MIRA: an effective imaging algorithm for optical

interferometry
Eric Thiebauta
a Centre de Recherche Astrophysique de Lyon, CNRS/UMR 5574, France.

ABSTRACT
This paper presents Mira, a Multi-aperture Image Reconstruction Algorithm, which has been specifically de-
veloped for image restoration from optical interferometric data. The sought image satisfies agreement with the
input interferometric data and with some a priori image properties (positivity, normalization and regularization).
The algorithm can cope with very limited amount of data; as an extreme case, Mira is able to restore images
without any Fourier phase information. This leads to the possibility to perform imaging with only 2 telescopes
or when the phase closures are corrupted.
Keywords: optical interferometry; image reconstruction; inverse problem.

1. INTRODUCTION
Optical interferometers (VLTI, IOTA, etc.) yield the best angular resolution in the visible and infrared. The
measures provided by these instruments are however not directly images and reconstruction algorithms are
mandatory tools to fully exploit their high angular resolution imaging capabilities. Image reconstruction in
radio-interferometry has a long history and is now well under control.1, 2 But, owing to the specifics of optical
interferometry, the methods developed for radio-interferometry are not directly usable and we need to develop
new ones. The reasons of additional difficulties in optical interferometry are a much sparser sampling of the
spatial frequencies, the so-called u-v plane, and the loss of most of the Fourier phase information.
An interferometer samples the u-v plane at discrete spatial frequencies given by:

j,k (t) = bj,k (t)/ (1)

where is the wavelength and bj,k (t) is the separation, at time t and projected onto the sky, between the j-th and
k-th interfering telescopes. Ideally, an interferometer measures the complex visibility x() which is the Fourier
transform of the normalized distribution of intensity x(a) of the observed object in angular direction a. Hence,
the observed complex visibilities are:

xj,k (t) = x bj,k (t)/ . (2)
For N telescopes (or antennae) combined simultaneously, the maximum number of different spatial frequencies
simultaneously measured by an interferometer is N (N 1)/2. With usually at least a dozen radio antennae
against only 3 or 4 optical telescopes, the radio interferometers have an overwhelming advantage. To rebuild an
image one must properly interpolate the missing spatial frequencies which are holes in the coverage of the u-v
plane; this is the purpose of a priori constraints such as positivity, regularization and, possibly, support imposed
on the sought image.
In addition to the sparsity of interferometric data, atmospheric turbulence is responsible for a random phase
shift between the telescopes. At optical wavelengths and in the absence of a phase reference source, the resulting
phase piston errors are difficult (or impossible) to measure and, a fortiori, to compensate for. To overcome
random delays due to turbulence, astronomers must integrate interferometric measurements that are insensitive
to this defect. The powerspectrum and the bispectrum are such estimators. The sampled powerspectrum is:
(2) 
2
xj,k (t) = x bj,k (t)/  , (3)

Further author information: [email protected]

which only needs two telescopes (here j and k), and provides no information about the phase of the complex
visibility. The bispectrum is the triple product of the complex visibilities measured by the interferences from
three telescopes:
(3)



xj,k,` (t) = x bj,k (t)/ x bk,` (t)/ x b`,j (t)/ , (4)
where j, k, and ` are the indices of the involved telescopes. The phase of the bispectrum is the so-called phase
closure:
def  (3)
j,k,` (t) = arg xj,k,` (t) = j,k (t) + k,` (t) + `,j (t) mod 2 , (5)
where is the phase of the complex visibility:
def 
j,k (t) = arg xj,k (t) . (6)

Hence the bispectrum provides some Fourier phase information from the three spatial frequencies j,k (t), k,` (t),
and `,j (t) = j,k (t) k,` (t) which form a closed triangle in the u-v plane. However, in the case of a 3-telescope
interferometer, the phase closure data provide only a single phase out of three spatial frequencies. Moreover this
is at the price of increased complexity for the data processing (non-linearity) and for the operation of the
instrument since it requires to make at least three telescopes to interfere.
The notation used here is intended to explicit the dependence of the data with the interfering telescopes and
with the time (i.e. orientation with respect to observed object due to earth rotation). In the remaining of the
paper, for sake of simplicity, the list L of sampled spatial frequencies may be indexed by a single index:
def
L = { k ; k = 1, . . . , m} = {bk,` (t)/; (k, `, t)} , (7)

where (k, `, t) formally means: for all baselines and exposures used during the observations. To simplify the
equations to come, we also introduce the following notations for the complex visibilities and squared visibilities:
def
vj,k (t) = xj,k (t) , (8)
def (2)
sj,k (t) = xj,k (t) . (9)

2. IMAGE AND COMPLEX VISIBILITY MODELS

The result of the image reconstruction is the distribution of intensity x(a) across the field of view in the
angular direction a. A practical mean to parametrize this distribution is to use a basis of functions {zj :
7
R; j = 1, . . . , n} and to approximate the brightness distribution by a linear expansion:
n n
FT
X X
x(a) ' xmodel (a) = xj zj (a) x(a) ' xmodel () = xj zj () (10)
j=1 j=1

where x Rn are the parameters of the model image, xmodel () and zj () are the Fourier transforms of the model
image and of the j-th basis function at spatial frequency . This description is very general, for example it can
be used to have a multi-resolution model of the image or to account for point sources over a diffuse background.3
Another common description is to use the same function z(a) on an evenly spaced grid of angular directions
Ga = {aj ; j = 1, . . . , n}, then:
n n
FT
X X
xmodel (a) = xj z(a aj ) xmodel () = z() xj e2 i aj . (11)
j=1 j=1

In this description, the basis function z(a) set the effective angular resolution of the image in the manner of
the clean beam in the Clean method.4 To avoid spectral aliasing, it is necessary that Shannon criterion be
respected and that the angular sampling step a obeys:

a , (12)
2 bmax
Figure 1. Rebinning (top) and interpolation (bottom) of spatial frequencies. Top: the data at sampled frequencies are
rebinned by operator G to match the grid of frequels. Bottom: model of complex visibilities is interpolated at sampled
spatial frequencies by operator R.

where bmax is the maximum length of the projected interferometric bases. Some approaches, as the building blocks
method5 or Wipe,6 use explicit expressions for z(a). Finally choosing any interpolation function as the basis
function z(a) makes the model in Eq. (11) the same as most other algorithms for which the parameters are
the intensities of the pixels of the image: xj = x(aj ).
The image models considered in Mira have exact Fourier transforms. Using matrix notation, the model of
the complex visibility is:
v model = A x (13)
where x Rn are the sought image parameters and v model Cm assuming there are m sampled spatial frequen-
cies. In words, vkmodel ' x( k ) is the model of the complex visibility for k-th sampled spatial frequency. The
complex coefficients of matrix A depend on the image model:

Ak,j = zj ( k ) (14)
= z( k ) e2 i aj k (15)

where expression in Eq. (14) results from the model in Eq. (10) while the model in Eq. (11) yields Eq. (15).
The exact linear transform can be used in the image restoration algorithm but may become computationally
untractable when the size of A, that is m n, becomes too large. The Fourier transform of Eq. (11) may be
approximated by a discrete Fourier transform which, under an additional circulant approximation, is efficiently
computed by means of a fast Fourier transform (FFT). In this case, as discrete spatial frequencies the so-called
frequels do not necessarily coincide with those measured, it is necessary to interpolate the Fourier spectrum
of the image.2
In radio interferometry, the so-called gridding technique7 consists in interpolating the complex visibility data
onto the rectangular grid of frequels G = { k ; k = 1, . . . , n} that correspond to spatial frequencies after fast
Fourier transform of the pixel image representation. After gridding, the re-sampled data write:

v data = G v data (16)

where v data Cm are the original complex visibility data. The linear operator G accounts for the combination
of different operations:7 local averaging which depends on the actual density of the u-v coverage, weighting to
account for the reliability of the data, and tapering to set the resolution. The model of the re-sampled complex
visibility data is then:
v model = T x = T F x (17)
where T is a truncation (or sampling) operator with all coefficients equals to 0 or 1 and which discards the
discrete frequencies outside the u-v coverage (the grayed area in the central part of top of Fig. 1) and x is the
FFT of the pixel map x:

x = F x with Fk,j = exp 2 i aj k . (18)
With this definition of the FFT operator F, the inverse transform writes:
 1  1

F j,k = exp +2 i aj k (19)
n
where n = n1 n2 is the total number of pixels in the image, n1 and n2 being the number of pixels along the
two axis (usually n1 = n2 ). Note that:
F1 TT v data
is the so-called dirty image equals to the principal solution8 of solving v model =v data without additional constraints.
In optical interferometry, complex visibilities are usually not available and most data consist in non-linear
quantities (powerspectrum, bispectrum or phase closure) which cannot be linearly remapped onto the frequel
grid. The gridding of the data is not possible, but the discrete Fourier transform of the image can be interpolated
at the frequencies sampled by the data. To benefit from FFT speedup, Mira makes use of an approximation of
A:
A'RF (20)
where F is the FFT matrix and R is a linear interpolation matrix which performs interpolation of the model
FFT x at measured spatial frequencies. This interpolation can be chosen so that R is a very sparse matrix which
is fast to apply and has light memory footprint. In Mira, we use complex bilinear interpolation, thus the model
of the complex visibility at each measured spatial frequency is a linear combination of 4 neighbors in x. The
number of floating point operations to apply R scales as 16 m.

3. INVERSE PROBLEM APPROACH

Once chosen the parametrization, image reconstruction can be seen as an inverse problem.9, 10 Adopting a
Bayesian viewpoint, the best parameters x+ can be chosen as the most likely ones given the data d (complex
visibilities, powerspectrum, phase closure, etc.):

x+ = arg max Pr(x|d) = arg max Pr(d|x) Pr(x) , (21)

x x

where the last expression comes from Bayes theorem and after having discarded the term Pr(d) which does not
depend on the sought parameters. This choice is termed the maximum a posteriori (MAP) solution and can be
recast as:
x+ = arg min f (x) (22)
x

where the penalty function f (x) writes:

f (x) = c0 c1 log Pr(x|d) = c0 c1 log Pr(d|x) c1 log Pr(x) (23)

| {z }| {z }
fdata (x) fprior (x)

where c0 and c1 > 0 are two real constants chosen for convenience. This equation shows that, to find the maximum
a posteriori solution, we must minimize a joint criterion which is the sum of two terms: a likelihood term
fdata (x) log Pr(d|x) which measures the compatibility of the parameters with the data, and a regularization
term fprior (x) log Pr(x) which imposes priors on the image.
 The expression of fdata (x) is derived from some approximations of the statistics of the data noise as explained
in Sect. 4. In practice, the prior statistics is however seldom known and the regularization must be derived from
some heuristics as we know discuss. The two main reasons to introduce the regularization in an inverse problem
are: (i) to provide additional information when the data alone cannot completely define a unique solution, and
(ii) to counter the amplification of noise for a poorly conditioned problem.10 The purpose of regularization is
then to select among all solutions compatible with the measures, whichever is closer to some priors since the data
alone are insufficient to provide a satisfactory solution. That is both stable (robust with respect to the noise)
and unique. Formally, this quest can be expressed as:
x+ = arg min fprior (x) s.t. fdata (x) (24)
x

where the inequality constraint fdata (x) imposes that the model be compatible with the data: the lower is
the closer will be the model and the measurements. Assuming that the inequality constraint is active (otherwise
the measurements are meaningless), the solution of the constrained minimization problem writes:
x+ = arg min [fprior (x) + ` fdata (x)] (25)
x

where ` > 0 is a Lagrange multiplier that have to be tuned so that fdata (x+ ) ' .
To be an effective regularization, fprior must have certain properties. In the case of image restoration in
interferometry, voids in the sampling of the u-v plane imply that the problem is most often ill-posed: solely
maximizing the likelihood of the model that is minimizing fdata (x) admits an infinite number of solutions.
In this context, regularization should help filling the lack of data at unsampled spatial frequencies, it is therefore
natural to require that regularization yields somewhat smooth interpolation between measured spatial frequen-
cies. This kind of spatial frequency smoothing can be achieved by imposing a simple quadratic constraint in the
image (see Appendix A) or by using one of the maximum entropy methods1, 2 which have proved their worth in
radio astronomy. Also note that building-block methods4, 5 yield a regularized solution by requiring that it has
only a limited number of significant components.

4. DATA PENALTY
In its current form, Mira can account for three different kind of interferometric measurements: complex visi-
bilities v data , powerspectrum sdata and phase closure data . Hence, the data penalty fdata (x) to fit these data
writes:
fdata (x) = fv (x) + fs (x) + f (x) , (26)
where fv (x), fs (x) and f (x) are respectively penalty terms with respect to complex visibility data, powerspec-
trum data and phase closure data. Note that this definition assumes that data of different types are uncorrelated.

4.1 Complex Visibility Data

In Mira it is assumed that interferometric data are independent and that complex visibility errors have Gaussian
distribution. The penalty with respect to complex visibility data then writes:
res

!T rr ri
! res

!
X X Re vk,` (x, t) Wk,` (t) Wk,` (t) Re vk,` (x, t)
fv (x) = res

ri ii
res

. (27)
t k<`
Im vk,` (x, t) Wk,` (t) Wk,` (t) Im vk,` (x, t)

where the complex visibility residuals are:

res def data model
vk,` (x, t) = vk,` (t) vk,` (x, t) (28)
and the weights:
rr
 rr ii ri
1 ii
Wk,` (t) = Ck,` (t) Ck,` (t) Ck,` (t)2 Ck,` (t) , (29)
ri 2 1
 rr ii ri
 ri
Wk,` (t) = Ck,` (t) Ck,` (t) Ck,` (t) Ck,` (t) , (30)
ii rr ii ri 1
(t)2 rr
 
Wk,` (t) = Ck,` (t) Ck,` (t) Ck,` Ck,` (t) , (31)
are computed from the covariances of the complex visibility data:
rr data


Ck,` (t) = Var Re vk,` (t) , (32)
ri data data



Ck,` (t) = Cov Re vk,` (t) , Im vk,` (t) , (33)
ii data


Ck,` (t) = Var Im vk,` (t) . (34)

For interferometric data, the so-called Goodman model can be a good approximation of the distribution of
complex visibility errors. In this case, the real and imaginary parts of the complex visibility are independant,
ri
hence Ck,` (t) = 0, and have the same variance; the penalty then simplifies to:
 2
 data model
(t) vk,` (x, t)

X X vk,`
fv (x) =  data , (35)
t k<`
Var |vk,` (t)|

rr ii
 data
where Ck,` (t) = Ck,` (t) = Var |vk,` (t)| .
OI-FITS11 is a file format which has became the standard for the storage and exchange of optical interfer-
ometry data. This format is very versatile but has a number of restrictions when processing OI-FITS data.
The data statistics is only provided by the standard deviation of the measurements, hence there is no means to
account for correlations. Since complex visibilities and bispectrum are provided in polar form, this means that
the amplitude and phase of complex data are independent. Simple geometrical considerations, yield a quadratic
approximation similar to that in Eq. (27) but with weights:

rr
cos2 data
k,` (t) sin2 data
k,` (t)
Wk,` (t) = + , (36)
Var{data
k,` (t)} data 2 data
k,` (t) Var{k,` (t)}
!
ri 1 1
Wk,` (t) = data 2 cos data data
k,` (t) sin k,` (t) , (37)
Var{data
k,` (t)} k,` (t) Var{ data (t)}
k,`

ii
sin2 data
k,` (t) cos2 data
k,` (t)
Wk,` (t) = + , (38)
Var{data
k,` (t)} data 2 data
k,` (t) Var{k,` (t)}

and with residuals computed for:

data
vk,` (t) = data data
k,` (t) exp(i k,` (t)) , (39)
where data data
k,` (t) and k,` (t) are the amplitude and phase of the measured complex visibilities. Convex approxi-
mations of the likelhood penalty have also been studied by Meimon et al.12

4.2 Powerspectrum Data

Assuming normally distributed errors for the powerspectrum, the term fs (x) in Eq. (26) write:
 2
X X sdata
k,` (t) smodel
k,` (x, t)
fs (x) = . (40)
t k<`
Var[sdata
k,` (t)]

4.3 Phase Closure Data

In order to account for phase wrapping and to avoid excessive non-linearity, the term related to the phase closures
data is defined by Mira to be the weighted quadratic distance between the complex phasors rather than between
the phases closures: 2
X X 1  data
 i j,k,` (t) model
i j,k,` (x,t) 
f (x) = data (t)] e e  . (41)
t j<k<`
Var[j,k,`
In the limit of small phase closure errors, the penalty becomes:
h i2
data model
X X j,k,` (t) j,k,` (x, t)
f (x) ' data (t)]
(42)
t j<k<`
Var[j,k,`

which is readily the 2 term that would be obtained for Gaussian phase statistics. This justifies the weighting
used in Eq. (41). Other methods have been proposed to cope with the phase wrapping13, 14 but we have found
that, in practice, they can slow down or prevent the convergence of the algorithm.

5. REGULARIZATION
Mira has been designed to be versatile in terms of input data and type of regularization. The Yorick version
of Mira let the user define its own regularization to match his priors. A number of different regularizations are
already built into Mira which are summarized in what follows.

5.1 Quadratic Regularizations

Mira implements a generic quadratic regularization:
T
fprior (x) = (Aprior x bprior ) Wprior (Aprior x bprior ) , (43)

where the matrices Wprior and Aprior and the vector bprior can be chosen to reproduce any quadratic regularization.
The weighting matrix Wprior must be symmetrical and positive semi-definite for fprior (x) to be convex.
For instance, taking Wprior = Aprior = I and bprior = 0 yields Tikhonovs regularization which is the most
simple quadratic one: X 2
fprior (x) = x2j = kxk2 , (44)
j

its effects are to limit the number of significant pixels (although not as well as with an `1 norm) and to introduce
some sort of smoothness.
In a Bayesian framework and assuming Gaussian statistics for the priors, the expected value xprior = hxi and
covariance matrix Cprior = h(x xprior ) (x xprior )T i are assumed to be known. The corresponding regularization
is quadratic and writes:
T
fprior (x) = (x xprior ) C1 prior (x xprior ) , (45)
which can be implemented thanks to the generic expression in Eq. (43).
The following quadratic regularization:
2
fprior (x) = kD xk , (46)

(with D a finite difference operator) enforces smoothness but is mostly interesting for ill-conditioned inverse
problems such as image deconvolution to avoid noise amplification. In effect, effective regularization for inter-
ferometric data should induce smooth interpolation of the missing frequencies or, similarly, limit the number
of significant pixels in the field of view. To that end, we have proposed the following quadratic separable
regularization for Mira and Wisard:15
X
fprior (x) = x2j /xprior
j , (47)
j

where, under the normalization constraint j xj = j xprior

P P
j , the default solution is xprior which is chosen to be
strictly positive and properly normalized as shown in Appendix A.
5.2 Maximum Entropy
Mira implements several regularization penalties to build the maximum entropy 1, 2 image from the interfero-
metric data:
X
fent1 (x) = xj (48)
j
X
fent2 (x) = log(xj ) (49)
j
X
fent3 (x) = xj log(xj ) (50)
j
X  
fent4 (x; xprior ) = xj log xj /xprior
j (51)
j
X h prior  i
fent5 (x; xprior ) = xj xj + xj log xj /xprior j (52)
j
X
fent6 (x; S) = xj log (xj /(S x)j ) (53)
j
X
fent7 (x; S) = [(S x)j xj + xj log (xj /(S x)j )] (54)
j

where S is a linear operator which defines a so-called floating prior 1618 xprior = S x that depends on x. Note
that if x and the default solution xprior are normalized to the same value, then fent5 (x; xprior ) in Eq. (52) is equal
to fent4 (x; xprior ) in Eq. (51) and fent7 (x; S) in Eq. (54) is equal to fent6 (x; S) in Eq. (53).

5.3 Other regularizations

Quadratic regularizations are known to somewhat oversmooth the resulting images. This is particularly incon-
venient for astronomical objects which have high dynamical range and high frequency contents due to point-like
sources or sharp edges. To let some sharp features appear in the restored image, an edge-preserving regularization
can be used:
X q 
fprior (x) = 2 2
(Dj x)k +   (55)
j,k

where Dj is a finite difference linear operator approximating the partial spatial derivative of its argument along
k-th direction (horizontal, vertical and, perhaps, diagonals) and  > 0 is a threshold. For small absolute finite
differences with respect to , the regularization is approximately quadratic (`2 ); while for large differences, the
regularization is approximately linear (`1 ).
For objects with a mixture of point-like structures and a smooth background, the following regularization
has proved effective:
X q 
2
fprior (x) = 0 x2j + 2  + 1 kD xk2 (56)
j

where the first term (with  a very small positive value) is approximately the `1 norm of the image and its
effect is to limit the number of bright pixels in the image, and where the second term (with D a finite difference
operator) enforces smoothness of the image to avoid spurious high frequencies.

6. ALGORITHM SUMMARY
The approach of algorithm Mira is to seek for the image x Rn , n being the number of pixels, by directly
20

minimizing a joint criterion under constraints of positivity and normalization:

X
f (x) = fdata (x) + fprior (x) s.t. x 0 and xj = (57)
j

where fdata see Eq. (26) enforces agreement with the measurements, fprior see Sect. 5 is a regularization
term which enforces other a priori constraints and > 0 is used to tune the weight of the priors. Taking = 1/`,
this definition is directly related to the Lagrangian in Eq. (25).
Mira implements various priors such as maximum entropy, quadratic (`2 ) smoothness, edged-preserving
(`2 `1 ) smoothness, etc. Moreover, the algorithm is designed so that the user can plug its own regularization.
Figure 2. Image reconstruction from simulated data for the 2004 Optical/IR Interferometry Imaging Beauty Contest.19
Form left to right: true image, u-v coverage, true image smoothed at the resolution of the interferometer, image restored
from powerspectrum and phase closure data, image restored without any phase data. The wavelength for the simulation
is = 0.55 m.

Practical means to automatically set the value of or, equivalently, to set the value of in Eq. (24) have
been proposed.2123 The most effective is the method proposed by Skilling & Bryan24 for maximum entropy
regularization and which has been adapted to other kind of regularizations.18 In the current version of Mira,
the tuning of the hyper-parameter = 1/` is done by the user. In the reconstructions shown in Fig. 2 and Fig. 3,
the regularization level is tuned so as to have fdata (x) m where m is the number of measurements.
Given the data, the regularization and its level, the criterion f (x) is multimodal unless all data consist in
complex visibilities. Ideally the solution should then be sought by means of a global optimization method. Owing
to the large number of parameters (the number n of pixels in the image x), global optimization would require
unpractical amount of computation. The strategy used by Mira is to perform only local optimization starting
from an initial image. The final image obtained by Mira therefore depends on the data and on the priors but
also on the initial image and on the path followed by the local optimization method.
To minimize the criterion, Mira uses the optimization method VMLMB,25 a limited variable metric algorithm
which accounts for parameter bounds. This last feature is used to enforce positivity of the solution. Only the
value of the cost function and its gradient are needed by VMLMB. Normalization of the solution is obtained by
a change of variables, the image brightness distribution becoming:

x0j
xj = P 0 , (58)
k xk

where x0 Rn are the variables seen by the optimizer with the constraints that x0j 0, j. Thus the image x is
both normalized and positive.
Examples of image reconstructions from simulated and real data are shown by Fig. 2 and Fig. 3. The two
rightmost images in Fig. 2 were restored with Mira from data simulated for the first Optical/IR Interferometry
Imaging Beauty Contest;19 note that the last image was recovered without any phase information demonstrating
the ability of Mira to cope with phaseless data. The rightmost image in Fig. 3 was recovered by Mira from
real IOTA data of the red giant star Arcturus.26 The regularization is the quadratic one given by Eq. (45) with
a prior set by a parametric fit of the data. This procedure was intended to check whether the interferometric
data were compatible with more features than a simple limb darkened star.

7. COMPARISON WITH OTHER METHODS

There exists a number of methods designed to cope with the kind of data provided by optical interferometry. The
self-calibration technique2731 has been developed for radio-astronomy and consists in deriving the Fourier phases
at measured frequencies so that they match the phase closure data and otherwise remain as close as possible to the
Fourier phase of the current image model. In an approach similar to the technique of self-calibration, Wisard32
explicitly deals with phase ambiguities introducing as few new unknowns as possible to convert the phase closure
data into Fourier phase pseudo-data. Bsmem33 and the building-blocks method5 attempt to reconstruct an
image such that its bispectrum is in agreement with the bispectrum data. These two methods differ in their
optimization strategy and in their regularization: the building-blocks method is a matching pursuit algorithm
with an implicit regularization imposed by limiting the number of building-blocks; whereas Bsmem uses Skilling
 = 1.53 m
0.4
20 = 1.55 m 30
= 1.57 m
= 1.59 m 20

Astronomical Units
0.2

relative Dec [mas]

North direction [M]
10 = 1.61 m
= 1.63 m 10
= 1.65 m
= 1.67 m 0 0.0
0
= 1.69 m
= 1.71 m
10
= 1.74 m
10 = 1.75 m 0.2
= 1.78 m
20
= 1.80 m
30
20
0.4
20 10 0 10 20 30 20 10 0 10 20 30 30 20 10 0 10 20 30
West direction [M] Relative RA [mas] Relative RA [mas]

Figure 3. Image of the red giant star Arcturus. Left: u-v coverage with IOTA interferometer. Middle: parametric
reconstruction by a limb darkening power law. Right: reconstruction by Mira algorithm. Data courtesy: S. Lacour.

& Bryan24 method to find the maximum entropy image which matches the phase closure data. The approach of
Mira is somewhat similar to Bsmem and the building-blocks method in that the algorithm directly fit the phase
closures. However Mira implements many different regularization methods and is able to account for any kind
of data. In particular, since Mira does not attempt to explicitly solve degeneracies, it can be used to restore an
image of course with at least a 180 orientation ambiguity from the power spectrum only, i.e. without any
phase information34 as shown in Fig. 2.

8. CONCLUSIONS
With the development of optical interferometers arise the needs for image restoration algorithms dedicated to cope
with this particular kind of data. In this framework, Mira (a Multi-Aperture image Reconstruction Algorithm)
has been designed to account for any kind of interferometric data (squared visibilities, complex visibilities, phase
closures, ...) and make uses of proper regularization and constrained non-linear optimization to seek for the
object brightness distribution. We have shown that Mira is able to produce an image from with very limited
amount of data, it is therefore specially efficient in the case of optical interferometry where the u-v coverage is
very poor and where Fourier phase information can be weaker and degenerated (e.g. because it is only provided
by phase closures). As an extreme case, Mira is able to restore images without any phase information. This
leads to the possibility to perform imaging with only 2 telescopes or when the phase closure data are corrupted.

ACKNOWLEDGMENTS
Mira algorithm has been implemented and tested with Yorick (http://yorick.sourceforge.net/) which is
available for free.

APPENDIX A. LIMITED FIELD OF VIEW

A general expression for a quadratic separable regularization is given by:
X
fprior (x) = wj x2j , (59)
j

where wj > 0, j, otherwise the criterion is degenerated. Note that this criterion can be seen as enforcing
a loose support constraint with a weighting w > 0 (this notation means wj > 0, j). The default solution
xprior is obtained by minimizing the cost function in the absence of data and subject to the normalization and
non-negativity constraints:
X
xprior = arg min fprior (x) s.t. x 0 and xj = , (60)
x j

where > 0 is the total flux of the solution. Assuming for the moment that all enequality constraints are inactive
at the solution, the Lagrangian for the constrained problem can be written as:
X 
L(x; `) = fprior (x) 2 ` xj , (61)
j
where ` is the Lagrange multiplier associated with the normalization constraint.35 Minimizing L(x; `) with
respect to x yields:
1
x+ (`) = arg min L(x; `) x+ j (`) = ` wj .
x
+
The optimal Lagrange multiplier ` is then identified by requiring the normalization:
X X
= x+ +
j (` ) = `
+
wj1 `+ = P 1 .
j wj
j
j

Finally the default solution is:

wj1
xprior
j = x+ +
j (` ) = P 1 , (62)
j wj

which is normalized and striclty positive since w > 0. Hence this validate our hypothesis that the inequality
constraints were all inactive at the solution. Replacing the weights by their values for a given default solution,
the regularization simply writes: X
fprior (x) = x2j /xprior
j , (63)
j

xprior
P
where xprior is striclty positive and properly normalized: j j = .

REFERENCES
[1] Narayan, R. and Nityananda, R., Maximum entropy image restoration in astronomy, Ann. Rev. Astron.
Astrophys. 24, 127170 (1986).
[2] Cornwell, T., Imaging concepts, in [ASP Conf. Ser. 82: Very Long Baseline Interferometry and the
VLBA ], Zensus, J. A., Diamond, P. J., and Napier, P. J., eds., 39+ (1995).
[3] Giovannelli, J.-F. and Coulais, A., Positive deconvolution for superimposed extended source and point
sources, Astron. & Astrophys. 439, 401412 (Aug. 2005).
[4] Hogbom, J. A., Aperture synthesis with a non-regular distribution of interferometer baselines, A&AS 15,
417426 (June 1974).
[5] Hofmann, K.-H. and Weigelt, G., Iterative image reconstruction from the bispectrum, Astron. & Astro-
phys. 278, 328339 (Oct. 1993).
[6] Lannes, A., Anterrieu, E., and Marechal, P., Clean and wipe, Astron. & Astrophys. Suppl. 123, 183198
(May 1997).
[7] Sramek, R. A. and Schwab, F. R., Imaging, in [Synthesis Imaging in Radio Astronomy ], Perley, R. A.,
Schwab, F. R., and Bridle, A. H., eds., Astronomical Society of the Pacific Conference Series 6, 117+
(1989).
[8] Cornwell, T. and Braun, R., Deconvolution, in [Synthesis Imaging in Radio Astronomy], Perley, R. A.,
Schwab, F. R., and Bridle, A. H., eds., Astronomical Society of the Pacific Conference Series 6, 167+
(1989).
[9] Tarantola, A., [Inverse Problem Theory and Methods for Model Parameter Estimation ], SIAM (2005).
[10] Thiebaut, E., Introduction to image reconstruction and inverse problems, in [Optics in Astrophysics ],
Foy, R. and Foy, F.-C., eds., NATO ASI, Kluwer Academic (2005).
[11] Pauls, T. A., Young, J. S., Cotton, W. D., and Monnier, J. D., A data exchange standard for optical
(visible/ir) interferometry, Publications of the Astronomical Society of the Pacific 117, 12551262 (Nov.
2005).
[12] Meimon, S., Mugnier, L. M., and Besnerais, G. L., Convex approximation to the likelihood criterion for
aperture synthesis imaging, J. Opt. Soc. Am. A 22, 23482356 (2005).
[13] Haniff, C., Least-squares fourier phase estimation from the modulo 2 bispectrum phase, J. Opt. Soc.
Am. A 8, 134140 (Jan. 1991).
[14] Lannes, A., Integer ambiguity resolution in phase closure imaging, J. Opt. Soc. Am. A 18, 10461055
(2001).
[15] Besnerais, G. L., Lacour, S., Mugnier, L. M., Thiebaut, E., Perrin, G., and Meimon, S., Imaging with
long-baseline optical interferometry, IEEE Journal of Selected Topics in Signal Processing (submitted in
2008).
[16] Horn, K., Images of accretion discs i. the eclipse mapping method, Mont. Not. R. Astron. Soc. 213,
129141 (1985).
[17] Lucy, L. B., Optimum strategies for inverse problems in statistical astronomy, Astron. & Astrophys. 289,
983994 (Sept. 1994).
[18] Pichon, C. and Thiebaut, E., Non-parametric reconstruction of distribution functions from observed galac-
tic discs, Mont. Not. R. Astron. Soc. 301(2), 419434 (1998).
[19] Lawson, P. R., Cotton, W. D., Hummel, C. A., Monnier, J. D., Zhao, M., Young, J. S., Thorsteinsson,
H., Meimon, S. C., Mugnier, L., Le Besnerais, G., Thibaut, E., and Tuthill, P. G., The 2004 optical/ir
interferometry imaging beauty contest, American Astronomical Society Meeting Abstracts 205, + (Dec.
2004).
[20] Thiebaut, E., Garcia, P. J. V., and Foy, R., Imaging with amber/vlti: the case of microjets, Ap&SS 286,
171176 (2003).
[21] Golub, G. H., Heath, M., and Wahba, G., Generalized cross-validation as a method for choosing a good
ridge parameter, Technometrics 21, 215223 (1979).
[22] Titterington, D. M., General structure of regularization procedures in image reconstruction, Astron. &
Astrophys. 144, 381387 (1985).
[23] Gull, S. F., [Maximum Entropy and Bayesian Methods ], ch. Developments in maximum entropy data anal-
ysis, 5372, Kluwer Academic (1989).
[24] Skilling, J. and Bryan, R. K., Maximum entropy image reconstruction: general algorithm, Mont. Not. R.
Astron. Soc. 211, 111124 (1984).
[25] Thiebaut, E., Optimization issues in blind deconvolution algorithms, in [Astronomical Data Analysis II],
Starck, J.-L. and Murtagh, F. D., eds., 4847, 174183, SPIE (2002).
[26] Lacour, S., Meimon, S., Thiebaut, E., Perrin, G., Verhoelst, T., Pedretti, E., Schuller, P. A., Mugnier, L.,
Monnier, J., Berger, J., Haubois, X., Poncelet, A., Besnerais, G. L., Eriksson, K., Millan-Gabet, R., Lacasse,
M., and Traub, W., The limb-darkened arcturus - imaging with the iota/ionic interferometer, Astron. &
Astrophys. accepted for publication (2008).
[27] Readhead, A. C. S. and Wilkinson, P. N., The mapping of compact radio sources from VLBI data,
Astrophys. J. 223, 2536 (July 1978).
[28] Cotton, W. D., A method of mapping compact structure in radio sources using VLBI observations, Astron.
J. 84, 11221128 (Aug. 1979).
[29] Cornwell, T. J. and Wilkinson, P. N., A new method for making maps with unstable radio interferometers,
Mont. Not. R. Astron. Soc. 196, 10671086 (Sept. 1981).
[30] Pearson, T. J. and Readhead, A. C. S., Image Formation by Self-Calibration in Radio Astronomy, Ann.
Rev. Astron. Astrophys. 22, 97130 (1984).
[31] Cornwell, T. and Fomalont, E. B., Self-Calibration, in [Synthesis Imaging in Radio Astronomy ], Perley,
R. A., Schwab, F. R., and Bridle, A. H., eds., Astronomical Society of the Pacific Conference Series 6,
185+ (1989).
[32] Meimon, S., Mugnier, L. M., and Besnerais, G. L., Reconstruction method for weak-phase optical inter-
ferometry, Optics Letters 30, 18091811 (2005).
[33] Buscher, D. F., Direct maximum-entropy image reconstruction from the bispectrum, in [IAU Symp. 158:
Very High Angular Resolution Imaging], Robertson, J. G. and Tango, W. J., eds., 91+ (1994).
[34] Thiebaut, E., Reconstruction dimage en interferometrie optique, in [XXIme Colloque GRETSI], GRETSI
(2007).
[35] Nocedal, J. and Wright, S. J., [Numerical Optimization ], Springer Verlag, second edition ed. (2006).