Very Long Baseline Interferometry

Why the baseline is the telescope

The angular resolution of any aperture is set by diffraction: θ ≈ λ / D, where D is the diameter and λ the wavelength of light being collected. To resolve angles small enough to see structure across a distant galaxy nucleus, you need either very short wavelength or very large D. At radio wavelengths — millimetres to metres — the required diameter for microarcsecond imaging is bigger than any single dish ever built, or ever buildable. The Arecibo dish, the largest fixed single radio aperture, was 305 m. To match what a 1.3 mm space telescope would do at the same resolution as the Event Horizon Telescope, you would need a dish about 12,000 km wide.

What you cannot build, you can synthesise. Two antennas separated by a baseline B observe the same source and record its electric-field voltages. When the wavefront arrives, it strikes one antenna slightly before the other — by a delay τ = (B/c) sin θ, where θ is the source's angle off the perpendicular bisector of the baseline. A small change in θ changes τ by a small fraction of a wavelength, which the correlator can detect as a phase shift. The pair of antennas behaves, for one Fourier component of the sky brightness, exactly like the two slits of a Young's experiment scaled up: the longest baseline sets the finest fringe spacing the array can resolve. That fringe spacing is λ/B — the same diffraction limit as a dish of diameter B.

This is the entire conceptual basis of interferometry. The shocking part of VLBI is that it works even when the two antennas are not physically connected to each other at all — and even when they sit on different continents.

Where VLBI came from

Radio interferometry was demonstrated by Martin Ryle in Cambridge in the 1940s and 1950s with cabled arrays; aperture synthesis using Earth rotation followed in the 1960s and earned Ryle a share of the 1974 Nobel Prize. But cables limit you to a few kilometres. The leap to intercontinental baselines required two engineering insights: clocks stable enough to keep phase coherent for long enough, and storage media wide enough to ship the data.

The first true VLBI demonstration came in 1967, in a roughly simultaneous burst of activity from several US groups — Bernard Burke, Norman Broten, Barry Clark, Marshall Cohen, David Jauncey, Kenneth Kellermann and others — using independent rubidium clocks at separated dishes, magnetic tape recorders, and post hoc cross-correlation. The earliest published fringe between widely separated stations was on 13 cm wavelength between Algonquin (Ontario) and the Dominion Radio Astrophysical Observatory (British Columbia) in May 1967. A series of papers in 1967-68 from MIT, NRAO, and the Canadian groups quickly extended baselines to thousands of kilometres and showed that the technique could resolve the cores of quasars, then-newly discovered and unresolved by any other instrument.

The result was a step change in resolution of about three orders of magnitude. Compact extragalactic sources that had been points became measurable objects with sub-milliarcsecond structure. Within a decade VLBI had discovered superluminal motion in 3C 273 and 3C 279 — apparent transverse velocities of several times c — interpreted as a relativistic-beaming projection effect in jets pointed nearly at Earth.

The resolution formula in practice

The working number for an interferometer is

θ_res ≈ λ / B_max  (radians)
     ≈ (λ_mm / B_km) × 206  microarcseconds

So at λ = 13 cm with B = 6,000 km, θ_res ≈ 4 milliarcseconds. At λ = 7 mm with the VLBA (B ≈ 8,600 km), θ_res ≈ 0.17 milliarcseconds. At λ = 1.3 mm on the EHT (B ≈ 12,000 km), θ_res ≈ 22 microarcseconds. The conversion table:

Network	Typical λ	B_max	Resolution	Era
Cambridge One-Mile	21 cm	1.6 km	~27 arcsec	1960s (connected)
VLA (D-config)	21 cm	1 km	~46 arcsec	1980
VLA (A-config)	21 cm	36 km	~1.3 arcsec	1980
EVN	6 cm	~9,000 km	~1.4 mas	since 1980
VLBA (10 dishes)	7 mm	8,600 km	~0.17 mas	since 1993
VSOP / HALCA	18 cm	~21,000 km	~1.7 mas	1997-2003
RadioAstron	1.3 cm	~350,000 km	~8 μas	2011-2019
EHT	1.3 mm	~12,000 km	~20 μas	since 2017

Twenty microarcseconds is small enough to read a credit card in Tokyo from a café in Madrid — or, more usefully, to resolve a structure of a few Schwarzschild radii around a 6.5 × 10⁹ M☉ black hole at 17 Mpc.

How the signal flows from sky to image

A VLBI observation has four stages. At each antenna, the radio signal from the source is mixed with the local oscillator down to baseband, sampled by an analog-to-digital converter, and written to a large-capacity recorder against a precise timestamp from the local hydrogen maser. The data rates are punishing: modern VLBI runs at 4–32 gigabits per second per station, which for a 6-hour observation is tens to hundreds of terabytes per dish. Network bandwidth between observatories is rarely sufficient, so for EHT and many VLBA campaigns the disk packs are physically flown to the correlator — usually by FedEx. This is the "FedExNet" joke that is more truth than humour.

At the correlator, each pair of stations has its streams cross-correlated. The correlator searches over a range of geometric delays τ — accounting for the predicted Earth-rotation contribution, source position, troposphere, and clock offsets — to find the value that maximises the cross-correlation. The peak gives a complex visibility V_ij at one point in the u,v-plane for one frequency channel at one time. Modelling, calibration, and self-calibration then turn the table of visibilities into a deconvolved image, usually with CLEAN or its variants.

The pipeline is unforgiving. A 1-millisecond clock error at one station washes out fringes entirely; a 1 mm change in the assumed station coordinate shifts the phase at 1.3 mm by 360 degrees. Atmospheric delays must be modelled or solved out at every minute, because troposphere water vapour swings the wet delay by several centimetres over weather timescales. The whole pipeline is therefore an exercise in chasing down phase budgets, and a successful VLBI observation is one where every error has been catalogued and absorbed into a calibrator solution.

Aperture synthesis: filling the dish with Earth rotation

A single pair of antennas at any instant samples one Fourier component of the sky brightness — one point (u, v) in the spatial-frequency plane. As Earth rotates beneath the source, the projected baseline traces out an ellipse in u,v-space, sweeping the sampled point across many components over the course of hours. An array of N antennas has N(N−1)/2 pairs, each tracing its own ellipse. Together they fill the u,v-plane with a sparse but elaborate set of samples; an inverse Fourier transform then yields an image, deconvolved against the "dirty beam" (the point-spread function of the sparse sampling) via the CLEAN algorithm of Högbom (1974).

V(u, v) = ∫∫ I(l, m) exp[ -2π i (u l + v m) ] dl dm   (van Cittert-Zernike)
I(l, m) ≈ F⁻¹ { S(u, v) · V(u, v) } * Beam_dirty

That second equation is the heart of every interferometric image you have ever seen: the apparent brightness is the Fourier transform of sparsely sampled visibilities, convolved with a beam dictated entirely by the array geometry and the duration of the observation. The whole point of Earth rotation is to thicken the sampling so the dirty beam shrinks.

The networks

VLBI is done by international consortia because no one institution can fund the global infrastructure. The major networks are:

• VLBA (Very Long Baseline Array). Ten identical 25 m NRAO dishes spanning the US from Mauna Kea, Hawaii to Saint Croix in the Virgin Islands. Maximum baseline 8,611 km. Operates continuously from 0.3 GHz to 86 GHz. Workhorse for AGN imaging, pulsar parallax (e.g. PSR distances at 10% precision out to ~10 kpc), and maser astrometry.
• EVN (European VLBI Network). A consortium of ~20 antennas across Europe, plus collaborators in South Africa, China, Russia, and Korea. Heterogeneous dishes (Effelsberg 100 m, Jodrell Bank Mk II, Westerbork, Medicina, Onsala, Yebes, etc.) operating from 0.3 GHz to 43 GHz. Correlated at JIVE in the Netherlands.
• LBA (Long Baseline Array, Australia). Southern-hemisphere counterpart based on the ATCA, Parkes, Hobart, and Mopra dishes, important for sources south of declination −20°.
• EHT (Event Horizon Telescope). Global 1.3 mm VLBI assembled ad hoc from millimetre observatories: ALMA, APEX, IRAM 30 m, JCMT, SMA, LMT, SMT, SPT. Operated as observing campaigns (April 2017 was the famous M87 observation), not as a year-round facility. The Greenland Telescope and Kitt Peak's KP12m have been added in later campaigns.
• Space VLBI. Japan's VSOP/HALCA (1997-2003, λ = 6, 18, 1.3 cm, baseline up to 21,000 km) and Russia's RadioAstron / Spektr-R (2011-2019, λ = 1.3, 6, 18, 92 cm, apogee ~360,000 km) extended baselines beyond Earth's diameter, briefly reaching ~7 microarcseconds.
• Geodetic VLBI (IVS). The International VLBI Service for Geodesy and Astrometry operates regular 24-hour sessions across a global geodetic network, using extragalactic quasars to measure baseline vectors and Earth orientation parameters at sub-millimetre and sub-microsecond precision.

Imaging black-hole shadows: M87* and Sgr A*

The defining EHT results are the silhouettes of two supermassive black holes, both ringed by their lensed accretion flow. The M87* image released in April 2019 — and Sgr A* in May 2022 — were the first direct images of structure on event-horizon scales.

M87* has mass M ≈ 6.5 × 10⁹ M☉ at distance D ≈ 16.8 Mpc, giving a Schwarzschild angular diameter

θ_S = 2 GM / (c² D)
    ≈ 7.6 microarcseconds

The photon-ring diameter for a non-spinning hole is 5.2 × θ_S ≈ 40 μas, matching the measured ring at 42 ± 3 μas. For Sgr A*, M ≈ 4.1 × 10⁶ M☉ at D ≈ 8.1 kpc gives θ_S ≈ 10 μas and predicted ring ~52 μas — measured at 51.8 ± 2.3 μas. In both cases the inferred ring diameter agrees with general relativity to within current calibration uncertainty.

The asymmetry of the rings — bright on one side, dim on the opposite side — is the signature of relativistic Doppler boosting in the accreting plasma orbiting the hole. The orientation of the bright crescent for M87* points to a black-hole spin axis aligned with the kpc-scale jet, and the polarisation map released in 2021 reveals an ordered magnetic field on horizon scales consistent with the magnetically arrested disk (MAD) regime.

Worked example: do we have the baseline we need?

Suppose you want to resolve a hotspot orbiting Sgr A* at the ISCO. The Sgr A* ISCO sits at roughly 3 GM/c² for a moderately spinning hole, which subtends about

θ_ISCO ≈ 3 × (4.1×10⁶ M☉) × G / (c² × 8.1 kpc)
       ≈ 15 microarcseconds

Required baseline at λ = 1.3 mm:

B = λ / θ_ISCO
  = 1.3×10⁻³ m / (15×10⁻⁶ × π/648000 rad)
  ≈ 18,000 km

That is longer than Earth's diameter (12,742 km) — so a ground-only array cannot resolve a single hotspot at the ISCO of Sgr A*. To break this barrier, the planned next-generation Event Horizon Telescope (ngEHT) is adding stations and, more ambitiously, the Black Hole Explorer (BHEX) concept proposes a 0.87 mm-capable space VLBI satellite that would push the baseline to ~30,000 km and the resolution to sub-10 microarcseconds. Movies of orbit-scale dynamics at Sgr A* — actual frames of plasma swinging around the hole — are the goal of the 2030s.

Non-imaging applications

• Geodesy. Daily 24-hour IVS sessions deliver baseline vectors with millimetre repeatability. Continental drift (Atlantic widening at ~25 mm/yr, Pacific convergence at ~70 mm/yr at the Andes) is mapped directly. The current International Terrestrial Reference Frame (ITRF) station velocities are anchored against VLBI.
• Earth orientation. VLBI is the only technique that gives UT1 (the difference between rotation-based time and atomic time) at high precision, because only VLBI provides an absolute reference to the extragalactic frame; GNSS solutions for polar motion and length-of-day must be tied to VLBI.
• Celestial reference frame. The ICRF is defined by the positions of ~300 extragalactic compact sources measured with VLBI. The current ICRF3 (adopted by the IAU in 2018) is accurate to about 30 microarcseconds in source positions.
• Pulsar parallax. Astrometric VLBI determines pulsar positions and proper motions with ~0.1 mas precision, enabling parallax distance measurements out to ~10 kpc and feeding the Galactic neutron-star kinematics.
• Spacecraft tracking. Delta-DOR (differential one-way ranging) at NASA's Deep Space Network uses VLBI of a spacecraft signal against a nearby quasar to measure angular position to ~1 nanoradian, essential for outer-planet navigation.

Where VLBI hits its limits

• Sensitivity. Collecting area of an array is the sum of individual dishes, not the synthesised aperture. For a sparse mm-VLBI array of 10 m-class antennas the brightness sensitivity is set by ~100 m² × bandwidth × integration — sufficient for bright quasar cores but inadequate for thermal extragalactic emission. This is why EHT targets are essentially limited to the brightest AGN and SMBHs.
• u,v-coverage. A handful of stations leaves the u,v-plane mostly empty. Images are unique only up to information held in the sampled visibilities; CLEAN and Bayesian methods like RML interpolate but cannot recreate genuinely missing baselines. This drives the push to add more stations.
• Atmospheric phase. At millimetre wavelengths the troposphere swings the phase by hundreds of degrees per minute. Without a phase reference (a nearby calibrator within the isoplanatic patch, or paired-antenna phase referencing) the phase information is lost — only closure quantities (closure phase, closure amplitude) survive, and these are harder to image with.
• Field of view. The primary beam of each dish (λ/D_individual) sets the maximum field of view. A 25 m VLBA dish at λ = 1.3 cm has FoV ~ 2 arcminutes; at 1.3 mm with a 12 m EHT dish, FoV is ~22 arcseconds. Mosaicking is mostly not done in VLBI — observations are single-pointing on a single compact source.
• Surface accuracy. The dishes must have a surface accuracy of ~λ/16 to image at wavelength λ. At 1.3 mm that is ~80 microns of figure error across the entire surface. This is one reason EHT involves only specially designed mm-wave observatories.

Common pitfalls and misconceptions

• Confusing baseline with diameter. The synthesised resolution θ ≈ λ/B treats B as the longest baseline, not the average. An array with one 10,000 km baseline and ninety-nine 100 km baselines has the resolution of the long one but the u,v-coverage of a small array — so resolution is achieved but image quality is poor.
• Treating phase calibration as automatic. Real VLBI calibration is a multi-week effort. Self-calibration only converges if the model is already close; for resolved sources the initial model has to come from independent observations or assumptions.
• Reading raw images literally. A VLBI image is the true brightness convolved with a beam dictated by sparse u,v sampling. Features at scales below the beam are not real; features above the largest angular scale (set by the shortest baseline) are filtered out. The image is a band-limited reconstruction, not a photograph.
• Equating VLBI resolution with EHT resolution. Most VLBI runs at centimetre wavelengths and milliarcsecond resolution. EHT's microarcsecond performance requires going to 1.3 mm, which restricts the network, the sources, and the season.
• Forgetting that data ships physically. The volume of raw VLBI data is huge enough that internet links are usually slower than aircraft. The 2017 EHT campaign produced about 5 petabytes; the South Pole disks travelled by sea after the Antarctic winter ended.