(2) |
The mirror actuators are adjusted each in turn, with repeated iterations performed faster than the seeing lifetime, to maximise the sharpness function. The convergence however tends to be trapped in secondary maxima. To overcome this problem, Ribak ([Ribak 1990]) used simulated annealing: noise is added to the actuator signals and gradually reduced until the system falls in a state of minimum energy, reached when the sharpness function is maximised.
Hamaker, O'Sullivan & Noordam [(1977)] have later remarked that, according to Parseval's theorem, maximising the sharpness function is equivalent to maximising the integrated square modulus in the image's Fourier transform, which is the pupil's auto-correlation function multiplied by the object's Fourier transform. The maximum occurs for uniform pupil phases. However, if the aperture is diluted, the intensity of the autocorrelation peaks remains invariant if they do not overlap. This indicates that the sharpness approach cannot work if the exit pupil aperture is highly diluted, as would be the case for long-baseline Fizeau interferometers. The densified exit pupils considered here can be made to meet the peak overlap requirement.
We have tried to improve the convergence and to avoid the simulated annealing, by using a hierarchy of aperture triplets. The 27-aperture interferometer is initially divided in 9 separate interferometers, each combining images from 3 adjacent apertures, then re-arranged as 3 separate interferometers having 9 apertures each and finally re-arranged as a single interferometer having 27 apertures. If observations are performed from Earth the steps of re-arrangements are cycled at high speed, within the life-time of seeing, by tilting elements of the active mirror support to ``explode'' the image into separated triplet components. More delicate tip-tilt and piston adjustments of the same elements provide the phase adjustments, as described in Fig. 2.
The monochromatic interference pattern from 3 sub-apertures is generally a honeycomb-like interference function, multiplied by the diffraction function. Adding 3 such patterns from adjacent groups creates a finer honeycomb pattern within the coarse one. The triplet grouping can be pursued at higher levels of the hierarchy to include 27, 81, 243, etc. apertures. The closest apertures are grouped first to benefit from the fact that the object is less resolved.
For a non-resolved star, phasing is directly achievable by centering a cell of the honeycomb pattern appearing at each level of triplet formation. This requires adjusting 2 phases, those of 2 among the 3 groups of apertures at every step of the routine (Fig. 3). In monochromatic light and for 3 groups of apertures, the honeycomb pattern can be centered by extracting the phase values in the 6 peaks of its Fourier transform. In polychromatic light, 3-dimensional Fourier transforms of the data cubes provided by the spectro-imaging camera can provide similar information with full removal of the modulo phase ambiguity. This algorithm has not been tested yet, but a 2-aperture analog is routinely used at the ``Grand Interféromètre à 2 Télescopes'' (GI2T) ([Morel & Koechlin 1998]) for fringe acquisition.
More simply, the central white cell appearing in a honeycomb pattern amidst the coloured lateral cells can also be identified, even with a broad-band monochrome camera, from its higher contrast. Centering all the white cells from the various triplet groups onto a common fiducial mark corrects directly the optical path errors. On resolved objects, such honeycomb centering should not generally succeed since de-centered honeycomb components can be present. Indeed, imaging an object through a diluted triplet aperture causes a convolution with the honeycomb-like spread function. If the object is itself a honeycomb-like grid pattern matching this function, then the convolution preserves the image contrast, a situation which may be qualified as ``honeycomb resonance''. With 3 different triplets now forming a single aperture, 3 different grid components of the object add their images, but they do not generally have a common centering. Hence the need for another criterion, sharpness for example, to adjust the tip-tilt of 2 pupil triplets before adjusting their global piston. Six parameters should therefore be adjusted at each level in the hierarchical procedure.
When calculating the sharpness integral, the diffraction envelope of the sub-images causes unwanted biases, which tend to center the brighter zones. The bias can be avoided by ``flattening'' the diffraction peak: each recorded sub-image must be first divided by the known diffraction function, within the peak area.
Somewhat surprisingly, our simulations (Fig. 4) on resolved objects worked well without these precautions, using a simplified version of the routine: only piston adjustments were made at each level, and no flattening of the diffraction envelope was carried out. It is however likely that some object shapes require the complete routine. Also, no trapping in secondary maxima of the sharpness function was observed, but the reasons are unclear. Further investigations are desirable in these respects.
A single camera can receive all partial images and the fully combined image, using either temporal or spatial separations as shown in Fig. 5. Rather than using the actuators for exploding the image into the various groups at a fast cycling rate, beam-splitters can provide a parallel display of these triplet, ninetuplet, etc. groups ([Labeyrie 1998]). Mapping the information is efficiently achievable with a spectro-imaging attachment such as the combination of a grism and a micro-lens array developed by Bacon, and co-workers ([1995]) at the Observatoire de Marseille.