Millimeter-Wave Adaptive Optics

• Millimeter-wave adaptive optics is a system that senses dynamic wavefront errors via aperture-plane interferometry to correct mirror deformations from wind, thermal, and gravitational loads.
• It employs a coherent broadband noise source and strategic radiators to measure excess path lengths with an accuracy of around 8 µm to monitor structural changes.
• Anti-windup PI control techniques are used to suppress low-frequency disturbances, maintaining the precise surface accuracy required for futuristic large-aperture telescopes.

Millimeter-wave adaptive optics (MAO), also called millimetric adaptive optics, is the analogue, at mm/submm wavelengths, of optical adaptive optics: a system that senses time-variable wavefront errors in real time and corrects them with actuators on the optics. In the telescope literature, MAO is formulated around real-time sensing of wavefront deformation with $\sim 10~\mu\mathrm{m}$ accuracy across the aperture, typically by measuring spatially discrete excess path lengths (EPLs) from characteristic positions on the primary mirror surface to the focal plane and using those measurements to compensate deformation of telescope optics induced by wind and thermal loads (Tamura et al., 2021, Nakano et al., 2022). A later control-theoretic treatment casts EPL compensation as an asymptotic disturbance suppression problem and supplies a closed-loop design for low-frequency disturbances such as thermal and wind-induced deformations (Jikuya et al., 8 Jun 2026).

1. Physical regime and motivating problem

For large single-dish millimeter/submillimeter telescopes, the dominant, rapidly varying wavefront errors are the telescope structure itself rather than the atmosphere across the aperture. The MAO literature identifies gravitational deformation of a large, segmented primary reflector and its backup structure, thermal deformation from solar heating, diurnal temperature gradients, and cooling at night, wind loading on the large reflector and subreflector support structures, and structural flexure or low-order modes of the backup structure and feed support as the principal error sources (Tamura et al., 2021). In the same context, large-aperture telescopes are described as being “largely affected by the external environment, such as wind, preventing them from achieving the expected performance” (Nakano et al., 2022).

This focus distinguishes MAO from optical/NIR adaptive optics. In the mm/submm single-dish setting, tropospheric water-vapor fluctuations are described as having spatial scales $\gtrsim 100$ m at the ALMA site, larger than a single dish aperture, so the atmosphere “across the aperture” is often relatively coherent and deformation of the primary mirror and secondary support is the main wavefront error source (Tamura et al., 2021). The relevant temporal scales are correspondingly modest. The Nobeyama 45 m telescope is reported to exhibit a characteristic frequency due to wind load at $\sim 0.9$ Hz, with additional wind-excited modes at $1.6$, $2.6$, $3.1$, and $4.1$ Hz, and MAO targeting structural deformations is therefore described as needing a control bandwidth of a few Hz rather than the kHz bandwidth typical in optical AO (Nakano et al., 2022).

The engineering target follows from the surface accuracy demanded by future $50$ m-class facilities. For LST and AtLAST, the required mirror surface accuracy is stated as $\sim 20\text{--}40~\mu\mathrm{m~r.m.s.}$, while the sensing problem is framed as real-time sensing of wavefront deformation with $\sim 10~\mu\mathrm{m}$ accuracy across the aperture (Nakano et al., 2022). In the earlier sensor concept paper, the performance requirement is given as relative surface measurements at the level of $\gtrsim 100$0 with temporal resolution of $\gtrsim 100$1 ms (Tamura et al., 2021).

2. Aperture-plane interferometry as the canonical MAO sensor

Tamura et al. proposed a millimeter wavefront sensor based on aperture-plane interferometry, borrowing techniques from radio interferometry to measure excess path length from characteristic positions on the primary mirror surface to the focal plane (Tamura et al., 2021). The central architecture uses a coherent broadband noise source, split into a direct reference branch and a branch routed to radiators located at specific positions on the primary mirror surface. Radiation emitted from a radiator is reflected by the telescope optics to a coherent receiver at the focus, and a correlator measures the complex cross-power spectrum between the reference branch and the received branch (Tamura et al., 2021).

The fundamental observable is the phase of the cross-correlation. For a radiator position $\gtrsim 100$2, if $\gtrsim 100$3 is the effective radio path length from that position on the primary to the focus, then the phase at frequency $\gtrsim 100$4 satisfies

$\gtrsim 100$5

and for changes in path length,

$\gtrsim 100$6

When multiple radiators are used, the same source is switched among them, so each radiator yields one EPL time series $\gtrsim 100$7, and the set $\gtrsim 100$8 provides a sparse sampling of the primary’s surface deformation projected onto the line of sight (Tamura et al., 2021).

In the frequency-domain formulation used for the broadband system, the calibrated phase slope versus frequency gives EPL directly: $\gtrsim 100$9 This is explicitly presented as a radio analogue of measuring a group delay (Tamura et al., 2021). The measurement is not a Shack–Hartmann estimate of local slope and not a curvature-sensor estimate based on defocused intensity. It is a coherent radio-interferometric delay measurement driven by an internal broadband noise reference and internal radiators on the primary, so it does not rely on astronomical guide stars or atmospheric phase information (Tamura et al., 2021).

A crucial conceptual feature is that MAO measures optical path through the actual mm/submm beam train. The sensor is described as measuring radio path length directly and integrating contributions from all optical elements, which is why it is positioned in the literature as complementary to radio holography, photogrammetry, and celestial-source phase correction rather than a replacement for them (Nakano et al., 2022).

3. Instrument architecture and measured performance on the Nobeyama 45 m telescope

Nakano et al. reported a prototype wavefront sensor for millimetric adaptive optics installed on the Nobeyama 45 m radio telescope (Nakano et al., 2022). The sensor operated at $\sim 0.9$0 GHz and, in the two-element prototype, sampled two positions on the primary mirror surface at radii of $\sim 0.9$1 m and $\sim 0.9$2 m at a sampling rate of $\sim 0.9$3 Hz. The measured quantity was the excess path length between the two positions, obtained by differentiating the two optical paths (Nakano et al., 2022).

The instrumental chain used a broadband noise source spanning $\sim 0.9$4 GHz, a coherent receiver at the focus, and a correlator with integration time per dump of $\sim 0.9$5 s. Radiators were switched at $\sim 0.9$6 Hz, so the resulting $\sim 0.9$7 time series was sensitive to frequency content between $\sim 0.9$8 and $\sim 0.9$9 Hz; the paper states explicitly that the sensor is not sensitive to deformations of frequencies faster than $1.6$0 Hz in this measurement (Nakano et al., 2022). This temporal bandwidth was chosen to match the dominant structural oscillations of the telescope.

The measured EPL power spectral density showed three components: a low-frequency drift $1.6$1, oscillations, and a white noise (Nakano et al., 2022). Comparison under moderate wind $1.6$2 and strong/gust wind $1.6$3 showed that the low-frequency component and oscillation peaks increased with wind speed, indicating structural, wind-driven origins (Nakano et al., 2022). The same study reported that peaks and envelopes of band-pass filtered $1.6$4 and accelerometer-derived displacement aligned in time, with cross-correlation peaks near zero delay, which was taken as confirmation that EPL probes the same deformation modes as mechanical sensors (Nakano et al., 2022).

The sensitivity result most often cited is the white-noise-limited EPL precision. Using a white-noise power density of $1.6$5 over the $1.6$6 Hz band, the integrated white-noise power is $1.6$7, corresponding to

$1.6$8

The abstract rounds this to an $1.6$9 $2.6$0 statistical error in EPL measurements (Nakano et al., 2022). Because this is significant with respect to the $2.6$1 mirror surface accuracy required by LST and AtLAST, the paper concludes that the technique is useful for future large-aperture submillimeter telescopes (Nakano et al., 2022).

The earlier laboratory demonstration of the aperture-plane interferometric sensor gave a complementary accuracy result. In a $2.6$2 m test path, EPL versus mechanical displacement was reported as highly linear, with residuals showing a mean offset of $2.6$3 and standard deviation of $2.6$4, below the stated $2.6$5 requirement (Tamura et al., 2021). At the same time, both the concept paper and the Nobeyama characterization note that the prototype used only a small number of radiators and did not yet demonstrate full $2.6$6D surface reconstruction (Tamura et al., 2021, Nakano et al., 2022).

4. Control-theoretic formulation of EPL compensation

The control problem in MAO has been formalized as a plant in which actuator commands drive optical elements and are observed through EPL measurements (Jikuya et al., 8 Jun 2026). In the main example, each drive axis is modeled as a first-order position servo,

$2.6$7

and for $2.6$8 decoupled axes,

$2.6$9

The EPL measurement model is

$3.1$0

where $3.1$1 is the actuator-to-sensor coupling matrix, $3.1$2 is the desired EPL pattern from homologous deformation at elevation $3.1$3, and $3.1$4 is the disturbance term containing thermal drifts, wind, modeling error, and residual structural vibration (Jikuya et al., 8 Jun 2026).

Subtracting the known elevation-dependent component yields the residual

$3.1$5

The geometric structure is central. Any $3.1$6 is decomposed uniquely into a component in $3.1$7 and a component in $3.1$8, and with the Moore–Penrose pseudoinverse,

$3.1$9

The desired asymptotic behavior for constant disturbance $4.1$0 is

$4.1$1

which makes explicit that only the suppressible component of the residual can be cancelled (Jikuya et al., 8 Jun 2026).

This suppressible-versus-insuppressible distinction addresses a common misunderstanding. MAO does not, in this formulation, guarantee full rejection of arbitrary EPL patterns. Disturbances outside $4.1$2 remain in the residual, and the control objective is asymptotic suppression of constant disturbance components lying in the controllable subspace (Jikuya et al., 8 Jun 2026).

5. AWPI control, disturbance rejection, and controllable modes

The controller proposed for this setting is an Anti-Windup Proportional-Integral (AWPI) law (Jikuya et al., 8 Jun 2026). In the unsaturated case, the core PI structure acts on $4.1$3: $4.1$4 The paper shows that a pure P controller cannot achieve asymptotic suppression of constant disturbances, which is why the integral term is required (Jikuya et al., 8 Jun 2026). Because the plant and gain matrices are diagonal in the chosen coordinates, the MIMO design reduces to loop shaping of decoupled scalar sensitivity functions. For one scalar channel,

$4.1$5

Under the gain conditions given in the paper, the control bandwidth is approximately $4.1$6, and low-frequency suppressible disturbances are attenuated while high-frequency components pass substantially unchanged (Jikuya et al., 8 Jun 2026).

To handle actuator saturation, the anti-windup mechanism augments the integrator: $4.1$7

$4.1$8

The purpose is to maintain control continuity during recovery from saturation and prevent undesirable discontinuities in the drive command (Jikuya et al., 8 Jun 2026). In numerical simulations incorporating a three-axis secondary reflector drive and five-point EPL measurements, the framework demonstrated direction-dependent disturbance rejection and suppression of von Karman-modeled wind turbulence (Jikuya et al., 8 Jun 2026).

The same paper adds two operational tools. One is a manual focus adjustment scheme that allows observer intervention without interfering with the feedback loop. The other is the cosine similarity index, introduced to quantify the suppressibility of specific Zernike modes: $4.1$9 A value of $50$0 means the discretized mode lies in $50$1 and is fully suppressible; a value of $50$2 means it is orthogonal to $50$3 and fully insuppressible (Jikuya et al., 8 Jun 2026). This gives a direct link between sensor geometry, actuator geometry, and which optical aberrations can actually be corrected.

6. Relation to other metrology methods and alternate usage of the term

MAO is positioned in the astronomical literature as complementary to radio holography, photogrammetry, optical metrology, and celestial-source phase correction rather than interchangeable with them. Radio holography and photogrammetry are described as methods for setting the static surface to $50$4 rms in calm conditions, but they are slow and not usable continuously during science observations (Tamura et al., 2021). The aperture-plane interferometric MAO sensor instead uses an internal $50$5 GHz reference signal, is independent of sky conditions and celestial source availability, and measures radio path length directly through the relevant optics (Nakano et al., 2022).

Another frequent misconception is that MAO is primarily a substitute for static alignment. The sensor papers state the opposite: the measured EPLs are relative to an already holographically tuned “ideal” surface, and MAO is intended to maintain that shape in the presence of wind and thermal loads rather than perform the initial coarse alignment of the telescope (Tamura et al., 2021).

By analogy, the expression “millimeter-wave adaptive optics” has also been used to describe adaptive control, sensing, and steering of millimeter-wave beams in dual-polarized MIMO systems (Song et al., 2014). In that reinterpretation, channel sounding supplies the sensor output, beam alignment is the correction step, and adaptive sounding time is chosen from an approximate probability of beam misalignment. The system employs dual-polarized antennas, a practical soft-decision beam alignment algorithm that exploits orthogonal polarizations, and a high Ricean $50$6-factor, poor-scattering channel model in which the spatial channel lies in a $50$7D angular subspace (Song et al., 2014). This suggests that the phrase now spans two related but distinct ideas: conventional MAO for large submillimeter telescopes, and an adaptive-optics analogy for beam steering in mmWave communication systems.