Bandwidth Smearing: How Wide Channels Blur a Radio Map

What it is and why it matters

A radio interferometer synthesizes an image by correlating pairs of antennas; each baseline measures one Fourier component (a visibility) of the sky brightness at a spatial frequency set by the antenna separation in wavelengths. The array assigns a single (u,v) coordinate per baseline as if the whole receiver operated at one frequency. But real receivers integrate over a channel of finite width Δν. Because the (u,v) coordinate is proportional to ν, every frequency inside that channel actually samples a slightly different spatial frequency. Averaging them together is like a lens focusing red and blue light at different points — chromatic aberration.

The consequence is bandwidth smearing: a compact source away from the delay-tracking (phase) center is smeared radially, its peak flux density reduced while its integrated flux is conserved. It matters enormously for deep, wide-field surveys — VLA-COSMOS, LOFAR Boötes, GMRT TGSS — where faint sources scattered across the primary beam must be photometered accurately. Miss the correction and you underestimate peak brightness, misclassify point sources as extended, and bias source counts.

The mechanism, step by step

Start with the visibility phase. For a source at position (ℓ,m) offset from the phase center, the fringe phase on a baseline is φ = 2π(uℓ + vm), and the components (u,v,w) are the baseline projection measured in wavelengths, so (u,v) = (b/λ)·(direction cosines) ∝ ν. Fix the physical baseline b and vary ν across the channel: the sampled (u,v) point slides radially outward with increasing frequency, tracing a short line rather than a dot.

Now the correlator averages visibilities over the passband. The delay-tracking hardware perfectly compensates the phase only at the band-center frequency ν₀ and only for a source exactly at field center. For an off-axis source, the residual phase 2π(w·)(ν − ν₀)/ν₀ varies linearly across the band. Averaging e^{iφ(ν)} over Δν partially cancels the signal — a sinc-like decorrelation. In the image plane this manifests as a smear along the radial direction, because the (u,v) displacement is radial. The farther off-axis, the larger the phase gradient, and the deeper the peak-flux loss.

The characteristic numbers and the key criterion

The severity is captured by a single dimensionless parameter combining the fractional bandwidth and the source offset measured in synthesized beamwidths: β = (Δν/ν₀) × (θ₀/θ_HPBW), where θ₀ is the angular distance from the phase center and θ_HPBW is the synthesized-beam half-power width. For a Gaussian passband and beam, the peak-response reduction (NRAO's working formula) is

I / I₀ = 1 / √(1 + 0.46 β²).

So smearing worsens with (i) wider channels, (ii) higher resolution (smaller θ_HPBW magnifies θ₀/θ_HPBW), and (iii) larger offsets. Two regimes bracket the physics. A low-resolution VLA D-configuration observation with a few-percent channel may lose only a few percent at the primary-beam edge. A VLBI experiment — beam of milliarcseconds, sources tens of beams off-axis — can be smeared to invisibility unless channels are made extremely narrow. The equivalent radial blur in the image is Δθ ≈ (Δν/ν₀)·θ₀, exactly proportional to fractional bandwidth times offset.

How it is detected and corrected

The observational signature is unmistakable: an off-axis point source appears radially elongated, elongation growing linearly with distance from field center, with reduced peak but conserved integrated flux. Fit a Gaussian and you find the deconvolved 'size' points straight at the pointing center. Surveys quantify this directly — at a primary-beam radius the VLA loses of order 3–4% peak to bandwidth smearing (and another 2–3% to time-average smearing).

The cure is spectral windowing: split the receiver band into many narrow channels and image each near-monochromatically, then combine — the modern technique of multi-frequency synthesis with per-channel gridding, or wide-band imaging in CASA/WSClean that grids each channel at its own (u,v). Correlators now deliver thousands of channels (LOFAR, MeerKAT, the ngVLA design) precisely so that each channel's fractional bandwidth is tiny. Where full spectral resolution is impractical, survey pipelines apply an analytic Bridle & Schwab (1989) correction as a function of offset, channel width, and integration time to restore peak fluxes.

Where it operates and what it is not

Bandwidth smearing operates wherever an interferometer averages a finite band: the VLA and JVLA, LOFAR at 150 MHz, the GMRT at 150/610 MHz, MeerKAT, ALMA continuum, and especially VLBI where beams are milliarcseconds. It grows toward the primary-beam edge, so it is the classic limiter of the usable field of view for point-source photometry, alongside the primary beam itself.

Distinguish it carefully from its sibling, time-average smearing (or time smearing): there the (u,v) point drifts because Earth rotation moves the baseline during a finite integration Δt, blurring sources tangentially in arcs rather than radially. Both conserve flux, both worsen off-axis, but their geometry is orthogonal. Neither is the primary-beam attenuation (a real sensitivity roll-off, not a blur), nor the delay/w-term distortions of non-coplanar wide fields (handled by w-projection). And unlike a fixed instrumental PSF, smearing is spatially variant — it cannot be modeled as convolution with one point-spread function, which is why ordinary CLEAN deconvolution does not remove it.

Open questions and significance

Bandwidth smearing is textbook physics — Bridle & Schwab formalized it for the synthesis-imaging community around 1989 — but it remains a live engineering constraint as arrays push to enormous fractional bandwidths and huge fields. The SKA, ngVLA, and wide-band VLBI want octave-plus bandwidths (Δν/ν approaching 1) for sensitivity, which is exactly the regime where a single average channel would smear the whole field into radial mush. The answer is brute-force channelization plus wide-band, wide-field imagers that grid every channel independently — computationally expensive at SKA data rates.

Open practical problems: correcting smearing self-consistently alongside direction-dependent calibration and the w-term in a single wide-field solver; propagating the residual smearing into flux-scale and source-count error budgets for μJy surveys; and, for VLBI mosaicking across a primary beam, choosing channel widths and dump times that keep every field position below a target peak-loss threshold. Because it caps the trustworthy field of view, controlling bandwidth smearing directly sets how many faint radio sources a survey can measure per pointing — and thus its scientific reach.

Bandwidth smearing vs. time-average smearing: two chromatic-aberration-like blurs in synthesis imaging, with their scalings and characteristic peak-flux losses at the primary-beam half-power point.

Property	Bandwidth smearing	Time-average smearing
Physical cause	Finite channel width Δν; (u,v) ∝ ν	Finite integration time Δt; (u,v) rotates with Earth
Blur direction	Radial (toward/away from field center)	Tangential (perpendicular, arc-like)
Scales with	Fractional bandwidth Δν/ν × offset	Integration time Δt × offset × sidereal rate
Peak loss at PB half-power (VLA typ.)	~3–4%	~2–3%
Flux conserved?	Yes (integrated flux invariant)	Yes (integrated flux invariant)
Cure	Narrower channels (spectral windowing)	Shorter integration / dumps