Space Instruments

Correlation (VLBI)

Q: What is the difference between an FX and an XF VLBI correlator?

An XF correlator first computes the cross-correlation in the time domain by multiplying and accumulating lagged sample products, then Fourier transforms the resulting lag spectrum to obtain frequency channels. An FX correlator reverses the order: it Fourier transforms each station's voltage stream first, then cross-multiplies the per-channel spectra. The FX approach scales far better for large station counts because its cost grows roughly as N log N per station plus N-squared per channel cross-multiplies, rather than the N-squared times lags cost of an XF lag machine. Modern systems such as DiFX, the VLBA software correlator, and the EHT correlators are almost all FX designs running on CPU or GPU clusters.

Q: Why must a delay model and fringe rotation be applied before integrating?

Each baseline sees the source at a different geometric delay because the wavefront reaches the two antennas at different times, and Earth rotation makes that delay change continuously. The correlator computes the geometric delay tau_g from the baseline vector and source direction, then shifts one station's samples by the integer-plus-fractional delay and counter-rotates its phase by the fringe rate 2 pi nu (d tau_g / dt). Without this, the rapidly spinning interferometric phase, which can exceed several kHz at centimeter wavelengths on long baselines, would average to zero over even a one-second integration and destroy the fringe. Accurate delay models use clock offsets, Earth orientation parameters, and tropospheric and ionospheric terms.

Q: How does coarse one-bit or two-bit quantization affect correlation sensitivity?

VLBI records enormous data volumes, so samples are quantized coarsely. One-bit (two-level) sampling at Nyquist recovers only about 0.637 of the correlation amplitude, costing roughly 1.96 dB of sensitivity, while two-bit (four-level) sampling with optimum thresholds recovers about 0.881, costing only about 0.55 dB. The correlator applies the Van Vleck correction to convert the measured digital correlation coefficient back to the true analog correlation. Two-bit sampling is the common modern compromise: it doubles the bit rate versus one-bit but reduces the quantization loss enough that wider recorded bandwidth gives a net sensitivity gain.

/kor-uh-LAY-shun, V-L-B-I/

At the heart of Very Long Baseline Interferometry, the correlator takes the independently recorded, digitized voltage streams from antennas separated by hundreds to thousands of kilometers and performs a time-aligned cross-correlation between every pair. Because the stations share no common cable, each must be aligned against a precise geometric delay model and counter-rotated to remove the fringe phase before integrating. The output of each baseline is a complex visibility: amplitude and phase samples of the source brightness distribution in the spatial-frequency (u, v) domain. Typical recorded bandwidths span 256 MHz to 4 GHz per polarization at recorded data rates of 0.5 to 32 Gbps per station, and after fringe rotation the residual rate is small enough that the phase turns only a fraction of a cycle over a 0.5 to 2 second accumulation.

Category: Space Instruments

Output: Complex visibility V(u,v)

Architecture: FX or XF correlator

From Recorded Voltages to Complex Visibilities

In a connected-element array the antenna signals travel down cables to a central building where they are combined in real time. VLBI cannot do this: the baselines are too long, so each station digitizes its intermediate-frequency band, time-stamps every sample against a local hydrogen-maser clock, and records to disk or transfers over fiber. The correlator is the offline machine that finally brings these streams together. Its job is to estimate, for each pair of stations, the time-averaged product of their electric-field voltages, which by the Van Cittert-Zernike theorem equals a single Fourier component of the sky brightness distribution sampled at the projected baseline coordinate (u, v).

The challenge is that the two recordings are not aligned in time. The wavefront from the source reaches the more distant antenna later by the geometric delay τ_g = −(b · s) / c, where b is the baseline vector and s the unit vector to the source. This delay can reach tens of milliseconds on intercontinental baselines and changes continuously as the Earth rotates. The correlator computes τ_g from a model that also folds in station clock offsets, polar motion, solid-Earth tides, and tropospheric and ionospheric path lengths, then applies an integer-sample shift plus a fractional-sample delay interpolation to one stream.

Once delay-aligned, the residual interferometer phase still spins at the fringe rate, the time derivative of the delay scaled by the observing frequency. The correlator multiplies one station by a complex sinusoid at the negative fringe rate so that, over the accumulation period, the cross-product phase is nearly stationary and survives integration. The accumulated complex number, after a quantization (Van Vleck) correction, is the calibrated visibility that downstream imaging and fringe-fitting software consume.

The Cross-Correlation and Fringe Relations

Geometric delay (baseline b, source direction s):
τ_g = −(b · s) / c  (seconds)

Fringe rate (phase spin to be removed):
f_fringe = ν × (dτ_g / dt) ≈ ν × (ω_E b_⊥ / c)  (Hz)

Measured complex visibility on baseline ij:
V_ij(u,v) ≈ ⟨ v_i(t) × v_j^*(t − τ_g) e^{−j2πf_fringet} ⟩

Coherence integration limit:
Δt < 1 / (2π f_fringe)  so the phase turns < 1 rad per accumulation

Where ν = observing frequency, ω_E = Earth rotation rate (7.29 × 10⁻⁵ rad/s), b_⊥ = projected baseline, c = speed of light, and v_j^* denotes the delayed, conjugated station-j voltage. Example: at ν = 8.4 GHz with a 6,000 km baseline, f_fringe can exceed 10 kHz, demanding millisecond-class fringe rotation before accumulation.

Correlator Architectures and Quantization

Parameter	XF (lag) correlator	FX correlator	Notes
Order of operations	Cross-multiply lags, then FFT	FFT per station, then cross-multiply	FX dominates modern systems
Compute scaling	∝ N² × N_lag	∝ N log N + N² per channel	FX wins for large N stations
Typical platform	Custom ASIC / FPGA	CPU & GPU cluster (DiFX)	Software correlation now standard
1-bit sampling loss	≈ 1.96 dB (recovers 0.637 of correlation)		Van Vleck correction applied
2-bit sampling loss	≈ 0.55 dB (recovers 0.881)		Common modern compromise
Recorded data rate	0.5 to 32 Gbps per station		VLBA, EVN, EHT class systems

Common Questions

Frequently Asked Questions

What is the difference between an FX and an XF VLBI correlator?

An XF design cross-multiplies lagged time samples first and accumulates a lag spectrum, then Fourier transforms it into channels; its cost grows as N² × N_lag. An FX design Fourier transforms each station first, then cross-multiplies the per-channel spectra, scaling roughly as N log N plus N² per channel. FX wins decisively for large station counts, which is why software correlators like DiFX and the EHT and VLBA systems are FX machines on CPU or GPU clusters.

Why must a delay model and fringe rotation be applied before integrating?

The wavefront reaches the two antennas at different times, and Earth rotation makes that geometric delay τ_g change continuously. The correlator shifts one station by the integer-plus-fractional delay and counter-rotates its phase at the fringe rate ν(dτ_g/dt), which can exceed several kHz at centimeter wavelengths. Without this, the spinning phase averages to zero over even a 1-second integration and the fringe is lost. Models include clock offsets, Earth orientation, and atmospheric path terms.

How does coarse one-bit or two-bit quantization affect correlation sensitivity?

To hold down data volume, samples are quantized coarsely. One-bit (two-level) Nyquist sampling recovers only 0.637 of the correlation amplitude, a 1.96 dB hit, while two-bit (four-level) sampling with optimum thresholds recovers about 0.881, only 0.55 dB. The correlator applies the Van Vleck correction to convert the measured digital coefficient back to true correlation. Two-bit is the usual modern choice because the wider recordable bandwidth it enables outweighs its small extra loss.

Related Terms

← Correlation Radiometer (Space) Correlative Interferometer →