Closure Phase: Beating Atmospheric Errors With Three Interferometer Baselines

What closure phase is and why the atmosphere forces you to use it

An interferometer measures a complex visibility V = |V|·e^{iφ} for each pair of telescopes — the Fourier component of the sky brightness sampled at that baseline. The amplitude |V| encodes how resolved the source is; the phase φ encodes its position and asymmetry. In principle the phases reconstruct the image. In practice, the turbulent atmosphere adds a rapidly varying, essentially random phase offset above each telescope.

Crucially, that corruption is station-based: a single error term θ_A belongs to telescope A and contaminates every baseline that includes A. The measured phase on baseline A–B is therefore φ_AB^obs = φ_AB^true + θ_A − θ_B (with a sign convention). No single baseline phase is trustworthy — the θ terms can be many radians and change on millisecond timescales.

• The insight: arrange telescopes into a closed triangle and add the three phases; the station terms telescope out.
• The payoff: a self-calibrating observable that needs no external phase reference and survives turbulence, instrumental drift, and even the ionosphere.

The mechanism: why station errors cancel around the loop

Write the observed phase on each leg of the triangle A→B→C→A, keeping the station error terms:

• φ_AB^obs = φ_AB + (θ_A − θ_B)
• φ_BC^obs = φ_BC + (θ_B − θ_C)
• φ_CA^obs = φ_CA + (θ_C − θ_A)

Sum them into the closure phase Φ_ABC = φ_AB^obs + φ_BC^obs + φ_CA^obs. Every station error appears exactly twice with opposite sign: +θ_A…−θ_A, +θ_B…−θ_B, +θ_C…−θ_C. They cancel identically, leaving

Φ_ABC = φ_AB + φ_BC + φ_CA, a purely source-intrinsic quantity.

Because phase noise wraps at 2π, in real data you don't literally add angles; you multiply the complex visibilities to form the triple product (or bispectrum) B = V_AB · V_BC · V_CA, and take Φ = arg(B). The bispectrum is the Fourier transform of the triple correlation, so it is the natural, wrap-safe carrier of the same cancellation. A point source has Φ = 0 exactly; any nonzero closure phase is an unambiguous signature of source asymmetry.

Key quantities: how much phase information you actually recover

For an array of N telescopes, the number of independent visibility phases is (N−1); the atmosphere destroys one per station, i.e. N unknowns constrained. The bookkeeping works out cleanly:

• Number of possible triangles / closure phases: N(N−1)(N−2)/6
• Number of independent (non-redundant) closure phases: (N−1)(N−2)/2
• Fraction of the true phase information recovered: (N−2)/N

Worked example: for N = 3 you get 1 closure phase and recover 1/3 of the phase information. For the CHARA Array's N = 6, that's 10 independent closure phases and 4/6 ≈ 67%. For the Event Horizon Telescope's ~8 stations you recover 6/8 = 75% of the phase. The lesson: closure phase never fully substitutes for the lost phases, but the deficit shrinks toward 100% as arrays grow — which is why big VLBI networks lean on it so heavily. Modern combiners reach exquisite precision: VLTI's four-telescope GRAVITY instrument delivers closure-phase accuracy near 1 degree, enough to detect companions at contrasts of a few×10⁻³.

How it is observed, from radio VLBI to optical aperture masking

Closure phase is measured wherever the atmosphere makes single-baseline phases useless:

• Radio/mm VLBI: The EHT and VLBA combine widely separated dishes where clocks and the troposphere ruin absolute phase. Closure phases (and closure amplitudes) are the primary robust data products fed to image reconstruction.
• Optical/IR long-baseline interferometry: VLTI (PIONIER, GRAVITY, MATISSE) and the CHARA Array (MIRC-X) form triangles among separate telescopes to image stellar surfaces, binaries, and disks.
• Sparse aperture masking (SAM): Place a mask with a handful of holes over a single large telescope's pupil behind adaptive optics. Each hole pair is a baseline; triangles of holes yield closure phases that beat residual AO wavefront errors, reaching contrasts at or inside the formal diffraction limit.

In every case the pipeline forms the triple product, averages the bispectrum to beat down noise, and passes the phases to algorithms such as CLEAN, MEM, or regularized-maximum-likelihood imagers (e.g. eht-imaging), often inside a self-calibration loop that iteratively solves the station terms.

How it differs from its cousins

Closure phase belongs to a small family of closure quantities, and it helps to keep them straight:

• Closure amplitude does for gain (amplitude) errors what closure phase does for phase, but needs a four-telescope loop: it is the ratio (|V_AB|·|V_CD|)/(|V_AC|·|V_BD|), in which multiplicative station gains cancel.
• Kernel phase (Martinache 2010) generalizes closure phase to a filled or redundant aperture, linearizing the small-error regime so that even a single AO-corrected dish yields self-calibrating observables.
• Phase referencing / fringe fitting is the alternative: chase and remove the atmospheric phase directly (e.g. with a bright nearby calibrator or a dual-feed system like GRAVITY's). It recovers absolute astrometry that closure phase cannot, but demands extra hardware and a suitable reference.

The trade-off is fundamental: closure phase throws away the absolute position information (the 'flux centroid' and overall astrometric phase) in exchange for immunity to errors. It is a relative, asymmetry-sensitive observable — perfect for structure, blind to where the source sits on the sky.

Significance, famous cases, and open questions

Closure phase is arguably the single idea that made ground-based aperture-synthesis imaging possible. Roger C. Jennison proposed it in 1958 at Jodrell Bank; it was adapted to very-long-baseline radio work in the 1970s (Rogers et al. 1974) and later carried into the optical/IR by groups building COAST, IOTA, VLTI, and CHARA.

• Event Horizon Telescope, 2019 & 2022: the ring images of M87* and Sgr A* rest on closure phases and amplitudes, because 230 GHz VLBI has no usable absolute phase. The measured closure phases constrained the ring's asymmetry and hence the emission geometry near the shadow.
• Stellar imaging: closure-phase interferometry resolved the surfaces and spots of stars like Altair and Betelgeuse and mapped interacting binaries.
• Exoplanet & companion hunting: nonzero closure phases flag faint asymmetric companions; SAM on Keck, VLT, and JWST/NIRISS (AMI mode) exploits this.

Open issues: closure quantities discard some information, so image reconstruction from them is formally under-constrained and prior-dependent — a live debate after the EHT results. Non-closing 'systematic' errors from polarization leakage, bandwidth, and non-station effects still limit precision, and optimal ways to weight and de-correlate closure data remain active research.

Closure phase versus a raw visibility phase and closely related interferometric observables

Observable	What it measures	Immune to station errors?	Info recovered / notes
Raw visibility phase	Phase on one baseline	No — corrupted by both telescopes' atmosphere	Unusable on the ground without phase referencing
Closure phase	Phase of triple product around 3-telescope loop	Yes — station terms cancel exactly	(N−2)/N of the phase information; N(N−1)(N−2)/6 measurable, (N−1)(N−2)/2 independent
Closure amplitude	Ratio of visibility amplitudes on a 4-telescope loop	Yes — gain errors cancel	Needs 4 telescopes; complements closure phase
Kernel phase	Linearized closure phase from a filled/redundant pupil	Yes — to first order in small errors	Generalizes closure phase to single-dish AO imaging