Astronomical Seeing: Why Stars Twinkle and Telescopes Blur

Temperature Eddies Are Weak, Invisible Lenses

Starlight crosses light-years of near-vacuum and reaches the top of the atmosphere as an almost perfect plane wave, flat to a minute fraction of a wavelength across any aperture we could build. Everything goes wrong in the final ~20 km.

Air is a weak refractor: refractivity n − 1 ≈ 2.7 × 10⁻⁴ in the visible at sea level, depending on density and hence temperature, with ∂n/∂T ≈ −1 × 10⁻⁶ per kelvin. Turbulence does not change a layer's mean temperature; it mixes parcels differing by a fraction of a kelvin, each acting as a lens of contrast δn ~ 10⁻⁷–10⁻⁶. Individually negligible — but integrated over kilometres the optical-path errors reach a few micrometres, several wavelengths of visible light, which is precisely the amount that destroys an image.

The statistics come from Kolmogorov's 1941 turbulent cascade, transplanted to optics by Valerian Tatarski in Wave Propagation in a Turbulent Medium (1961). Energy enters at an outer scale L₀ of tens of metres (median ~22 m at Paranal) and dissipates at an inner scale of a few millimetres. Between them the refractive-index structure function is a clean power law, D_n(r) = C_n² r^2/3. That structure constant C_n² runs from ~10⁻¹⁷ m^−2/3 in the quiet free atmosphere to ~10⁻¹³ m^−2/3 in the churned air near the ground. The turbulence is layered, not uniform: a surface layer below ~1 km (often half the total), a jet-stream layer at 10–12 km, and — embarrassingly — turbulence generated inside the dome and over the warm primary mirror itself.

The Fried Parameter: One Number That Fixes Everything

In 1965–66 David L. Fried showed that the whole vertical profile collapses into a single length, r₀ = [0.423 k² sec ζ ∫ C_n²(h) dh]^−3/5, with k = 2π/λ and ζ the zenith angle. Physically, r₀ is the aperture over which the wavefront is still flat to about one radian: the phase variance across a circular pupil is σ² = 1.03 (D/r₀)^5/3 rad², which equals ~1 exactly when D = r₀. Below r₀ you have a telescope; above it, a mosaic. Everything else follows.

• Seeing angle. Long-exposure FWHM ε ≈ 0.98 λ/r₀. At 500 nm, r₀ = 10 cm gives 1.01″ and r₀ = 20 cm gives 0.51″ — the entire range quoted for the world's best sites.
• Wavelength. r₀ ∝ λ^6/5, so ε ∝ λ^−1/5. From 500 nm to 2.2 µm the blur shrinks only 25%, but r₀ grows 5.9× and the patch count (D/r₀)² falls ~35×. That, not the blur size, is why infrared adaptive optics was conquered first.
• Airmass. r₀ ∝ (cos ζ)^3/5, so at 60° from the zenith the seeing disc is 2^0.6 = 1.52× larger. Observers really do wait for targets to transit.
• Time. The pattern is frozen and blown across the pupil (Taylor's hypothesis), giving τ₀ = 0.314 r₀/v̄ with v̄ the C_n²-weighted wind of 10–20 m s⁻¹: a few milliseconds. The Greenwood frequency f_G = 0.426 v̄/r₀ ≈ 30–100 Hz sets how fast a correction loop must run.
• Angle. Since layers sit at different heights, neighbouring stars look through partly different air. The isoplanatic angle θ₀ = 0.314 r₀/h̄ is only ~2″ at 500 nm for h̄ ≈ 5 km — the harshest constraint in adaptive optics.

Speckles: Aperture Buys Photons, Not Resolution

When D < r₀ — the naked eye, binoculars, a small refractor — the aperture sees one coherent patch that is merely tilted. The image stays a sharp diffraction pattern but dances: pure image motion, which is why small telescopes show a jittering point rather than a smear.

When D ≫ r₀ the pupil tiles into roughly (D/r₀)² independent patches. Each forms its own diffraction blob of size λ/D, and because they arrive with random relative phases they interfere into a speckle pattern: thousands of grains scattered inside a halo about λ/r₀ wide. On a VLT unit telescope at 500 nm with r₀ = 15 cm that is ~3,000 speckles of 13 milliarcseconds spread across a 0.67″ disc. The pattern reshuffles every τ₀; expose for a second and hundreds of realisations average into a smooth, broad-winged profile (well fitted by a Moffat function) whose width is the seeing.

The consequence is stark: an uncorrected 8-metre telescope has the angular resolution of a 15-centimetre one. The extra aperture buys collecting area — ~3,000 more photons — plus, once the wavefront is corrected, the diffraction limit you paid for. Newton diagnosed this in the Opticks without any of the mathematics: telescopes “cannot be so formed as to take away that confusion of the Rays which arises from the Tremors of the Atmosphere. The only Remedy is a most serene and quiet Air, such as may perhaps be found on the tops of the highest Mountains above the grosser Clouds.” Three centuries later, that is exactly where the observatories are.

Twinkling Is a Different Effect From Blurring

Naked-eye scintillation is an intensity fluctuation, not a phase one, and it comes from a different part of the atmosphere. Turbulence initially only wrinkles the phase; over a propagation distance h those wrinkles Fresnel-diffract into bright and dark patches, as the eddies focus and defocus the beam and throw shadows on the ground. Their natural size is the Fresnel scale r_F = √(λh) — about 7 cm for 500 nm light from 10 km up. The ground is carpeted with 7-centimetre flying shadows sweeping past at 10–30 m s⁻¹, so the intensity flickers at v/r_F ~ 10²–10³ Hz; the eye, integrating over tens of milliseconds, registers only the slower envelope of that flicker. In the weak-fluctuation regime σ_I² = 19.2 λ^−7/6 sec^11/6ζ ∫ C_n²(h) h^5/6 dh; note the h^5/6 weighting, so high layers dominate twinkling while the ground layer that dominates seeing barely contributes. Good seeing and steady stars are not the same condition.

• Aperture averaging. Your ~6 mm pupil samples one shadow cell at a time and sees full modulation. A 20 cm telescope spans only ~(D/r_F)² ≈ 8 cells, but since σ_I² falls roughly as (D/r_F)^−7/3 that already cuts the fluctuation several-fold, so the twinkling largely vanishes — while the blur remains untouched. Stars twinkle to the eye but only shimmer in an eyepiece.
• Source averaging. A source of angular size θ smears the pattern from height h over a length hθ, so averaging wins once θ > √(λ/h) ≈ 1.5″. Stars fail badly: even Betelgeuse, among the largest stellar discs, is ~44 mas, and most stars are under 1 mas. Planets pass easily — Jupiter 30–50″, Venus 10–66″, Mars 3.5–25″ — so hundreds of shadow patterns superpose and the light is steady. Uranus at ~3.5″ sits on the boundary and does flicker faintly.

Twinkling also explains astronomy's most-reported false UFO: dispersion spreads a low star into a tiny vertical spectrum (~1″ between 400 and 700 nm at 45° zenith angle, several arcseconds near the horizon) and each colour scintillates independently, so Sirius near the horizon flashes red, green and blue. Scintillation also imposes a photometric noise floor, σ ≈ 0.09 D^−2/3 X^1.75 e^−h/8000 (2t)^−1/2 (Young's relation, with D in cm, observatory altitude h in m and exposure t in s; recalibrated site by site by Osborn et al. 2015) — a large part of the case for Kepler, TESS, CHEOPS and PLATO.

How Seeing Is Actually Measured

The workhorse is the DIMM — Differential Image Motion Monitor, standardised by Marc Sarazin and François Roddier in 1990. A small telescope is masked to two ~10 cm subapertures ~20 cm apart, one behind a wedge so the star forms two images. The variance of their separation depends only on the atmosphere, cancelling tracking errors, wind shake and vibration; inverting it yields r₀. Published medians at 500 nm: Paranal ≈ 0.69″, Mauna Kea ≈ 0.6″, La Palma ≈ 0.76″. The extreme case is Dome C, Antarctica, where Lawrence et al. (Nature, 2004) measured free-atmosphere seeing of 0.27″ — above a ~30 m surface inversion that is itself dreadful.

Profilers unpack the vertical structure. MASS exploits the h^5/6 scintillation weighting through four concentric annuli to retrieve C_n² in ~6 layers. SLODAR triangulates a profile by cross-correlating Shack–Hartmann slopes from the two components of a binary. Stereo-SCIDAR recovers both layer altitudes and their wind velocities — the inputs a predictive adaptive-optics controller needs. Balloon-borne microthermal thermosondes measure C_T² in situ as ground truth.

One subtlety matters operationally: DIMM seeing is not delivered image quality. The monitor samples free air a few metres up; the science image also carries dome seeing, mirror seeing (an aluminium primary just 1 K warmer than the air adds a few tenths of an arcsecond of convective blur; empirical scalings run ≈0.4 ΔT^1.2 arcsec), wind buffeting and residual aberrations. Hence ventilated enclosures and actively cooled primaries — an advance as important as moving to high sites. The Vera C. Rubin Observatory takes the opposite route to adaptive optics: its 8.4 m survey telescope is deliberately seeing-limited, with an expected median delivered FWHM near 0.8″, because over a 3.5° field anisoplanatism makes wavefront correction hopeless.

Beating It: Lucky Imaging and Adaptive Optics

Speckle interferometry. Antoine Labeyrie showed in 1970 that the ensemble-averaged power spectrum of many millisecond exposures retains information out to λ/D even though long exposures wash it out; bispectrum “speckle masking” recovers the phases too. Many close binaries and the first resolved stellar diameters were measured this way.

Lucky imaging. Fried calculated in 1978 that a short exposure is near-diffraction-limited with probability P ≈ 5.6 exp[−0.1557 (D/r₀)²] for D/r₀ ≳ 3.5 — 0.3% of frames at D/r₀ = 7. Cambridge's LuckyCam (Law, Mackay & Baldwin, 2006) on the 2.56 m Nordic Optical Telescope routinely delivered ~0.1″; combined with adaptive optics on the Palomar 5 m it reached 35 mas in i-band, sharper than Hubble at that wavelength. The exponential in (D/r₀)² is also why the technique never scales to 8-metre apertures.

Adaptive optics. Horace Babcock proposed sensing the wavefront and cancelling it with a deformable element in PASP in 1953; Vladimir Linnik independently in 1957. The idea matured inside classified US defence programmes, declassified in 1991. The first astronomical system was COME-ON (ESO/ONERA/Observatoire de Paris) on the 1.52 m at Haute-Provence in 1989, then the ESO 3.6 m at La Silla. A wavefront sensor — Shack–Hartmann, curvature, or increasingly pyramid — samples the pupil in r₀-sized subapertures at 500–2,000 Hz, well above the Greenwood frequency; a reconstructor turns slopes into commands; a deformable mirror applies the conjugate shape. Performance is scored by the Strehl ratio, tied to residual phase variance by Maréchal's S ≈ exp(−σ²). Actuator counts scale as (D/r₀)²: the ELT's M4 is a 2.4 m adaptive mirror with 5,316 actuators, about what 39 m demands at 1–2 µm.

The hard part is the reference. Sensing needs roughly V ≲ 13–14 within θ₀, so natural-guide-star sky coverage in the visible is under 1%. Foy and Labeyrie's 1985 answer was the laser guide star: a 10–25 W laser on the sodium D₂ line at 589 nm excites the mesospheric sodium layer at 90–100 km (column ~4 × 10¹³ atoms m⁻²) to make an artificial star anywhere. Two limits are irreducible — the cone effect (light from 90 km samples a cone, not the cylinder a real star illuminates) and tip-tilt indeterminacy (the beam traverses the same air up and down, so its return carries no absolute position), which still forces a faint natural star for tilt. The field has branched into GLAO (ground layer only, wide field — the VLT's GRAAL feeding HAWK-I, and GALACSI feeding MUSE), MCAO (mirrors conjugated to several altitudes — ESO's MAD in 2007, Gemini's GeMS from 2011), tomographic LTAO, and extreme AO — VLT/SPHERE, Gemini/GPI, Subaru/SCExAO — which reaches ~90% Strehl in H band on bright stars in good conditions and images planets directly, as in HR 8799 and β Pictoris b. Adaptive optics tracking stars orbiting Sgr A* underpinned the half of the 2020 Nobel Prize shared by Andrea Ghez, who used Keck, and Reinhard Genzel, who used ESO's NTT and VLT.

What Seeing Is Not, and What Remains Unsolved

Several atmospheric effects are routinely mislabelled as seeing. Transparency is independent: the clearest, driest nights often have the worst seeing because a fast jet stream sits overhead, while humid, stagnant, laminar nights can be superb. Refraction and dispersion displace rather than scramble — mean refraction lifts an object ~34′ at the horizon and the differential across a bandpass stretches the image into a short spectrum, cured by an atmospheric dispersion corrector of counter-rotating prisms, not by adaptive optics. Extinction with airmass dims a source without affecting sharpness. And active optics is not adaptive optics: active optics, pioneered on ESO's New Technology Telescope in 1989, corrects gravitational and thermal flexure of the primary at ~0.05 Hz, while adaptive optics corrects the air at kilohertz rates.

Real open problems remain. Servo lag is now often the dominant error term, and machine-learned predictive control is being demonstrated at Keck, SCExAO and MagAO-X to anticipate frozen-flow motion rather than chase it. The low-wind (island or petal) effect at the VLT and Subaru, in which cold spider vanes create piston steps between pupil fragments that a Shack–Hartmann cannot sense, still spoils the sharpest images on the calmest nights. The sodium layer's altitude and density vary on minute timescales, injecting spurious focus into laser systems. Non-common-path aberrations set the contrast floor for exoplanet imaging. And nobody yet knows how close 30–40 m telescopes will come to their few-milliarcsecond diffraction limits, still less whether visible-light adaptive optics is tractable there, where (D/r₀)² approaches 10⁵ correction zones. After three centuries, Newton's “tremors of the atmosphere” are managed, quantified and partly cancelled — but not defeated.

What atmospheric turbulence costs telescopes of different sizes (λ = 500 nm, r₀ = 15 cm, seeing 0.67″)

Aperture	Diameter D	Diffraction limit λ/D	Speckles ≈ (D/r₀)²
Dark-adapted human eye	7 mm	15″	1 — smaller than r₀
Amateur reflector	0.2 m	0.52″	~2
Hubble Space Telescope	2.4 m	0.043″	none — above the atmosphere
VLT unit telescope	8.2 m	0.013″ (13 mas)	~3,000
Keck I / II	10 m	0.010″ (10 mas)	~4,400
ESO Extremely Large Telescope	39 m	0.0026″ (2.6 mas)	~68,000