(chap:05)=
# Spatial coherence and intensity interferometry
:::{admonition} Chapter opening
:class: chapter-opening
Chapter {ref}`chap:04` asked whether one beam of light is correlated with itself at different times. Now place two telescopes tens of meters, hundreds of meters, or even kilometers apart. Their mean intensities may look almost identical, but whether their intensity fluctuations rise and fall together depends on whether the source has already been resolved on the corresponding angular scale. This idea comes from the van Cittert--Zernike theorem, Hanbury Brown--Twiss intensity interferometry, and the Narrabri Stellar Intensity Interferometer. It has become practical again because atmospheric Cherenkov telescope arrays such as VERITAS, MAGIC, H.E.S.S., and CTAO already have large mirrors, fast detectors, and long baselines {cite:p}`1938Phy.....5..785Z,1939Phy.....6.1129V,1956Natur.177...27B,1957RSPSA.242..300B,1958RSPSA.248..222B,1967MNRAS.137..375H,1967MNRAS.137..393H,1974iiia.book.....H,2006ApJ...649..399L,2013APh....43..331D,2020NatAs...4.1164A,2024MNRAS.529.4387A`.
:::
## Spatial coherence and the van Cittert--Zernike theorem
Let two telescopes be located at $\mathbf{x}_1$ and $\mathbf{x}_2$. Their baseline projected onto the plane perpendicular to the line of sight is
```{math}
\mathbf{B}_\perp=\mathbf{x}_{2,\perp}-\mathbf{x}_{1,\perp}.
```
At wavelength $\lambda$, define the spatial frequencies
```{math}
:label: eq:ch05-uv
u=\frac{B_x}{\lambda},
\qquad
v=\frac{B_y}{\lambda}.
```
$B_x$, $B_y$, and $\lambda$ are all measured in meters, so $u$ and $v$ are dimensionless baseline lengths in wavelengths. If $B=100\,\mathrm{m}$ and $\lambda=500\,\mathrm{nm}$, then $B/\lambda=2\times10^8$. That number is the sampled point in the Fourier plane of the sky brightness. Its inverse gives the angular scale, $\lambda/B\simeq1\,\mathrm{mas}$.
The first-order spatial coherence between the two telescopes is
```{math}
:label: eq:ch05-gamma12
\gamma_{12}^{(1)}
=
\frac{\left\langle E_1^\ast(t)E_2(t)\right\rangle}
{\sqrt{\left\langle |E_1(t)|^2\right\rangle
\left\langle |E_2(t)|^2\right\rangle}} .
```
$\gamma_{12}^{(1)}$ is complex. Its modulus gives the coherence amplitude; its argument gives the Fourier phase. An amplitude interferometer combines the two optical fields with a delay line and measures fringe visibility directly. An intensity interferometer does not combine the light beams. It compares intensity fluctuations after detection, so the direct observable is $|\gamma_{12}^{(1)}|^2$.
For a spatially incoherent, quasi-monochromatic, far-field source in a small field of view, the van Cittert--Zernike theorem gives
```{math}
:label: eq:ch05-vcz
\gamma^{(1)}(u,v)
=
\frac{
\int I_\nu(l,m)\,
\exp[-2\pi i(ul+vm)]\,dl\,dm
}{
\int I_\nu(l,m)\,dl\,dm
} .
```
$I_\nu(l,m)$ is the specific intensity on the sky, commonly in $\mathrm{W\,m^{-2}\,Hz^{-1}\,sr^{-1}}$, or in an equivalent photon spectral brightness. The small angular coordinates $l$ and $m$ are measured relative to the phase center and can be treated as radians. The denominator normalizes the total flux, so $\gamma^{(1)}(0,0)=1$. The assumptions matter. The source must be mutually incoherent during propagation, the bandpass must not mix different $u,v$ points too strongly, the small-field approximation must make the $w$-term negligible, and the source brightness must not change substantially during the averaging time {cite:p}`1938Phy.....5..785Z,1939Phy.....6.1129V,2000stop.book.....G,2007itcp.book.....W`.
```{figure} ../_static/figures/generated/chapter_05/ch05_vcz_visibility_models.png
:name: fig:ch05-vcz-models
:width: 88.0%
Different sky-brightness models leave different structures in |V|2. A uniform disk has Bessel-like nulls, a Gaussian source declines smoothly, an equal-brightness binary produces cosine oscillations, and a thin ring oscillates more rapidly. Intensity interferometry measures the squared modulus of these curves, not the complex visibility itself.
```
## The scale of stellar angular diameters
The standard ruler is the uniform disk. If the stellar angular diameter is $\theta$, the visibility is
```{math}
:label: eq:ch05-uniform-disk
V(B)=\frac{2J_1(x)}{x},
\qquad
x=\frac{\pi\theta B}{\lambda},
```
where $J_1$ is the first-order Bessel function, $\theta$ is in radians, and $B$ is the projected baseline. The first null is at $x=3.8317$, so
```{math}
:label: eq:ch05-first-null
B_0\simeq \frac{1.22\,\lambda}{\theta}.
```
In the blue band used by VERITAS and MAGIC, take $\lambda=416\,\mathrm{nm}$. A star with $\theta=0.5\,\mathrm{mas}=2.42\times10^{-9}\,\mathrm{rad}$ has $B_0\simeq210\,\mathrm{m}$. A $1\,\mathrm{mas}$ star reaches the first null near $105\,\mathrm{m}$. A $0.2\,\mathrm{mas}$ target needs about $520\,\mathrm{m}$. This explains why Narrabri, with 10--188 m baselines, could measure bright sub-milliarcsecond stars, while kilometer-scale CTAO baselines can reach tens of microarcseconds {cite:p}`1967MNRAS.137..375H,1967MNRAS.137..393H,1974MNRAS.167..121H,1974iiia.book.....H,2006ApJ...649..399L,2014SPIE.9146E..0ZD`.
```{figure} ../_static/figures/generated/chapter_05/ch05_uniform_disk_resolution.png
:name: fig:ch05-uniform-disk-resolution
:width: 86.0%
First-null baseline of a uniform disk at λ = 416 nm. A 0.5 mas star reaches its first null at about 210 m, while 0.1 mas targets need kilometer-scale baselines. This is why Cherenkov arrays are attractive for blue optical intensity interferometry.
```
Real stars are not uniform disks. Hot stars have limb darkening, rapid rotators have equatorial bulges and gravity darkening, Be stars have disks, binaries have two brightness centers, and Cepheids change radius with phase. The uniform-disk diameter $\theta_{\rm UD}$ is therefore a convenient comparison parameter, not the final physical radius. Converting to a limb-darkened diameter $\theta_{\rm LD}$ requires an atmosphere model and the bandpass response. The first four-telescope VERITAS intensity-interferometry measurements of $\beta$ CMa and $\epsilon$ Ori gave $\theta_{\rm UD}=0.523\pm0.017\,\mathrm{mas}$ and $0.631\pm0.017\,\mathrm{mas}$, with corresponding limb-darkened diameters $0.542\pm0.018\,\mathrm{mas}$ and $0.660\pm0.018\,\mathrm{mas}$ {cite:p}`2020NatAs...4.1164A`. Later observations of $\beta$ UMa and the rapid rotator $\gamma$ Cas show what happens as $u,v$ coverage improves: intensity interferometry can constrain the direction-dependent shape of the photosphere, not just one radius {cite:p}`2024ApJ...966...28A,2025ApJ...995..191A`.
## What HBT intensity interferometry measures
The previous sections translated sky brightness and stellar angular diameter into visibility $V(B)$. Intensity-interferometry data analysis adds one more step. It does not measure the complex visibility directly. It estimates $|V(B)|^2$ from intensity fluctuations at two telescopes. Applying the Siegert relation from Chapter {ref}`chap:04`, Eq. {eq}`eq:ch04-siegert`, to spatial channels gives
```{math}
:label: eq:ch05-hbt-spatial
g_{12}^{(2)}(0)-1
=
\zeta\,\left|\gamma_{12}^{(1)}\right|^2 .
```
Here $g_{12}^{(2)}$ is the zero-delay correlation between the intensities at the two telescopes. The factor $\zeta$ contains finite timing response, spectral bandwidth, polarization, background, and instrumental normalization. In practice the model is often written as
```{math}
:label: eq:ch05-correlation-model
C_{12}(B)
\equiv g_{12}^{(2)}(0)-1
=
N_0\,f_1 f_2\,|V(B)|^2 ,
```
where $N_0$ is the zero-baseline correlation amplitude, $f_1$ and $f_2$ are the fractions of the total count rate contributed by the target in the two telescopes, and $V(B)$ is the visibility in Eq. {eq}`eq:ch05-vcz`. If the effective detector timing resolution is $\Delta t_{\rm eff}$, the optical bandwidth is $\Delta\nu$, and the thermal light has unresolved polarization, the scale of $N_0$ is roughly
```{math}
:label: eq:ch05-n0
N_0\sim \frac{1}{2\,\Delta\nu\,\Delta t_{\rm eff}},
\qquad
\Delta\nu\simeq\frac{c\,\Delta\lambda}{\lambda^2}.
```
For VERITAS, with $\lambda=416\,\mathrm{nm}$, $\Delta\lambda=13\,\mathrm{nm}$, and $\Delta t_{\rm eff}\sim4\,\mathrm{ns}$, this estimate gives an $N_0$ of a few $10^{-6}$. The fitted values were $(1.23\pm0.05)\times10^{-6}$ and $(1.26\pm0.06)\times10^{-6}$. Optical bandwidth, electronic response, and system throughput together push the correlation peak down to the part-per-million level {cite:p}`2020NatAs...4.1164A`.
```{figure} ../_static/figures/generated/chapter_05/ch05_zero_baseline_calibration.png
:name: fig:ch05-zero-baseline
:width: 92.0%
Zero-baseline calibration directly changes the inferred |V|2. The left panel sketches a night-to-night drift in N0 measured from calibrator stars. The right panel shows that an 8 percent error in N0 stretches all baselines in |V|2, biasing fitted angular diameters, limb darkening, or oblateness.
```
```{figure} ../_static/figures/generated/chapter_05/ch05_sii_signal_vs_baseline.png
:name: fig:ch05-sii-signal
:width: 86.0%
Intensity-interferometry signal estimated with C12 = N0|V|2, using N0 = 1.25 × 10−6, typical of modern blue optical systems. Larger angular diameters make |V|2 fall earlier with baseline and reach a null sooner. The same set of baselines is therefore more sensitive to a 0.8 mas star than to a 0.3 mas star.
```
Intensity interferometry gives up phase. With two telescopes it measures $|V|^2$, not the Fourier phase, so it cannot recover the same information as an amplitude interferometer. A two-telescope intensity measurement has a centrosymmetric ambiguity in the image. The gain is a much looser optical-path requirement. Amplitude interferometry needs optical paths stable to a fraction of a wavelength. Intensity interferometry needs the electronic signals aligned to a fraction of the detector response time. For a $1\,\mathrm{GHz}$ electronic bandwidth, path errors of several centimeters are needed before the correlation is strongly reduced. Atmospheric piston destroys optical phase, but it is not the main noise source for nanosecond intensity correlations. That is why Cherenkov telescopes, whose optical image quality is rough compared with purpose-built interferometers, can still be useful intensity interferometers {cite:p}`2006ApJ...649..399L,2008AIPC..984..205L,2013APh....43..331D,2014SPIE.9146E..0ZD`.
```{figure} ../_static/figures/generated/chapter_05/ch05_phase_ambiguity.png
:name: fig:ch05-phase-ambiguity
:width: 92.0%
Measuring only |V|2 loses Fourier phase. The source and its mirror image in the left panel have different brightness distributions. Their |V|2 curves in the right panel are identical. Intensity-interferometric imaging needs two-dimensional u, v coverage, physical priors, third-order correlations, or phase-retrieval algorithms to break such degeneracies.
```
## Signal-to-noise ratio, background, and systematic error
Hanbury Brown's usual signal-to-noise scaling for intensity interferometry is
```{math}
:label: eq:ch05-snr
\left(\frac{S}{N}\right)_{\rm rms}
=
A\,\alpha\,n\,|\gamma_{12}^{(1)}|^2
\left(\frac{\Delta f\,T}{2}\right)^{1/2}.
```
$A$ is the collecting area of each telescope in $\mathrm{m^2}$. $\alpha$ is the total photon-detection efficiency, including mirrors, filters, quantum efficiency, and readout losses. $n$ is the source photon spectral flux in the observing band, in $\mathrm{photons\,s^{-1}\,m^{-2}\,Hz^{-1}}$. $\Delta f$ is the electronic correlation bandwidth in Hz, $T$ is the integration time in seconds, and $|\gamma|^2=|V|^2$ is the squared visibility on the baseline. The expression does not explicitly contain the optical filter bandwidth, because narrowing the filter reduces photon rate and increases coherence time; in the limit where electronics are much slower than the optical coherence time, the two effects nearly cancel. It does contain electronic bandwidth, because faster detection dilutes the same optical correlation peak less.
Equation {eq}`eq:ch05-snr` is useful as a scale. Le Bohec and Holder estimated that two $100\,\mathrm{m^2}$ telescopes with $\alpha=0.3$, $\Delta f=1\,\mathrm{GHz}$, and $T=5\,\mathrm{hr}$ could detect a $V=6.7$ star at $5\sigma$ when $|\gamma|^2=0.5$; a $V=5$ star could yield angular-diameter precision of a few percent {cite:p}`2006ApJ...649..399L`. Narrabri used two $6.5\,\mathrm{m}$ light collectors, a 443 nm center wavelength, a 10 nm filter, and about 60 MHz electronic bandwidth. Its final sample was mostly limited to bright stars with $B<2.5$ {cite:p}`1967MNRAS.137..375H,1974MNRAS.167..121H,1974iiia.book.....H`. VERITAS used four 12 m telescopes, 250 MS/s sampling, and offline correlation to obtain several-hour stellar-diameter measurements during bright-moon time that was otherwise poor for gamma-ray observing {cite:p}`2020NatAs...4.1164A,2024ApJ...966...28A,2025ApJ...995..191A`.
```{figure} ../_static/figures/generated/chapter_05/ch05_snr_scaling.png
:name: fig:ch05-snr-scaling
:width: 86.0%
Relative scaling of intensity-interferometry signal-to-noise ratio. Each magnitude of dimming reduces the photon spectral flux by a factor of about 0.398. Integration time and electronic bandwidth improve SNR only as square roots. The curves omit collecting area, efficiency, background dilution, and |V|2, and show why bright-star selection and fast electronics set the practical reach.
```
Background first increases shot noise and then multiplies the amplitude in Eq. {eq}`eq:ch05-correlation-model` by $f_1f_2$. If the target contributes only 70 percent of the photons in each telescope, the correlation amplitude is already reduced to about 0.49. Recovering the same significance requires longer integration, and systematic errors become easier to see. Moonlight, night-sky background, a broad PSF, nearby stars, and detector dark current enter this way. Le Bohec and Holder estimated that the $\sim0.05^\circ$ optical PSF of a Cherenkov telescope makes night-sky light a sensitivity boundary for faint targets. Modern VERITAS analyses also correct night-sky background and dark current; without those corrections, angular diameters can be biased at roughly the 10 percent level {cite:p}`2006ApJ...649..399L,2020NatAs...4.1164A,2025ApJ...995..191A`.
Systematic correlations are more dangerous than white noise because the target signal itself is only $10^{-6}$--$10^{-3}$. Common checks are direct and unromantic: the correlation far from zero delay should vanish; the peak should disappear after channel swaps or artificial time shifts; different telescope pairs, nights, moon separations, and electronic chains should give consistent $N_0$; and $|V|^2$ for the same target should follow projected baseline, not PMT current, temperature, or electronic gain. MAGIC-SII uses standard camera pixels, narrowband filters, VCSEL analog fiber transmission, about 110 MHz aggregate bandwidth, and a real-time GPU correlator with no dead time. It has reported 22 stellar diameter measurements. In such systems the limiting factor is an engineering and calibration budget, not just mirror area {cite:p}`2024MNRAS.529.4387A,2022SPIE12183E..0DK,2019ICRC...36..740M,2019ICRC...36..714K`.
```{figure} ../_static/figures/generated/chapter_05/ch05_error_budget.png
:name: fig:ch05-error-budget
:width: 82.0%
Components of the angular-diameter error budget in intensity interferometry. Statistical error falls with photon number and integration time, but zero-baseline calibration, background dilution, stellar modeling, weather, and selection functions set a systematic floor. Modern arrays are useful only if these terms are carried into the likelihood together with photon noise.
```
## Multiple baselines, Earth rotation, and imaging
$N$ telescopes provide
```{math}
:label: eq:ch05-baseline-number
N_{\rm base}=\frac{N(N-1)}{2}
```
baselines. Intensity interferometry has a practical advantage here: electronic signals can be copied to many correlators to form all telescope pairs. In amplitude interferometry, splitting the same coherent light among many beam combiners costs photons. Earth rotation changes the projected baseline, so a single physical telescope pair sweeps a track in the $u,v$ plane during the night. Saving correlations every few minutes turns one spatial-frequency point into a track. Several nights, several hour angles, and many telescope pairs together produce two-dimensional $u,v$ coverage {cite:p}`2008AIPC..984..205L,2012MNRAS.419..172N,2012MNRAS.424.1006N,2013APh....43..331D`.
```{figure} ../_static/figures/generated/chapter_05/ch05_uv_coverage.png
:name: fig:ch05-uv-coverage
:width: 82.0%
A fixed telescope baseline projects to different u, v points as the Earth rotates. Each colored track represents one telescope pair; the negative conjugate points also constrain the Fourier transform of a real brightness distribution. Denser coverage separates disks, ellipses, binaries, and surface spots more cleanly.
```
Measuring only $|V|^2$ can still support low-dimensional imaging or parameter reconstruction. Uniform disks, elliptical disks, binaries, and limb-darkening models are already strongly constrained by $|V|^2$. Less model-dependent imaging needs either phase information or image constraints. Third-order intensity correlations contain ${\rm Re}(\gamma_{12}\gamma_{23}\gamma_{31})$, which provides a closure-phase like constraint. Another route is to recover phase from the two-dimensional $|V|$ distribution using Cauchy--Riemann, Gerchberg--Saxton, MIRA, or related post-processing algorithms. Simulations by Nunez et al. show that bright hot stars, rapid rotators, binaries, and spotted stellar surfaces can be recovered with physically useful shape parameters under CTA-like $u,v$ coverage, but the answer depends on SNR, short-baseline coverage, prior masks, and image regularization {cite:p}`1966JOSA...56..441G,2012MNRAS.419..172N,2012MNRAS.424.1006N,2014SPIE.9146E..0ZD,2024NJPh...26f3014B`.
The third-order expression in Eq. {eq}`eq:ch04-g3-closure`, applied to three telescopes $1,2,3$, contains
```{math}
2\,{\rm Re}(\gamma_{12}\gamma_{23}\gamma_{31}).
```
This term carries the phase sum around the closed triangle. It is therefore conceptually close to closure phase in amplitude interferometry, because it is not affected by the piston phase of any one telescope. The difficulty is statistical. The third-order signal is weaker than the second-order signal, and its error propagation is harsher. It belongs on large arrays and very bright targets before it becomes a default data product for intensity interferometry.
## From Narrabri to modern Cherenkov arrays
The Narrabri Stellar Intensity Interferometer was the first instrument to show the scientific power of the method in a sustained way. It used two $6.5\,\mathrm{m}$ light collectors on variable 10--188 m baselines, with a filter near 443 nm, 10 nm bandwidth, PMT quantum efficiency around 25 percent, and an effective electronic bandwidth near 60 MHz. During the 1960s and 1970s it measured angular diameters of 32 stars and several binary systems, down to diameters of about $0.4\,\mathrm{mas}$ {cite:p}`1967MNRAS.137..375H,1967MNRAS.137..393H,1970MNRAS.148..103H,1971MNRAS.151..161H,1974MNRAS.167..121H,1974MNRAS.167..475H,1974iiia.book.....H`. It was eventually shut down because the technology around it was limiting: large movable light collectors, low-noise fast electronics, and multibaseline imaging were all hard at the time.
Modern Cherenkov arrays changed the hardware boundary. They were built for nanosecond Cherenkov flashes, so they already have 10 m class apertures, large mirror areas, fast PMTs, long baselines, and tolerable optical image quality. VERITAS has demonstrated intensity interferometry on all six baselines of its four 12 m telescopes. Its $\beta$ CMa and $\epsilon$ Ori angular diameters were measured to better than 5 percent in only a few hours, using time close to full Moon that was not ideal for gamma-ray observations {cite:p}`2020NatAs...4.1164A`. MAGIC uses two 17 m parabolic telescopes and the existing camera pixels; after the SII upgrade it can switch quickly between normal gamma-ray observing and intensity-interferometry mode, and its GPU correlator has reported 22 stellar diameters {cite:p}`2024MNRAS.529.4387A`. Later VERITAS results on $\beta$ UMa and $\gamma$ Cas moved the field from feasibility demonstration to specific stellar physics. For $\gamma$ Cas, six baselines and more than 160 pair-hours of data measured the flattening of the visible photosphere, giving a minor-axis diameter of about $0.43\,\mathrm{mas}$, an axial ratio of about 1.28, and a rotation-axis position angle near $116^\circ$ {cite:p}`2024ApJ...966...28A,2025ApJ...995..191A`.
CTAO is attractive because of both baseline length and baseline count. Kilometer baselines at 350--450 nm correspond to tens of microarcseconds, and dozens to more than a hundred telescopes can provide thousands of baselines with dense $u,v$ coverage. It will not replace CHARA, VLTI, or other amplitude interferometers, because the sensitivity, phase information, and infrared capability are different. It is a complementary blue-optical, very-long-baseline tool for bright hot stars and rapidly changing targets. Its scientific output will depend on detector bandwidth, time synchronization, correlator throughput, background control, reproducible zero-baseline calibration, and data products that can connect to optical-interferometry standards {cite:p}`2009A&A...507.1719F,2010SPIE.7734E..1CN,2010SPIE.7734E..1TJ,2013APh....43..331D,2014SPIE.9146E..0ZD,2022icrc.confE.803K,2024SPIE13095E..0IT`.
The chain is compact. Sky brightness enters the complex visibility $V(u,v)$ through the van Cittert--Zernike theorem. An HBT correlator does not return that complex number. It measures $g_{12}^{(2)}-1=N_0f_1f_2|V|^2$. Zero-baseline calibration and background modeling decide whether the instrumental amplitude can be converted into astrophysical $|V|^2$. Multiple baselines and Earth rotation then turn a set of $|V|^2$ points into stellar diameters, ellipticities, binary parameters, or image constraints. A later chapter on detectors and clocks returns to the hardware question: how $10^{-6}$-level correlation peaks are actually recorded.