A detail can emerge from the difference between two recordings

Two radio antennas point at the same galaxy. In each antenna’s individual record, two compact sources may blend together. Yet a wave from one direction reaches the antennas with one delay, while a wave from a slightly different direction has another. The difference between recordings contains information that the broad response of a single antenna does not reveal on its own.

An interferometer compares aligned signals from several telescopes to measure a source’s spatial structure. The separation and direction between a pair of antennas define its baseline. A long projection of that baseline onto the sky plane makes the comparison sensitive to small angular changes. The empty space between antennas still collects no radiation.

We need one starting point from the previous article on radio telescopes: the antenna and receiver first produce an electrical signal, not a finished photograph. Those are the signals an interferometer uses. A radio wave remains electromagnetic even when its recording resembles noise. We will follow how comparing recordings leads to an image, and why that image needs checking.

Radio sources first had to be separated on the sky

Early radio maps lacked resolution: neighbouring sources blended, and their positions were difficult to match to optical galaxies. In his 1974 Nobel lecture, Martin Ryle explained how that problem drove his group toward instruments with greater resolving power. Building a precise, continuous kilometre-wide surface was far harder than separating smaller receiving elements.

Aperture synthesis became crucial: different antenna pairs and positions measure different pieces of information, which computation combines. Ryle describes a 1960–1961 experiment with Ann Neville testing Earth rotation and map computation. Cambridge’s One-Mile Telescope began observations in 1964. These are stages in the method’s development, not a claim that all interferometry appeared on one day or belonged to one person. Ryle’s primary lecture, PDF pages 1–6.

Another advance removed the need for a shared cable between widely separated stations: record signals against a time reference and compare them later. Successful experiments by Canadian and US groups in 1967 marked the practical beginnings of very-long-baseline interferometry, VLBI. For the particular VLBA system, observations with all ten stations began on 29 May 1993; that is not VLBI’s invention date. GSI: the method’s history, NRAO archive timeline.

Delay the earlier signal until the later one arrives

Imagine a distant point source. Across two antennas, its wavefront can be treated as planar. If that front is tilted relative to the ground baseline, one antenna meets it earlier. What matters is not the baseline’s entire length but its component toward the source: that gives the extra path.

In the diagram, antenna B has already received a particular front, while that same front must still travel 50 m to reach A. Hold back B’s recording for the corresponding time and we compare the same part of the wave. It is the early channel that must be delayed. Accidentally delaying the later channel only worsens the mismatch.

The same wave reaches B before A

Toward sourceOne wavefrontQ→A: 50 mQABb = 100 m · τg = 166.782 ns
Our exact planar example: A=(0,0), B=(100,0) m, source direction s=(0.5,√3/2), both axes 3.2 px/m. The dashed line is one wavefront through B and Q; Q→A adds 50 m, or 166.782 ns. Black arrows show wave arrival; olive points toward the source. Antenna symbols are schematic phase centres, not ray traces of their reflectors. Delay the earlier B recording. Equations and limits.

Earth rotates during an observation, changing both the source direction relative to the baseline and the required delay. The correlator tracks that change. For a broad band, rotating the phase at just one frequency is insufficient: the time records must be aligned and residual phase accounted for as well. A timestamp helps locate corresponding samples, but does not replace a stable frequency reference and calibration. NRAO: geometric and compensating delays, §3.7.1 and §3.7.3.

Branch · With equationsOptional calculation: delay sign, baseline direction and the limits of λ/B.

Path difference and angular sensitivity involve different projections

Let vector b point from A to B, and unit vector s from the array toward the source. Define a positive delay as later arrival at A. For a plane wave in vacuum:

\Delta L=\mathbf b\cdot\mathbf s,\qquad \tau_g=t_A-t_B=\frac{\Delta L}{c}.

Here ΔL is in metres, c=299792458 m/s and τg in seconds. In the diagram, θ=30° from vertical and b=100 m horizontally, so ΔL=b sinθ=50 m and τg≈166.782 ns. For a source directly above the baseline, ΔL=0. Reversing the antenna order reverses the sign.

Geometric phase φg=2πΔL/λ changes with angle. For this two-dimensional example, dφg/dθ=2πb cosθ/λ. Thus the local angular fringe period is set by another projection, B⊥=b cosθ:

δθ_(fringe) ≈ (λ)/(B_(⊥)).

This is a scale in radians, not a universal criterion for separating two objects or measuring positions. It applies along the relevant direction; the synthesized response also depends on all baselines and their weights. Near the direction along the baseline itself, B⊥ approaches zero even though the delay is large. A large time difference alone does not mean the best resolution in every direction.

A correlator retains the shared part of the signals

A celestial radio signal is often not a pure sinusoid. It can be noise-like, while its electric field at two antennas remains statistically related after delay correction. Correlation measures that relationship: aligned signals in corresponding frequency channels are multiplied and averaged. Two quadrature components preserve both magnitude and phase information. Together they form the complex visibility.

Sinusoids provide a useful foothold: aligned rises and falls give one average product; a shift by part of a cycle changes it. This is a teaching example, not a requirement that a galaxy transmit a stable musical note. Radiation from different independent sky regions contributes to the visibility; those regions need not radiate in synchrony with one another.

Independent receiver noise does not contribute the same stable mutual signal, but a finite correlation estimate still has noise. Longer averaging reduces random uncertainty only under suitable conditions. It cannot repair receiver saturation or shared radio interference. NRAO: complex correlation and sensitivity, §3.7.2 and §3.7.6.

There is another operation: add aligned voltages to form one narrow receiving beam. This phased sum can supply a pulsar time series or an input to a larger VLBI array. Pairwise correlation preserves a set of information for imaging; one completed sum is not that same dataset. One array can support both modes. ALMA’s phasing system does exactly this, turning local antennas into one VLBI station. Matthews et al.: the ALMA Phasing System, §III and §V.2.

From antenna voltages to two different array outputs

Stage functionWhat moves onward
  1. Antenna and feed at each stationRadio-frequency voltage · V
  2. Amplification, selection and mixingIntermediate-frequency voltage · V
  3. Digitization of each channelNumerical samples · ADU; sampling rate in sample/s
  4. Transmission or time-stamped recordingADU samples + time (s) + frequency (Hz) + polarization
  5. Delay and phase compensationAligned ADU samples for the chosen direction
  6. Choice of subsequent operationThe same aligned channels can feed different modes

Separate reference inputs: LO sets frequency conversion; the sample clock sets sampling instants; timestamps place samples on a time scale. A connected array may share a reference; VLBI uses local references with residual differences estimated. These are not extra celestial signals.

Phased sum

Add aligned voltages → one aggregate ADU stream → power detection/time series or VLBI recording. That stream alone does not retain the complete set of separate visibilities.

Pairwise correlations

Multiply by the complex conjugate and average → complex visibilities in relative ADU² → calibrate the scale in Jy → initial map and reconstruction, usually in Jy/beam. N stations provide N(N−1)/2 distinct cross-pairs.

Our functional diagram, not the exact wiring of a particular instrument. ADU is a numerical sample unit; real chains require defined digital scaling and calibration coefficients. Writing units does not turn an uncalibrated product into flux density in Jy. One array may support both modes. Equations and limits.

Each baseline tests its own scale and direction

A single antenna pair does not draw an entire galaxy. It tests how strongly the brightness distribution matches a particular spatial pattern: alternating brighter and darker bands. A long projected baseline corresponds to finer bands; a short one to broader bands. Rotating the baseline changes the pattern’s orientation.

Four antennas provide six distinct pairs. If they all stand on one line, the range of orientations is limited. A two-dimensional arrangement helps, while Earth rotation gradually changes the projections seen from the source. This accumulates additional spatial frequencies. The source must remain sufficiently stable for the chosen combination method; rapid variability cannot simply be treated as a static sky.

These samples appear as points in the uv plane. It maps measured scales and directions, not ground positions or galaxy pixels. Coordinates u and v are projected baseline components in wavelength units. Even many tracks leave gaps between them. NRAO: Earth rotation and spatial frequencies, §3.7.4–3.7.5.

Ground positions and spatial-frequency samples are different maps

Ground plane · E,N (m)

ABCD01000200010002000E (m)N (m)

Sky plane · u,v (kλ)

-10-1001010u (kλ)v (kλ)

H=0 · highlighted points · Samples over 6 h

Our hypothetical geometry, not an observatory layout. (E,N) m: A(0,0), B(60,0), C(500,300), D(1700,1300); λ=0.21 m, latitude 45°, source declination +30°. Six baselines are sampled at H from −3 to +3 h in 0.5 h steps; colour highlights H=0. Each point has a conjugate counterpart: 78 samples plus 78 conjugates, not 156 independent measurements. Lines join selected times without filling missing frequencies. The line-of-sight component w is included in the geometry but not plotted. Equations and limits.

Branch · Mathematical modelOptional branch: the quantities being compared and when a two-dimensional formula applies.

Why visibility is related to a Fourier transform

For a narrow frequency channel, a small field, a distant source and proper phase-centre tracking, we use the approximation:

V(u,v) = ∬A(l,m)I_(ν)(l,m); × (e)^( − 2πi(ul + vm)) dl dm.

Iν is spectral brightness, for example Jy/sr; A is the dimensionless primary-beam response; l,m are small direction cosines relative to the centre. The product dl dm approximates solid angle, so V has units of Jy. One jansky equals 10⁻²⁶ W·m⁻²·Hz⁻¹. Coordinates u,v are projected baseline lengths divided by λ. The minus sign defines our complex-visibility convention: the preceding branch’s geometric φg is positive, and its corresponding phase factor here is e⁻ⁱφg.

Each sky region’s contribution is rotated by its own phase, and the contributions are summed. For a real A Iν distribution, opposite visibilities are conjugates: V(−u,−v)=V*(u,v). Both points are therefore often plotted, although they are not two new independent measurements of one pair.

A wider field or a large line-of-sight w component needs an additional phase term. A rough check of the flat-field approximation is small wθ², with θ in radians. Source spectral changes, polarization, direction-dependent errors and non-coplanar baselines can also require a richer model. The formula does not promise that a single “inverse Fourier transform” yields the true sky. CASAdocs: the measurement operator and reconstruction.

Missing data leave traces in the image

To see what incompleteness does, first define a sky we know exactly. Our separate computational example contains three point pixels and a uniform background. Visibilities are calculated from that scene, some spatial frequencies selected, and a direct initial reconstruction performed. These are synthetic data, not an observatory recording or a physical observation of our own.

The reconstruction from 12 modes has strong repeated structures. A larger artificial set of 136 modes makes the point pattern more recognizable, but does not recover the missing zero frequency. Blue negative regions are reconstruction artifacts, not sources of “negative light”. The scale displays them rather than clipping them to zero.

Known input and initial reconstructions from incomplete data

Injected scene

mJy / pixel

-12-12001212Offset · pixels

12 modes · dirty image

mJy / dirty beam

-12-12001212Offset · pixels

136 modes · dirty image

mJy / dirty beam

-12-12001212Offset · pixels

Colour: fixed numerical range −0.7…1.1

−0.7 · Negative0 · Zero1.1 · Positive
Our periodic 25×25 scene: 1/0.65/0.4 mJy point injections at pixels (8,15)/(17,11)/(13,5), plus 0.02 mJy/pixel background. A separate artificial DFT selection: 12 modes and 136 modes with 0<kₓ²+kᵧ²≤41; these are not tracks from the preceding geometric model. Axes give offsets from central pixel (12,12). Input is mJy/pixel; maps have unit-peak dirty-beam normalization, in mJy/dirty beam. Negative values remain visible; fixed numerical colour range −0.7…1.1, no automatic contrast. No observational noise, CLEAN or deconvolution. The missing zero mode is not recovered. Equations and limits.

The initial map is a dirty image, and the response to an ideal point is the dirty beam, or synthesized point-spread function. Its sidelobes distribute part of a source’s signal elsewhere. Deconvolution separately seeks a source model that explains the data while accounting for that response. CLEAN or regularized reconstruction methods introduce additional assumptions; they do not turn unmeasured frequencies into measured ones. Wilner: sampling, known scene and dirty image, slides 26–28.

Short baselines matter for extended structures. A compact array and properly calibrated single-dish measurements help recover large scales. That is why ALMA includes both more closely spaced 7 m antennas and separate 12 m total-power antennas. These make different contributions; all 66 antennas do not operate in one identical mode. ALMA: main array, ACA and total power.

Branch · Checking reconstructionOptional discrete example: normalization, sidelobes and an unrecoverable frequency.

An unknown background can leave the same data

The example has N=25 pixels per axis, with the origin at the central pixel. The forward transform sums pixel fluxes with phase weights. The sparse set contains 12 conjugate-symmetric integer modes; the dense set contains all 136 nonzero integer pairs satisfying kₓ²+kᵧ²≤41. These are deliberately selected teaching masks, separate from the physical uv geometry above. They are not rounded samples of its tracks.

For M selected modes, the initial reconstruction is:

D_{pq}=\frac1M\sum_{(k_x,k_y)\in S} V_{k_xk_y}\,e^{2\pi i(k_xp+k_yq)/N}.

Normalization by M gives a unit dirty-beam peak for a unit point injection. The input is therefore labelled mJy/pixel and the dirty map mJy/dirty beam: identical colours do not imply identical pixel photometry. All three panels have an explicit fixed numerical scale, with negative values retained. Periodic boundaries, no noise and equal weights are model assumptions. No CLEAN, regularization or final restored map is calculated here.

Input: points of 1,0.65,0.4 mJy above a 0.02 mJy/pixel background, totalling 14.55 mJy. Add another 0.05 mJy to every pixel. The zero mode changes by 25²×0.05=31.25 mJy, but none of the selected nonzero modes changes. These data cannot determine that added uniform background. This is an exact ambiguity in the periodic model, not a claim that every finite real cloud produces strictly zero signal on every baseline.

An independent check is circular convolution of the input scene with the dirty beam. It gives the same initial maps; the full set of all 625 modes returns every input pixel. This checks the calculation even when the picture already looks persuasive.

A large baseline does not replace a large collecting area

Compare four hypothetical 10 m dishes on a 2 km baseline with a filled 2 km aperture. At λ=0.21 m, λ/B⊥ for a suitably projected baseline is about 21.66 arcseconds. But the four dishes have only about 314 m² of geometric area, equivalent to a single circle 20 m across. A filled 2 km circle has roughly 3.14 million m².

Four dishes: a 2 km baseline, the area of a 20 m circle

Four 10 m dishes

Geometric area

314.16 m²

Circle with the same area

Geometric area

314.16 m²

D = 20 m

Filled circle 2 km across

Geometric area

3,141,592.65 m²
A separate geometric example, not an ALMA specification or the array in the preceding diagram. This is a numerical area comparison without a scale drawing. Four D=10 m circles have 0.01% of a filled D=2000 m circle’s area. At λ=0.21 m and B⊥=2000 m, λ/B⊥≈21.66″; this does not give equal collecting area or universal sensitivity. Equations and limits.

Actual effective area also includes losses and aperture illumination. Sensitivity further depends on system noise, bandwidth, time, usable pair count, polarizations and processing weights. The geometric areas are therefore not a finished sensitivity specification. Moving existing antennas farther apart does not increase their area, although it changes the available details.

A faint extended cloud may benefit more from a short baseline or single dish; a compact galactic nucleus may need a long baseline. “More kilometres” is not a universal answer to every observing problem.

An array is built as a network of measurement channels

Each station needs a suitable antenna of its own: verified panels or another receiving structure, feed, support and pointing system. Accuracy must suit the wavelength. Synthesis does not remove material and surface requirements. For example, GSI’s historical 32 m Tsukuba antenna had aluminium panels and a steel support structure. That describes a particular instrument, not a common specification for all VLBI stations. GSI: antenna materials.

The receiver, frequency conversion, digitizer and time reference are then integrated. Boards and connectors must carry the required band without missing data, swapped polarizations or uncontrolled delays. A correlator contains actual electronic modules, connections and cooling, not just a formula. Fifty simultaneously cross-correlated antennas have 50×49/2=1225 distinct pairs; that is not the number of all possible pairs among 66 antennas. ALMA: correlator construction.

Antenna locations are also part of the instrument. Supports must fix position and orientation, and coordinates must enter the delay model. ALMA documents concrete pads with three-point attachments and connections for power, reference signals and data. Optical fibre carries encoded signals and frequency references; the original celestial radio photon does not travel through it between dishes. ALMA: foundations and optical fibre.

Acceptance has separate levels: antenna surface and pointing; receiver noise, band and linearity; timestamps, packets and electrical delays; then correlations on a known source. Detecting fringes confirms that channels can be aligned and correlated, but alone does not establish correct photometry, coordinates or a complex reconstruction. This is an engineering explanation based on documentation; the author has not manufactured or physically tested an array.

Separate the shared sky from channel errors

Unequal gains change amplitudes, while unaccounted delays rotate phases. The atmosphere adds path length that can change differently above different antennas. A constant phase offset need not reduce a visibility’s magnitude: it can rotate it. Rapid changes during averaging, however, cancel parts of the complex signal and reduce its coherent response.

A calibrator is a source with a suitable known or testable model. Observations of it help estimate channel responses over time and frequency. Flux scale, polarization leakage, timing corrections and data quality are checked separately. A weak or resolved calibrator cannot automatically be treated as an ideal point. Corrections follow a model and usable signal, not arbitrary adjustment toward an attractive picture. CASAdocs: calibration stages.

At millimetre wavelengths, ALMA also uses water-vapour radiometers to estimate rapid changes in added path length. This is an auxiliary atmospheric measurement, not communication between stars. It does not make the atmosphere perfectly known. VLBI also requires estimates of residual time and frequency differences between independent stations; a stable maser and correct date do not remove every correction. ALMA Phasing System: maser and phase corrections, §IV.1 and §V.2.

Radio interference can correlate too. Flagging spoiled channels or time intervals removes data and changes coverage; that must enter the response and uncertainty assessment. Useful checks include comparisons across stations, times and frequencies, inspection of residuals, and reconstruction with other justified settings. Self-calibration does not measure missing scales. NRAO: RFI and calibration limits.

Branch · Measurement checksOptional branch: station phases, closure and loss of coherence.

What cancels around a triangle, and what remains

In a simple scalar model, one frequency/polarization channel has complex gain gi. For pair i,j:

(V_(ij))^(obs) = g_(i)(g_(j))^(∗)(V_(ij))^(true) + n_(ij).

The asterisk denotes complex conjugation; nij is noise. Temporarily ignoring additive noise, phase contains the station-phase difference ψi−ψj. Around an oriented closed triangle:

Φ_(123) = arg⁡(V_(12)V_(23)V_(31)).

Station phase errors cancel in the product of gi gj*. This is the closure phase; it need not be zero for an arbitrary complex scene. It does not remove errors belonging to a particular baseline or statistical uncertainty at low signal strength. Closure consistency does not prove an absence of RFI either. NRAO: equations on slides 10–11 and caveats on 12–13.

Another check concerns averaging. For a uniform rectangular band Δν and residual timing error δτ, the coherent response magnitude is multiplied by |sinc(Δνδτ)|, with sinc(x)=sin(πx)/(πx). At Δν=1 MHz and δτ=0.25 μs this is about 0.9003; at δτ=1 μs it is zero in the ideal model. For an independent zero-mean Gaussian phase fluctuation with RMS σφ in radians, the factor is exp(−σφ²/2); RMS=1 rad gives about 0.607. These are different coherence-loss models, not universal instrument specifications.

At large separations, the way signals are brought together changes

In VLBI, each station receives at coordinated times, digitizes against its own stable frequency reference and retains timestamps. Data can be sent over a network or delivered as recordings for later correlation. There is no need to wait for a wave to “reach the other antenna by cable”: both observe the same sky, and computation matches their recordings. A hydrogen maser stabilizes the local reference; residual clock differences still have to be accounted for. IVS: recordings, clocks and correlation.

Long baselines reveal compact galactic nuclei and jets, radio-source positions and motions. Geodetic quasar observations also help determine station positions, crustal motion and Earth orientation. Source and geometry assumptions depend on the task; imaging is not an interferometer’s only output. NASA GSFC: VLBI geometry and geodetic applications.

ALMA can first form a local phased sum and supply it as one sensitive element of an international array. Such a station needs dedicated calibration and quality control. The paper on the 2017 campaign describes that special mode; it should not be confused with an ordinary local ALMA image. Goddi et al.: calibrating phased ALMA.

In optical and infrared interferometry such as the VLTI, light beams are physically brought to a shared instrument. Moving optics change the internal path length; turbulence, vibration and much shorter wavelengths make coherence harder to maintain. ESO documents a particular delay line with gold-coated aluminium mirrors, a carriage, fine piezo motion, laser metrology and monitoring of thermal rail distortion. These are real mechanical and measurement components, not simply a digital delay applied to a file. ESO: VLTI Delay Lines.

Segments of one large mirror also need wavefront alignment, but form a shared aperture with a different optical path and platform. Dividing a surface into segments does not itself turn a telescope into a VLBI network with independent recordings.

Trust features that survive checks against the data

An interferometer resolves fine details because the relative signals of separated antennas are highly sensitive to arrival direction. Imaging needs a set of baselines, calibration and reconstruction with a known incomplete response. Antenna area determines the actual signal collected; baseline determines an accessible angular scale; coverage and errors determine how reliably that scale can be used.

A scientific example is the EHT work on M87 images from April 2017 observations, published in April 2019. The authors compared independent reconstructions and tested methods on synthetic data with known scenes. They assessed whether main features survived changes of reconstruction method; an attractive picture alone was not the test. Our teaching example uses none of their data and does not reproduce that entire analysis. EHT, Paper IV: reconstruction methods and limits.

You can now make two checks using our diagrams: identify which channel must be delayed, and explain why an added uniform background did not return from the nonzero modes. Both answers follow from what was actually measured. This also helps interpret reports of an “Earth-sized telescope”: a large baseline really provides a fine angular scale, while the gaps remain empty as far as collecting radiation is concerned.