock, Science 136 (1962); Ross & Plug, The Mystery of the Moon Illusion (2002).
• Levi Miller, Globe Deconstruction? (prerelease draft, 2026), section “Celestial Globes Exposed,” printed pp. 124–152. The Polaris height argument is at pp. 143–145; the angular-size calculation at p. 136; the parallel-light claim at p. 140.
• Oriani's theorem and the stability of refraction above 10–15° altitude: A. T. Young, “Astronomical Refraction”, San Diego State University — the standard modern reference on the subject.
• Record atmospheric extremes used for the ceiling in the method notes: highest sea-level pressure 1083.8 hPa (Agata, Siberia, 31 December 1968); lowest recorded temperature −89.2 °C (Vostok Station, 21 July 1983).
• Companion pages: the standing refraction argument on the Method page, sunlight bends in air, and the oval-Sun consequence, the angular-resolution chapter answered, and does the Earth spin, or does the sky.