Interferometry: Turning Scattered Dishes Into One Giant Eye

The diffraction wall, and the loophole in it

Every telescope is a hole cut in a wavefront, and every hole diffracts. A circular aperture of diameter D spreads a point source into an Airy pattern whose first dark ring lies at θ ≈ 1.22 λ/D radians. That is not a polishing defect to be engineered away; it is the wave nature of light insisting that you cannot pin down a direction more finely than the width of the wavefront you sampled.

The wall is easy to see in numbers. Hubble's 2.4 m mirror gives about 0.05″ in the visible, and JWST's 6.5 m about 0.08″ at 2 µm. But radio waves are hundreds of thousands of times longer, so the 100 m Green Bank Telescope at 21 cm delivers a beam roughly 9 arcminutes across — a third the width of the full Moon. To reach one milliarcsecond at 21 cm with a single dish you would need an aperture of about 53,000 km, four Earth diameters.

The loophole is that D is not really “the area of glass.” It is the largest separation between two points of the wavefront that you sample coherently. Take only two small patches, separated by a distance B, and let them interfere: the angular scale printed on the sky is λ/B. The middle is missing, but the finest detail survives. This splits the two jobs a telescope does, and the split is the whole bargain of the field:

• Resolution is set by the baseline B, the longest separation between elements.
• Sensitivity is set by total collecting area, the sum of the individual dishes and nothing more.

Two 25 m dishes 8,600 km apart resolve like a telescope the size of a continent and collect like two 25 m dishes. Interferometers are ferociously sharp and, for their sharpness, comparatively blind.

Fringes, visibility, and the theorem that makes it imaging

Point two apertures separated by B at the same star and combine the beams. Light arriving from an angle θ off the baseline normal travels an extra B sinθ to the far aperture, so the combined intensity rises and falls as cos² with an angular fringe period of λ/B. A true point source gives fringes of perfect contrast: bright to black.

Now make the source extended. A star is spatially incoherent — every patch of photosphere radiates independently — so the patches add in intensity, each contributing fringes shifted by its own position. Patches separated by more than half a fringe period contribute stripes in antiphase, and the contrast washes out. Define the fringe visibility V = (I_max − I_min) / (I_max + I_min), and it becomes a direct measure of source size compared with λ/B.

The van Cittert–Zernike theorem (Pieter van Cittert, 1934; Frits Zernike, 1938) makes this exact and turns contrast into imaging: for a distant, spatially incoherent source the complex visibility on a projected baseline is the normalised two-dimensional Fourier transform of the sky brightness,

• V(u, v) = ∫∫ I(l, m) e^{−2πi(ul + vm)} dl dm ÷ ∫∫ I(l, m) dl dm, with u and v the projected baseline components measured in wavelengths.

Each pair of telescopes, at each instant, returns one complex number: the fringe contrast is one Fourier amplitude of the sky, and the fringe position is its phase. Long baselines probe fine structure, short ones coarse. For a uniform disk of diameter θ_UD the transform is 2J₁(x)/x with x = πθ_UDB/λ, and the fringes vanish entirely at θ_UD = 1.22 λ/B. You need no image to measure a star's diameter — only the baseline at which the stripes disappear.

Making the wavefronts arrive together: delay, coherence, correlators

Interference happens only if the two wave trains overlap in time. A source at unit vector š reaches one telescope earlier by the geometric delay τ_g = (B · š)/c, up to about 33 milliseconds on a 10,000 km baseline and changing at up to ΩB/c ≈ 2.4 microseconds per second as the Earth turns. Every interferometer must continuously subtract a predicted delay computed from a model that includes Earth orientation, solid-Earth tides, the troposphere and even plate motion.

Radio does this in electronics. Below a terahertz the electric field itself can be sampled with its phase intact, so each antenna heterodynes the sky signal down to baseband against a local oscillator locked to a hydrogen maser — fractional stability of order 1 part in 10¹⁵, less than a picosecond of drift over a ten-second integration — then digitises and timestamps it. In VLBI the telescopes never touch: the Event Horizon Telescope recorded petabytes onto helium-sealed disk packs that were physically flown to correlators at MIT Haystack and the Max Planck Institute for Radio Astronomy in Bonn, with the South Pole disks waiting months for the first flight of the austral summer.

Optical and infrared interferometry cannot cheat this way. Any phase-preserving amplifier must add noise (the quantum limit formalised by Carlton Caves in 1982), so the light has to be piped to a common beam-combiner through vacuum tunnels using movable delay lines — at the VLTI, carriages rolling on rails in a 100 m-plus underground tunnel — matching path lengths to a few tens of nanometres. The tolerances are brutal: the coherence length λ²/Δλ is only about 12 µm for a broadband near-infrared channel at 2.2 µm, and the atmosphere shuffles the relative piston by microns every ~10 ms, so fringe trackers must close a servo loop at kilohertz rates. Hence dozens of radio interferometers and only a handful of optical ones.

Aperture synthesis: letting the Earth do the scanning

One baseline gives one Fourier component; an image needs many. Martin Ryle and his Cambridge group realised in the 1950s that they come almost free. As the Earth turns, the baseline projected onto the sky as seen from the source rotates and foreshortens, sweeping an elliptical arc through the uv-plane over a night. Elements help combinatorially: N antennas give N(N−1)/2 simultaneous baselines, so the VLA's 27 antennas give 351 and ALMA's fifty 12 m dishes give 1,225. The One-Mile Telescope (1964) and 5 km Ryle Telescope (1971) proved the method, and Ryle shared the 1974 Nobel Prize in Physics with Antony Hewish, his half of the citation naming aperture synthesis explicitly (Hewish's named the discovery of pulsars).

Reconstruction is an inverse Fourier transform of the sampled visibilities, but because the sampling is sparse the output is the dirty image: the true sky convolved with the dirty beam, the transform of the sampling pattern, riddled with sidelobes. Deconvolution is unavoidable. Jan Högbom's CLEAN algorithm (1974) repeatedly finds the brightest peak, subtracts a scaled dirty beam there, records a delta-function component, and finally restores the component list with a clean Gaussian beam. Modern practice adds self-calibration, maximum-entropy and regularised-maximum-likelihood methods; the EHT deliberately ran independent pipelines with blind parameter surveys because reconstruction from sparse data is choice-laden.

Phase is the fragile part, corrupted by atmosphere and clocks. The rescue is the closure phase, introduced by Roger Jennison in 1958: because the corrupting errors are station-based, the sum of visibility phases around a triangle of telescopes cancels them exactly. N stations yield (N−1)(N−2)/2 independent closure phases out of N(N−1)/2 baselines, so a fraction (N−2)/N of the phase information survives calibration errors untouched. The southward brightening of the M87* ring is a closure-phase result, which is why it was believed.

From a 20-foot beam on Mount Wilson to an Earth-sized telescope

Hippolyte Fizeau proposed the method in 1868, and Édouard Stéphan tried it at Marseille in 1874 with a mask over an 80 cm telescope, finding every star still unresolved — which correctly implied that all stellar diameters were below about 0.16″. The breakthrough came on 13 December 1920, when Albert A. Michelson and Francis G. Pease mounted a 20-foot steel beam carrying movable flat mirrors across the top of the 100-inch Hooker telescope. Betelgeuse's fringes died away at a mirror separation of 121 inches (3.07 m), giving θ = 1.22λ/B ≈ 0.047″ — the first angular diameter ever measured for a star other than the Sun. Modern near-infrared interferometry returns a uniform-disk diameter of roughly 42–45 mas. A parallel branch ran through Robert Hanbury Brown and Richard Twiss, whose Narrabri intensity interferometer (1963–1974) correlated intensity fluctuations rather than field amplitudes and measured diameters for 32 hot stars.

Radio then took over. The VLA, dedicated in 1980, spread 27 antennas along a Y with a 36 km maximum baseline; the VLBA (1993) placed ten 25 m dishes from Mauna Kea to St. Croix, 8,611 km apart, resolving 0.17 mas at 7 mm. ALMA (2013), 66 antennas at 5,000 m on Chajnantor, imaged the protoplanetary disk of HL Tauri in 2014 with a ~35 mas beam, resolving concentric rings and gaps a few au wide. In the near-infrared, CHARA on Mount Wilson combines six 1 m telescopes over baselines to 331 m and has imaged stellar surfaces, including the flattened, gravity-darkened photosphere of the rapid rotator Altair. ESO's VLTI feeds four 8.2 m Unit Telescopes into GRAVITY, which tracked the star S2 through its 2018 pericentre passage around Sgr A* with astrometry at the few tens of microarcseconds level and detected both the gravitational redshift and, in 2020, the Schwarzschild precession of its orbit.

The summit so far is the Event Horizon Telescope. In April 2017 eight stations — ALMA, APEX, the IRAM 30 m, JCMT, SMA, SMT, the LMT and the South Pole Telescope — observed at 1.3 mm on baselines reaching about 10,700 km, a nominal resolution near 20 µas. On 10 April 2019 the collaboration published the ring of M87*, 42 ± 3 µas across, implying 6.5 × 10⁹ solar masses at 16.8 Mpc; on 12 May 2022 came Sagittarius A*, a ring of about 52 µas around a ~4 × 10⁶ solar-mass black hole 8.3 kpc away. Both match the ~10 gravitational radii predicted for a shadow. Sgr A* was far harder: its gravitational radius corresponds to only ~20 light-seconds, so the source changes visibly during the hours needed to fill the uv-plane.

What interferometry is not, and where it fails

It is not a bigger bucket. The most stubborn misconception is that the EHT is “an Earth-sized telescope” in every sense. It has Earth-sized resolution and the sensitivity of eight dishes, which is why only two black holes on the entire sky are simultaneously bright enough and angularly large enough to image.

It is not photography. An interferometer never records pixels, only a sparse scattering of Fourier components; everything between them is supplied by an algorithm plus a prior. The EHT ring is robust because closure quantities and several independent pipelines agree, but the fine texture inside any such image is inference rather than measurement.

It is blind at both ends. The shortest baseline sets the largest angular scale an array can see; smooth emission broader than λ/B_min is “resolved out,” and the total flux — the u = v = 0 component — is never measured at all. ALMA fixes this missing-short-spacings problem by folding in its compact 7 m array and four total-power dishes. At the other end, the field of view is capped by each dish's own primary beam (λ/D_dish) and shrunk further by bandwidth smearing and time-average smearing away from the phase centre.

Look-alikes. LIGO and Virgo are Michelson interferometers, but they measure a single scalar path-length difference and localise sources by timing between sites, not by synthesising an aperture. Adaptive optics corrects atmospheric wavefront distortion on one aperture, restoring λ/D but never beating it — which is exactly why the VLTI runs AO on each Unit Telescope and interferes them. Speckle imaging is likewise capped at λ/D. Intensity interferometry is a genuinely different measurement: it correlates photon-rate fluctuations, discards phase, needs bright sources, and is nearly immune to path-length errors, which is why it is being revived on Cherenkov arrays such as VERITAS and MAGIC.

Open questions and the next baselines

The obvious frontier is a longer baseline, and the Earth is now the constraint. Space VLBI has already flown — Japan's HALCA/VSOP from 1997 and Russia's RadioAstron (Spektr-R, 2011–2019), whose apogee near 350,000 km gave baselines close to 30 Earth diameters — but never at millimetre wavelengths. Millimetre-wave space concepts aim to resolve the sharp, lensed photon ring around M87*, a feature whose diameter and shape are fixed by the spacetime metric almost independently of the messy astrophysics of the accretion flow.

Nearer term, the EHT is extending to 345 GHz (0.87 mm) and adding stations under the ngEHT programme, chasing not a sharper still frame but a movie of Sgr A* — which demands imaging algorithms that treat the source as time-variable rather than static across an eight-hour track. On the ground, the SKA (SKA-Low in Western Australia, SKA-Mid in South Africa with baselines to ~150 km) and the proposed ngVLA will raise sensitivity by orders of magnitude, and calibration at those dynamic ranges is itself an unsolved problem.

In the optical, the open question is whether the technique can scale at all. One idea is quantum-repeater-assisted interferometry — distributing entanglement so photon states can be combined without a physical light path, proposed by Gottesman, Jennewein and Croke in 2012 — which would do for optical astronomy what disk packs and hydrogen masers did for radio. A separate line, nulling interferometry (Ronald Bracewell, 1978), deliberately places a destructive fringe on the star so that a planet a fringe-width away survives; the LBTI has used it to measure exozodiacal dust around nearby stars, and mid-infrared space concepts such as LIFE are designed to hunt ozone and methane in temperate rocky atmospheres. In every case the physics is one century-old relation — resolution equals wavelength over baseline — pushed until the engineering, not the wave, is what gives way.

Angular resolution: filled apertures versus interferometers. Resolution follows the largest separation sampled; sensitivity follows total collecting area.

Instrument	Wavelength	Aperture or maximum baseline	Angular resolution
Hubble Space Telescope	550 nm (visible)	2.4 m filled aperture	~0.05″
Green Bank Telescope (single dish)	21 cm (1.4 GHz)	100 m filled aperture	~9′ (about 530″)
Karl G. Jansky VLA, A configuration	7 mm (43 GHz)	36 km baseline (27 × 25 m)	~0.04″
ALMA, extended configuration	1.3 mm (230 GHz)	16 km baseline (up to fifty 12 m antennas)	~20 mas
Very Long Baseline Array	7 mm (43 GHz)	8,611 km baseline (10 × 25 m)	~0.17 mas
Event Horizon Telescope, 2017	1.3 mm (230 GHz)	~10,700 km baseline (8 stations)	~20–25 µas