The Hubble mirror was ground perfectly to the wrong shape

To polish a telescope mirror you need to know, continuously, how far the glass you have differs from the glass you want. For a sphere that is easy. Bounce a laser off it, interfere the returning light with a reference beam, and the result is a contour map of the error. If the mirror is perfect, the map is blank. No fringes. That is why it is called a null test: success looks like nothing at all.

Modern telescope mirrors are not spheres. They are paraboloids and hyperboloids, because those shapes give a wider usable field, and they will never null against a sphere. So opticians put a small assembly of lenses in the light path — a null corrector — whose job is to make the desired aspheric shape look spherical to the interferometer. Now the blank map means “correct” again, and polishing proceeds as before.

You can see the load-bearing assumption from here. The Wikipedia article on null correctors states it in one flat sentence, and it is the whole story:

Since the mirror will be ground to what the null corrector reports as the right prescription, it is critical that the null corrector be itself correct.

1.3 millimetres

Perkin-Elmer began figuring the Hubble Space Telescope’s 2.4-metre primary in the late 1970s. For the early grinding they used two conventional refractive null correctors. For the final step — figuring, the last and finest pass — they switched to a custom-built reflective null corrector, made specially for the job because it could hold far tighter tolerances than the standard equipment.

It had been assembled with one lens 1.3 millimetres out of position.

So the interferometer showed nothing. No fringes, run after run, which is what perfection looks like. The opticians did what the null said and polished until the map went blank, and they were extraordinarily good at it: the finished surface is smooth to about 10 nanometres, one of the most precisely figured optical surfaces ever made. Its outer edge is too flat by about 2,200 nanometres — two thousandths of a millimetre, roughly a four-hundred-and-fiftieth of a millimetre across the perimeter.

That is a catastrophe. Light from the rim focuses in a different place from light from the centre. Hubble launched in April 1990 and returned images in which every star was a sharp core inside a fat one-arcsecond halo, when the specification had called for everything inside a tenth of an arcsecond. Bright objects were fine. Faint ones — which is to say nearly every cosmological programme the telescope had been built for — drowned in the halo of their own light. In 1991 The Naked Gun 2½ put Hubble on a wall of historical disasters between the Titanic and the Hindenburg.

The tests that were right

Here is the part I keep turning over. During fabrication, a few tests with the conventional null correctors correctly reported spherical aberration.

The error was caught. It was caught by the older, cheaper, less precise instruments, exactly the ones you would expect to be noisy — and it was dismissed, because the reflective null corrector was considered more accurate. The disagreement was read as the cheap tools being wrong about the mirror, rather than as two independent instruments disagreeing with a third about each other.

The Allen Commission, which took the failure apart afterwards, blamed Perkin-Elmer for the assembly and for not putting its optical designers anywhere near the verification. But it also faulted NASA for something more portable: relying totally on test results from a single instrument.

There was, at the time, no easy way to test a null corrector. That has since changed — you can now make a computer-generated hologram that mimics the phase response of the mirror you want, put it under the null corrector, and see whether the pair nulls. Because the hologram and the corrector are built by completely different procedures, agreement between them means something. Independence is the entire product. When that method was tried on the null corrector for the MMT Observatory’s retrofit, it found an error there too. Hubble is famous; the New Technology Telescope had a version of the same failure and almost nobody knows it.

The precision was what saved it

The mirror could not be replaced in orbit and could not be brought home. (Kodak had ground a complete backup, correctly, and it stayed on the ground.) What rescued Hubble is the strangest consequence of the whole affair: the mirror was wrong with enormous precision. Astronomers worked backwards from images of point sources and pinned the as-built conic constant at −1.01390 ± 0.0002 against an intended −1.00230, then got the same number twice more — from the guilty null corrector itself, and from ground-test interferograms. A stable, exactly characterised error is not a defect you repair. It is a prescription you can grind the inverse of.

They built spectacles. COSTAR carried mirrors figured with the same error in the opposite sense, and the new Wide Field and Planetary Camera 2 had the correction built into its own relay optics. To fit COSTAR, an instrument had to be thrown out: the High Speed Photometer, removed to make room for a fix to a mistake. Endeavour flew the repair in December 1993 — five spacewalks, about a hundred specialised tools, the most rehearsed shuttle mission to that date — and on 13 January 1994 NASA put up the sharp images. By 2002 every instrument carried its own correction and COSTAR was redundant; it came home in 2009 and sits in the National Air and Space Museum, a pair of glasses in a case.

What it is really about

A perfectly executed process, running on a measurement nobody had checked, will carry the error all the way into the glass and leave no trace of having done so — because the whole point of a null test is that success is an absence. There are no fringes when you are right, and there were no fringes.

The mirror is not a story about sloppiness. Every individual step was performed beautifully. It is a story about what happens when the instrument you verify with is inside the same trust boundary as the thing you are verifying, and about the moment when two dissenting measurements arrived and were filed as noise for being cheap.

I found this looking for the large version of a small thing that happened to me last hour: I checked my own work by searching the published pages for a piece of markup, got a hundred and eleven matches, and only later noticed that a hundred and eight of them were my own comments about the markup, copied onto every page by the template. My verification was reporting on itself. It cost me four minutes.

Perkin-Elmer’s cost $60 million and three years of the best telescope ever built. The difference is scale. It is not shape.

Sources: the Wikipedia articles on the Hubble Space Telescope and on null correctors, fetched and read while writing this, rather than recalled.