Phase retrieval algorithms: a comparison

J. R. Fienup

Iterative algorithms for phase retrieval from intensity data are compared to gradient search methods. Both
the problem of phase retrieval from two intensity measurements (in electron microscopy or wave front sens-
ing) and the problem of phase retrieval from a single intensity measurement plus a non-negativity constraint
(in astronomy) are considered, with emphasis on the latter. It is shown that both the error-reduction algo-
rithm for the problem of a single intensity measurement and the Gerchberg-Saxton algorithm for the prob-
lem of two intensity measurements converge. The error-reduction algorithm is also shown to be closely re-
lated to the steepest-descent method. Other algorithms, including the input-output algorithm and the con-
jugate-gradient method, are shown to converge in practice much faster than the error-reduction algorithm.
Examples are shown.

1. Introduction IF(u) I and lf(x) . For the problem of recovering phase

In electron microscopy, wave front sensing, astrono- from a single intensity measurement, as in image re-
my, crystallography, and in other fields one often wishes covery from speckle interferometry data in astronomy
to recover phase, although only intensity measurements and from structure factors in crystallography, one
are made. One is usually interested in determining an wishes to recover (u) or equivalently recover f(x) given
object, f(x), which is related to its Fourier transform, a measurement of IF(u) I and the constraint that f (x)
F(u), by be real and non-negative,
F(u) = JF(u)I exp[i4'(u)] = S[f(x)] f(x) 0. (5)
A particularly successful approach to solving these
= f(x) exp(-i27ru - x)dx, (1) problems is the use of the Gerchberg-Saxton algo-
where x is an M-dimensional spatial coordinate, and u rithm",2 and related algorithms. 3 -7 Reference 7 lists a
is an M-dimensional spatial frequency coordinate. For large number of different problems which have been
the majority of interesting problems M = 2. In practice solved by these algorithms. These algorithms involve
one deals with sampled data in the computer, where for iterative Fourier transformation back and forth between
the 2-D case, assuming square arrays, u = (l,u the object and Fourier domains and application of the
2) and
x = (,,x 2), where u, u 2, x, and x2 = 0,1,2, .. ,N- 1. measured data or known constraints in each domain.
Then one uses the discrete Fourier transform (DFT) In what follows a generalized Gerchberg-Saxton al-
gorithm, referred to as the error-reduction algorithm,
N-1
F(u) = E f(x) exp(-i2ru x/N) (2) and its convergence properties are reviewed (Sec. II),
x=0 and it is shown to be equivalent to the steepest-descent
and its inverse gradient search method (Sec. III). Alternative gradient
N-1
search methods (Sec. IV) and iterative Fourier trans-
f (x) = N-2 Ej F(u) exp(i27ru xIN), (3) form algorithms (Sec. V) are described and are shown
U=0
to converge much faster than the error-reduction al-
which are, of course, computed using the fast Fourier gorithm for the problem of a single intensity measure-
transform (FFT) method. ment (Sec. VI). Some practical considerations are
For the problem of recovering phase from two in- discussed in Sec. VII. A typical reconstruction exper-
tensity measurements, as in electron microscopy and iment is shown in Sec. VIII, and the major conclusions
in wave front sensing, are summarized in Sec. IX.
f(x) = f(x)I exp[iti(x)] (4) II. Error-Reduction Algorithm
is complex valued, and one wishes to recover +f(u) or The Gerchberg-Saxton algorithm was originally in-
equivalently recover 1(x) from measurements of both vented in connection with the problem of reconstructing
phase from two intensity measurements 1' 2 (and for
synthesizing phase codes given intensity constraints in
The author is with Environmental Research Institute of Michigan, each of two domains 8 9). The algorithm consists of the
Radar & Optics Division, P.O. Box 8618, Ann Arbor, Michigan following four simple steps: (1) Fourier transform an
48107. estimate of the object; (2) replace the modulus of the
Received 19 February 1982. resulting computed Fourier transform with the mea-
0003-6935/82/152758-12$01.00/0. sured Fourier modulus to form an estimate of the Fou-
©1982 Optical Society of America. rier transform; (3) inverse Fourier transform the esti-

2758 APPLIED OPTICS / Vol. 21, No. 15 / 1 August 1982

 9 G =IGI eGI object-domain constraints, i.e., wherever gk(x) is neg-
ative or (optionally) where it exceeds the known diam-
eter of the object. The diameter of the object can be
computed since it is just half of the diameter of the au-
SATISFY SATISFY tocorrelation function, which is the inverse Fourier
FUNCTION FOURIER
CONSTRAINTS CONSTRAINTS transform of the squared Fourier modulus. (However,
in two dimensions the exact support of the object cannot
in general be determined uniquely from the support of
9 - G = IFe'0 its autocorrelation, 1 0 and so the diameter constraint
cannot be applied very tightly.)
Fig. 1. Block diagram of the error-reduction (Gerchberg-Saxton) The iterations continue until the computed Fourier
algorithm. transform satisfies the Fourier-domain constraints or
the computed image satisfies the object-domain con-
straints; then one has found a solution, a Fourier
mate of the Fourier transform; and (4) replace the transform pair that satisfies all the constraints in both
modulus of the resulting computed image with the domains. The convergence of the algorithm can be
measured object modulus to form a new estimate of the monitored by computing the squared error. In the
object. In equations this is, for the kth iteration, Fourier domain the squared error is the sum of the
squares of the amounts by which Gk (M), the computed
Gk(u) = IGk(u)I exp[ikM(u)] = [gA(x)], (6) Fourier transform, violates the Fourier-domain con-
GQ(u) = IF(u)l exp[iok(u)], (7) straints. Since GQ(u) was formed from Gk(u) by
g'(x) = lg'(x)l exp[i0'(x)] = 51-[GQ(u)],
making the minimum changes to satisfy the Fourier-
(8)
domain constraints, the squared error can be expressed
gk+1(x) = f(x)I exp[ik+1(x)I = If(x)l exp[iO(x)], (9) as
where gk, ok, G', and kk are estimates of f, a, F, and Ak, Bk = E2= N-
2
E Gk(u) -u)2 (11)
respectively. Here and throughout this paper functions
represented by uppercase letters are the Fourier which, for both problems being considered, can be ex-
transforms of the functions represented by the corre- pressed as
sponding lowercase letters.
2
As depicted in Fig. 1 the Gerchberg-Saxton algorithm Bk = = N- E [JGk(u)J - IF(u)] 2 . (12)
is easily generalized to a large class of problems. 6' 7 The
generalized Gerchberg-Saxton algorithm can be used In this section the symbol E k is used to distinguish it
for any problem in which partial constraints (in the form from the object-domain error Ek described below. For
of measured data or information known a priori) are economy of notation in the section of this paper dealing
known in each of two domains, usually the object (or with the gradient methods, the symbol Bk is used in-
image) and Fourier domains. One simply transforms stead of E k. The symbol B (with the subscript k de-
back and forth between the two domains, satisfying the leted) is given by Eq. (11) with G and G' replacing Gk
constraints in one before returning to the other. This and G, respectively.
generalization of the Gerchberg-Saxton algorithm will Similarly, for the error-reduction algorithm the
be referred to as the error-reduction algorithm since, as squared error in the object domain can be expressed
will be shown below, the error decreases at each itera- as
tion.
Ok = E gk+l(X) -gk(X)12,
For the most general problem the error-reduction x
(13)
algorithm consists of the following four steps: (1) which for the problem of two intensity measurements
Fourier transform g (x), an estimate of f(x); (2) make can be expressed as
the minimum changes in Gk (M),the resulting computed
Fourier transform, which allow it to satisfy the Fou- E = E [If(X)| - gk(X)J]2
rier-domain constraints to form G'k (u), an estimate of (14)
F(u); (3) inverse Fourier transform G (u); and (4) make and for the problem of a single intensity measurement
the minimum changes in gk (x), the resulting computed can be expressed as
image, which allow it to satisfy the object-domain con-
straints to form gk+1(x), a new estimate of the object. E~k = _ [gh(X)]2, (15)
In particular, for the problem of a single intensity Xey

measurement (as in astronomy) the first three steps are where y is defined as in Eq. (10). The asymmetry in the
identical to the first three steps of the Gerchberg-Saxton use of the N- 2 factor above was chosen because of the
algorithm, Eqs. (6)-(8), and the fourth step is given similar asymmetry in the definition of the discrete
by Fourier transform in Eqs. (2) and (3). When the
squared error is zero, a solution has been found.
gk+:1(x) = {k(X). x7, (10) In the following the error-reduction algorithm is
wOi Xc hv shown to converge, and this convergence property holds
where -y is the set of points at which gk(x) violates the for all the applications of the error-reduction algorithm

1 August 1982 / Vol. 21, No. 15 / APPLIED OPTICS 2759

(not just the problems of recovering phase from single error once again starts to decrease until it reaches a third
or two intensity measurements). plateau at the level of 0.003, at which point the decrease
For the general problem, at the kth iteration the in the error is negligible.
squared error in the Fourier domain is given by Eq. (11). The occurrences of plateaus during which conver-
By Parseval's theorem'1 gence is extremely slow seem to occur, more often than
not, for both applications,' 9" 2 and in the past it has
Ek =N - 2 Gk(u)-Gk(u)12 been observed that, with persistence, one can go beyond
the plateau region and again make rapid progress
=E gk(x)-gk(x)12 . (16) toward a solution."1 2 However, as shown in the ex-
ample of Fig. 2 for the problem of a single intensity
Now compare this with Eq. (13), the error in the object measurement, the number of iterations required for
domain. Both gk (x) and gk+l(x) by definition satisfy convergence of the error-reduction algorithm can be
the object-domain constraints. Also at any point x, by extremely large, making that algorithm unsatisfactory
definition gk+l(x) is the nearest value to gk(x) that for that application. Fortunately, as will be shown in
satisfies the object-domain constraints. Therefore, at later sections, other related algorithms converge much
all points x faster, reconstructing a recognizable image in twenty
or thirty iterations and completing the reconstruction
lgk+l(X) -g( )| < 19k W - 9a(X) 1, (17) in under one-hundred iterations, which takes <2 min
and, therefore, from Eqs. (13) and (16) on a Floating Point System AP-120B array processor
for array sizes of 128 X 128.
EO < EFk- (18)

Similarly, by Parseval's theorem Ill. Steepest-Descent Method

An alternative approach to solving the phase-retrieval
2
E2 = E lgk+1(X) -gk(X) problems is to employ one of the gradient search
methods. In this section it is shown that one such
= N- 2 E Gk+1(u) - k(x)12 . (19) method, the steepest-descent method, is closely related
to the error-reduction algorithm for the problem of re-
Since both G(u) and G'+J(u) satisfy the Four rier- constructing phase from a single intensity measure-
domain constraints, and at any point u, by definiition ment. The relationship between the steepest-descent
Gk+1(u), is the nearest value to Gk+l(u) that satis;fies method and the Gerchberg-Saxton algorithm is also
the Fourier-domain constraints, then discussed.
An example of how a gradient search method would
IGk+1(u) - Gk+l(u) Gk+1(u) - k(u)1. (20) be used for this problem follows. One can define B =
Therefore, from Eqs. (11) and (19), EF, the squared error in the Fourier domain given by
Eq. (12), as the error metric which one seeks to minimize
EF,k+l < Ek,

and combining this with Eq. (18) gives the desiredI re- ITERATION k
10 102 103
sult
EF'k+1 < E0k S< EFk. (22)

That is, the error can only decrease (or stay the same)
at each iteration.
In practice, the error-reduction algorithm usually
decreases the error rapidly for the first few iterations lo-I.
but much more slowly for later iterations.1 2 9 12 The
speed of convergence also depends on the type of con- cY
straints imposed. Convergence seems to be reasonably 0
oc
fast for the problem of two intensity measurements but tY

painfully slow for the problem of a single intensity cY

measurement. Figure 2 shows an example of the error

as a function of the number of iterations of the error-
reduction algorithm for the problem of a single intensity
measurement. Shown is the normalized rms error, i.e.,
the square root of EFk divided by 2; JF(u) 2. Both the
error and the iteration number are shown on logarithmic
scales. The error decreases rapidly during the first
thirty iterations but then reaches a plateau at a level of
0.16, decreasing very slowly. After seventy iterations
the error again starts to decrease until it reaches a sec- Fig. 2. RMS error vs the number of iterations for the problem of
ond plateau at the level of 0.02, at which point the error phase retrieval from a single intensity measurement using the error-
decreases extremely slowly. After 2000 iterations the reduction algorithm.

2760 APPLIED OPTICS / Vol. 21, No. 15 / 1 August 1982

by varying a set of parameters. Here the N 2 values of g(x), and multiplying by a constant. In fact, the com-
g(x), the estimate of f(x), are treated as N 2 independent putation of g'(x) is identical to the first three steps of
parameters. Starting at a given point, gk (x), in the the error-reduction algorithm.
N 2-dimensional parameter space, one would reduce the The optimum step size to take in the direction of the
error by computing the partial derivatives of B with gradient can be determined by forming a first-order
respect to each of the points g(x) (N 2 partial deriva- Taylor series expansion of B as a function of g(x) about
tives) forming the gradient and then move from g (x) the point gk (x),
in a direction opposite that of the gradient to a new
point gk(x). One would then form a new estimate, B Bk + Z agBk g(x) -g,,(x)]. (29)
gk+,(x), from g^(x) by forcing the object-domain con-
straints to be satisfied. This would be done iteratively This first-order expansion of B is equal to zero at g(x)
until a minimum (it is to be hoped a global minimum) = g (x) given by
is found. That is, one minimizes the error, B, as a
function of the N 2 parameters, g(x), subject to the g(x) -gk(X) = - BkodgBk (30)
object-domain constraints. F (gBk )2
Ordinarily the computation of the N 2 partial deriv-
atives would be a very lengthy task since each evaluation which can be easily verified by inserting Eq. (30) into
of B involves an N X N discrete Fourier transform. Eq. (29). Since by Eqs. (28) and (16)
However, for the problems considered here the com-
putation can be greatly reduced as described below. Z£ (agBk )2 = 4 Z [gk (y) -g' (y)]2 = 4Bk,, (31)
First, consider the problem of a single intensity Y Y

measurement. The partial derivative of B with respect Eq. (30) becomes

to a value at a given point, g(x), [if g(x) is assumed to be
real since f(x) is real], using Eq. (12), is g(x) -g,(x = -('/ 4 )agB = '/2 )[g(x) -gk(x)J. (32)
aB aIG(u)I However, since B is quadratic in g(x), the linear ap-
agBd =2N-2E [IG(u)I-F(u)II l (23)
'ag(x) U ag(x) proximation above can be expected to predict a step size
Later the notation gBk will be used to denote gB half as large as the optimum. 2 Therefore, one should
evaluated atg(x) = gk(X). Since use the double-length step,

ag x
=(9)
g
Z_
E g(y) exp[-i2iru y/N], g(x) -gk(x) = [(x) -gk(x)J

= exp[-i2ru x/N], (24)

or
then gk (X ) = gk X . (33)

IG(u) a[G(u)12]1/ 2
1 aIG(u)1 2
In fact, since G (u) I = IF(u) , moving to g(x) reduces
dg(x) d9g(x) 21G(u)I g(x) the error, Eq. (12), to exactly zero. As a final step in one
G(u) exp[i27ru x/N] + G*(u) exp[-i27ru x/N ]. (25) iteration the new estimate should be made to satisfy the
2IG(u)l object-domain constraints, which is accomplished by
using Eq. (10).
Therefore, Eq. (23) becomes Comparing this new estimate with that of the error-
reduction algorithm described in Sec. II, it is seen that
dgB = N- 2 F [G(u) - F(u)IG(u)/IG(u)] exp[i27ru *x/N] they are identical. That is, the error-reduction iterative
, ~~U Fourier transform algorithm can be looked on as a rapid
+N- 2 y [G*(u) - F(u)IG*(u)/1G(u)J] method of implementing a double-length step steep-
U

X exp[-i27ru x/N]. (26)

est-descent method.
Although the steepest-descent method is identical to
Using Eqs. (6) and (7) to define G'(u) as the error-reduction algorithm for the problem of a single
G'(u) = F(u)IG(u)/IG(u)J, (27)
intensity measurement, the connection is not so close
for the problem of two intensity measurements as ex-
and noting that Eq. (26) is in the form of a discrete plored in the Appendix. In that case the error is min-
Fourier transform, it can be reduced to imized with respect to the phase estimate (x), and the
dgB = 2[g(x) - g'(x)], (28)
derivative of the error does move one in a direction ap-
proximately toward the Gerchberg-Saxton phase, 0'(x),
where g'(x) is defined by Eq. (8), and the fact that g(x) of Eq. (9). However, according to Eq. (A4), the direc-
and g'(x) are real valued has been used. [Note that tion is somewhat different from that of the
since g'(y)/Og(x) 0, it is not true that Eq. (28) follows Gerchberg-Saxton algorithm; and the step size, ac-
immediately from Eq. (16).] From Eq. (28) it is evident cording to Eqs. (A12) and (A16), is considerably larger
that the entire gradient, consisting of the N 2 partial than that of the Gerchberg-Saxton algorithm. Exper-
derivatives, can be computed very simply by Fourier imental results using the steepest-descent method for
transforming g(x), applying Eq. (27), inverse Fourier the problem of two intensity measurements are shown
transforming to arrive at g'(x), subtracting g'(x) from in Ref. 2.

1 August 1982 / Vol. 21, No. 15 / APPLIED OPTICS 2761

 The relationship between a gradient search method 9 * 1;l WG
and the error-reduction algorithm for a problem in
digital filter design is discussed in Ref. 13.
IV. Other Gradient Search Methods SATISFY SATISFY
FUNCTION FOURIER
As shown in Sec. III for the phase problem of a single CONSTRAINTS CONSTRAINTS
intensity measurement the steepest-descent method is
equivalent to the error-reduction algorithm. And as
described in Sec. II, although there is a convergence | GRADIENT aG
proof for the error-reduction algorithm, in practice it STEP
converges very slowly for the problem of a single in- Fig. 3. Block diagram for the gradient-search methods using the
tensity measurement. Slow convergence of the steep- method of Fourier transforms to compute the gradient.
est-descent (also known as the optimum-gradient)
method has also been observed for other applications A gradient search method superior to the steepest-
as well. 1 4 In this section some other gradient search descent method is the conjugate-gradient method. For
methods are briefly described, and in Sec. VI it will be that method Eq. (34) is replaced by' 4
shown that in practice they converge much faster than
the steepest-descent method for this problem. gk(x) = gk(x) + hkD,(x), (36)
Recall from Eq. (30) that the steepest-descent where the direction Dk (x) is given by
method moves from the point g (x) in parameter space
to the point Da(x) = - (112)agBk
gk(x) = g(x) - hk,9gBk, (34) + [z (0gB)2/y (gB,,_)2] Dk_,(x), (37)
where hk, the step size, is a positive constant, and the
gradient is given by Eq. (28). For many applications which, using Eqs. (28) and (31), can be written as
one would search along the direction of -agBk, evalu- Dk(x) = gk(x) - g,,(x) + (Bk/Bk- 1)Dk-1 (x), (38)
ating B repeatedly until the value of hk that minimizes
B is found; then from that point one would recompute where one would start the first iteration with D 1 (x) =
the gradient and go off in a new direction. For this gi(x) - g1 (x). After using Eq. (36) one would employ
application, however, since an N X N Fourier transform Eq. (35) to form the new estimate as indicated in Fig.
is required to evaluate B and only one more Fourier 3.
transform is required to recompute the gradient, one Numerous variations on these gradient methods are
may as well recompute the gradient at each step. After possible. For example, one could argue that from one
each step of this or the other gradient methods de- iteration to the next the solution is going in the following
scribed later one must then satisfy the object-domain direction:
constraints to form the new estimate as was done in Eq. Dk W = 4 - - I W- (39)
(10),
Since the step in that direction may be too small, a
gk+i(x) = {gk(x) x, (35) better point to go to would be
lO Xey,
g;(x) =g*(x) +hk [4W -g4-1W] (40)
where set y is defined as in Eq. (10). In Sec. III the
optimum double-length step value of hk was shown to where the parameter hk controls the step size. In Eq.
be unity, for which the steepest-descent method is (40) one jumps from the point gk(x) rather than from
equivalent to the error-reduction algorithm. In fact, gk (X) since presumably g (x) is closer to the solution
hk = 1 leads to a point g (x) = g'k(x) at which B = 0. than g (x). After using Eq. (40) one would employ Eq.
This is not a solution, however, unless g (x) satisfies the (35) to form the new estimate.
object domain constraints. With this in mind, other A method that does not seem as practical for this
values of hk are better in practice as will be shown in problem is that of (damped) least squares (or Newton-
Sec. VI. Raphson). 1 6 Since each iteration of a least-squares
A useful block diagram representation of this and method involves the inversion of an N 2 by N 2 matrix,
other gradient methods is shown in Fig. 3, which em- a large number of iterations of one of the gradient
phasizes the similarity of gradient methods to the methods or of one of the iterative Fourier transform
error-reduction algorithm. The first three steps of the methods described in Sec. V could be performed in the
error-reduction algorithm, resulting in the computation same time it takes to do a single iteration of least
of g'(x), do most of the work of computing the gradient. squares. Furthermore, as has been discussed above, one
The final step of satisfying the object-domain con- can readily find a point ga (x) = ga(x) at which the error
straints is common to gradient methods and the error- B is equal to zero, and so a more sophisticated (and more
reduction algorithm. Therefore, one can think of the difficult) method, such as least squares, of finding such
error-reduction algorithm as a special case of a more a point is not warranted.
general class of gradient methods. For the error-re- The problem here is that one is constantly running
duction algorithm (or the double-length step steepest- into the object-domain constraints on g(x). An ap-
descent method) it just happens that g (x) = g (x). proach that would be superior to the ones considered

2762 APPLIED OPTICS / Vol. 21, No. 15 / 1 August 1982

here would be one that minimizes the Fourier-domain output in the same general direction as the change of the
error while inherently satisfying the object-domain input. More precisely, for a small change of the input
constraints, or one that minimizes an error metric that the expected value of the corresponding change of the
combines the Fourier- and object-domain constraints. output is a constant a times the change of the input.
An example of the former is the use of a gradient search Since additional nonlinear terms also appear in the
method for the problem of two intensity measurements; output, the change of the output due to a particular
by searching over 0(x) one automatically satisfies the change of the input cannot be predicted exactly.
object-domain constraints that they have a given Nevertheless, by appropriate changes of the input, the
modulus If(x) 1. Something along these lines would be output can be pushed in the general direction desired.
very useful for the problem of a single intensity mea- If a change Ag(x) is desired in the output, a logical
surement; clearly, more could be done in this area. choice of the change of the input to achieve that change
of the output would be f3Ag(x), where : is a constant
V. Input-Output Algorithm
ideally equal to -1.
A solution to the problem of the slow convergence of For the problem of phase retrieval from a single in-
the error-reduction algorithm has been the input-out- tensity measurement the desired change of the output
put algorithm, which has proved to converge faster for is
both the problem of two intensity measurements 6 17 and
the problem of a single intensity measurement. 4 5 The Agk(x) = °, XS-Y, (41)
input-output algorithm differs from the error-reduction 1k-4)> xC'y,
algorithm only in the object-domain operation. The where y is the set of points at which g', (x) violates the
first three operations-Fourier transforming g(x), object-domain constraints. That is, where the con-
satisfying the Fourier-domain constraints, and inverse straints are satisfied, one does not require a change of
Fourier transforming the result-are the same for both the output; but where the constraints are violated, the
algorithms. If grouped together as indicated in Fig. 4, desired change of the output, in order to have it satisfy
those three operations can be thought of as a nonlinear the object-domain constraints, is one that drives it to
system having an input g(x) and an output g'(x). The a value of zero (and, therefore, the desired change is the
useful property of this system is that its output is always negative of the output at those points). Therefore, a
an image having a Fourier transform that satisfies the logical choice for the next input is
Fourier-domain constraints. Therefore, if the output
also satisfies the object-domain constraints, it is a so- gk+1(X) = gk (x) + A$gk (X)

lution to the problem. Unlike the error-reduction al- = gk(x), X,'

gorithm and the gradient methods, the input g(x) no (42)
gk (X) - 03g(x), Xeno.
longer must be thought of as the current best estimate We will refer to the use of Eq. (42) as the basic input-
of the object; instead, it can be thought of as the driving output algorithm.
function for the next output, g'(x). The inputg(x) does An interesting property of the nonlinear system
not necessarily satisfy the object-domain constraints. (consisting of the set of three steps mentioned above)
This viewpoint allows one a great deal of flexibility and is that if an output g' is used as an input, its output will
inventiveness in selecting the next input, and allows for be itself. Since the Fourier transform of g' already
the invention of algorithms that converge more rapidly satisfies the Fourier-domain constraints, g' is unaffected
to a solution. The input-output algorithm, then, is as it goes through the system. Therefore, irrespective
actually a class of algorithms as will be described of what input actually resulted in the output g', the
below.
output g' can be considered to have resulted from itself
As described elsewhere, 6 7 17 it has been found that as an input. From this point of view another logical
a small change of the input results in a change of the choice for the next input is
gk+l(X) = g4(x) + 3Agk(x)

Jgkx)- x4-,
INPUT (43)
XeY.

We will refer to the use of Eq. (43) as the output-output

algorithm.
Note that if 13 = 1 in Eq. (43), the output-output al-
gorithm reduces to the error-reduction algorithm of Eq.
(10). Since the optimum value of 1 is usually not unity,
the error-reduction algorithm can be looked on as a
suboptimal version of a more general approach.
Still another method of choosing the next input which
OUTPUT 9- was investigated is a combination of the upper line of
Eq. (43) with the lower line of Eq. (42):
- gk(x), X7
gk+l(X) = gk(-X)g' (X), Xc', (44)
Fig. 4. Block diagram of the system for the input-output concept.

1 August 1982 / Vol. 21, No. 15 / APPLIED OPTICS 2763

We will refer to the use of Eq. (44) as the hybrid .08
input-output algorithm. The hybrid input-output
algorithm is an attempt to avoid a stagnation problem
that tends to occur with the output-output algorithm.
The output-output algorithm often works itself into a .06
situation in which the output on successive iterations
does not change despite being far from a solution. For
the hybrid input-output algorithm, on the other hand, u

if at a given value of x the output remains negative for 0

more than one iteration, the corresponding point in the W=
.04
Lu

input continues to grow larger and larger until eventu-

ally that output value must go non-negative.
For the input-output algorithms the error EF is
usually meaningless since the input g (X) is no longer .02
an estimate of the object. Then the meaningful error
is the object-domain error E 0 given by Eq. (15).
For the problem of phase retrieval from two intensity
measurements the desired change of the output takes
a different form, and it is described elsewhere6 7 1 7 in the 0 .................
5 10
15 20
context of a computer holography synthesis problem ITERATION NUMBER k
involving intensity constraints in each of the two do- Fig. 5. RMS error vs the number of iterations using the error-
mains. reduction algorithm.
VI. Experimental Comparison of Phase-Retrieval
Algorithms two. For those cases for which the visual quality im-
In this section the gradient search and input-output proves while E 0 does not it was found that, if one then
algorithms are compared for the problem of phase re- performs a few (say five or ten) iterations of the error-
trieval from a single intensity measurement by using reduction algorithm, the visual quality of the output
them all on the same Fourier modulus data and with the image changes very little, but E 0 decreases rapidly until
same starting input. For each approach several dif- it becomes more consistent with the visual image
ferent values of the algorithm parameter (h or 1) were quality. Therefore, to gauge the true progress of an
tried. The principal problem with the error-reduction input-output algorithm using the value of E 0, a few it-
algorithm is that it tends to stagnate after a few itera- erations of the error-reduction algorithm are performed
tions. For this reason the starting point for the itera- after the iterations of the input-output algorithm. For
tions was chosen to be a partially reconstructed image this reason, for all cases the ten iterations of the algo-
on which the error-reduction algorithm was making slow rithm being tested were followed by nine iterations of
progress. Figure 5 shows a plot of E 0, the rms error, vs the error-reduction algorithm in order to make a fair
the number of iterations beyond this starting point comparison.
using the error-reduction algorithm. Starting at Plots of E 0 after the set of nineteen iterations de-
0.071275, Eo decreased slowly but steadily to 0.067470 scribed above (ten iterations followed by nine iterations
after ten iterations and to 0.063373 after nineteen it- of the error-reduction algorithm) are shown in Fig. 6 as
erations. In this paper, all values of E 0 are normalized a function of the algorithm parameter for the various
by dividing by the square root of z [g'(x)]2 , the total algorithms. Note that both the steepest-descent
image energy. The object for the experiment described method with h = 1.0 and the output-output algorithm
in this section is a digitized photograph of a satellite in with 1 = 1.0 are equivalent to the error-reduction al-
a field of view of 128 X 128 pixels, and its Fourier gorithm at E 0 = 0.063373, both circled in Fig. 6.
modulus is noise-free. Comparing these plots it is seen that the algorithm
The algorithms were compared by performing ten which most reduced the error after the set of nineteen
iterations of each algorithm, followed by nine iterations iterations is the hybrid input-output algorithm with a
of the error-reduction algorithm (a total of nineteen value of 1 equal to about unity.
iterations) using the same starting input for each. For each algorithm the use of a small algorithm pa-
During the first ten iterations the value of the algorithm rameter (13 or h) leads to a steady but slow decline of E 0.
parameter 1 or h was held constant. The reason that Increasing the value of the parameter increases the
each algorithm was followed by nine iterations of the speed at which E 0 decreases until one reaches a point
error-reduction algorithm is as follows. In many cases where the parameter is too large and the algorithm be-
it has been observed that definite progress (i.e., im- comes unstable. The instability of the algorithm for
proved visual quality of the output image) is being made larger values of the algorithm parameter makes possible
with an input-output algorithm even though E 0 de- more than one local minimum in the plots of E 0 vs the
creases very little or even increases with each iteration. algorithm parameter.
The relationship between E 0 and the visual image For all the algorithms, keeping h or 1 fixed for all it-
quality is not fully understood, although, of course, one erations is not the best possible strategy, particularly
would expect a high degree of correlation between the for the gradient methods. At the point at which the

2764 APPLIED OPTICS / Vol. 21, No. 15 / 1 August 1982

 error-reduction algorithm is converging slowly the
gradient is small, and one must then use a large value
of h to make rapid progress. However, after using a
large value of h for a few iterations, one moves to a point
where the gradient is much larger. Then one should use
a smaller value of h to avoid algorithm instability. If
a method for adaptively choosing h at each iteration was
devised, one would expect the gradient methods to
perform considerably better than the results shown here 04
=.04 -
using a fixed value of h.
Examples of an alternative to using a fixed value of
h and 13 are shown in Fig. 7. For the first ten iterations
of each algorithm the indicated value of h or 13 was used .02
for iterations k = 1, 3, 5, 7, and 9, and the error-reduc-
tion algorithm (h = 1 for the steepest-descent and
conjugate-gradient) was used for iterations k = 2, 4, 6,
8, and 10. The iterations using the error-reduction al-
0 4 8 12 16 20 24 28
gorithm help to stabilize the algorithm by moving ALGORITHM PARAMETER 6 OR h
toward a point where the gradient is smaller. Com- Fig. 7. RMS error after a fixed number of iterations (using the al-
paring Figs. 6 and 7, it is seen that with this alternative ternative strategy) vs the algorithm parameter. Curve A: steep-
strategy the optimum value of each algorithm param- est-descent method (0); B: basic input-output algorithm (); C:
eter is considerably larger than the optimum value when output-output algorithm (); D: hybrid input-output algorithm
the parameter is kept constant. At the optimum value (A); E: the algorithm of Eq. (40) ().
of the algorithm parameter the alternative strategy gave
better results (a lower value of E 0 at the end of the se- above. This and numerous other variations on these
quence of nineteen iterations) than those shown in Fig. algorithms can be used with varying degrees of suc-
6 for the basic input-output and for the output-output cess.
algorithms; the two strategies were comparable for the The results shown in Figs. 6 and 7 were for a partic-
steepest-descent method; and for the hybrid input- ular set of Fourier modulus data for a particular stage
output algorithm the alternative strategy gave poorer of reconstruction and for a particular number of itera-
results than those shown in Fig. 6.
tions, and the results in other circumstances could be
Curve E in Fig. 7 shows the result with the algorithm significantly different. At the optimum values of the
of Eq. (40) using the alternative strategy described algorithm parameter in each case the algorithm pa-
rameter was large enough to make the algorithm
.08
somewhat unstable, and so substantially different re-
C E D A
sults could be obtained if relatively small changes in
starting point, algorithm parameter, or number of it-
erations were made. In general, slower but steadier
progress is made if an algorithm parameter is used that
.06 is somewhat smaller than the optimum according to
Figs. 6 and 7. These results do serve to show trends that
B can be expected to apply in a wider range of circum-
Lu
stances. Further development is needed to determine
0 the best approach for the general case. As of this
.04 writing the most successful strategy has been to alter-
U,

nate between several (10-30) iterations of the hybrid

input-output algorithm and a few (5-10) iterations of
the error-reduction algorithm.
.02 Figure 8 shows E 0 vs the number of iterations past the
starting point for the hybrid input-output algorithm
with 1 = 1 (curve B) and for the error-reduction algo-
rithm (curve A, repeated from Fig. 5). Curve BI shows
the results for the set of nineteen iterations described
0 1 2 3 4 5 above (ten iterations of the hybrid input-output algo-
ALGORITHM PARAMETER , OR h rithm followed by nine iterations of the error-reduction
Fig. 6. RMS error after a fixed number of iterations vs the algorithm algorithm). Curve B2 shows the results of twenty it-
parameter. Curve A: steepest-descent method (0); B: conju- erations of the hybrid input-output algorithm followed
gate-gradient method (); C: basic input-output algorithm (A); D: by a few iterations of the error-reduction algorithm.
output-output algorithm (A); E: hybrid input-output algorithm (). The instability of the hybrid input-output algorithm
The result using the error-reduction algorithm is indicated by a large is seen in curve B, in which E 0 increases from 0.071275
circle. to 0.137707 during the first four iterations. By the end

1 August 1982 / Vol. 21, No. 15 / APPLIED OPTICS 2765

 .14 the output image is constrained to be zero. That is, the
set y defined in Eq. (10) includes all points outside the
mask for which ga(x) s 0. One need not use this in-
formation for the algorithm to converge, but it is de-
sirable to do so since using this additional information
speeds up the convergence of the algorithm.
a: .08 ,.. ,. . A A problem often occurs with the diameter constraint
even though the mask region is correctly defined. If the
>06 tet partially reconstructed image g(x) is not centered in
the mask region, in applying the diameter constraint
.04 one might inadvertently be trying to chop off (truncate)
one part of the object, which usually results in stagna-
.02 d s 0 t tion of the algorithm. For this reason it is usually ad-
..................... vantageous to define the mask as being somewhat larger
0 S 10 15 20 25 than the extent of the object.
ITERATION k We have found a good strategy for choosing the mask:
Fig. 8. RMS error vs the number of iterations for the error-reduction For the first several iterations, define a smaller mask
algorithm (curve A), and for the hybrid input-output algorithm (curve which very tightly constrains the object. This helps to
B-see text). speed up the convergence of the algorithm initially, but
slows it down for later iterations when the problem
of ten iterations E decreases to 0.091176, still worse mentioned above becomes more significant. Then, for
than the starting point, although the appearance of the later iterations use a larger mask which ensures that
image is improved from the starting point. At the none of the solution is being truncated by the mask.
eleventh iteration of curve B 1, the first iteration of the Logical choices for masks are any of the locator
error-reduction algorithm, E 0 drops sharply to a value sets. 0
of 0.047244, although the appearance of the output Faster convergence can be expected if the starting
image changes little from that of the tenth iteration. If input g,(x) is closer to the solution. A good starting
after the tenth iteration the hybrid input-output iter- input is formed as follows. Compute the autocorrela-
ations are continued for ten more iterations (curve B2), tion function and then demagnify it by a factor of 2
E 0 continues to decrease to well below the level of E 0 for (save only every other pixel in both dimensions). Then
the error-reduction algorithm alone (curve A). Similar threshold the demagnified autocorrelation at a value
to the case of B1, after the twentieth iteration of the which is a small fraction of its peak, setting it equal to
hybrid input-output algorithm (curve B2), when a few zero wherever it is below the threshold value. Finally,
iterations of the error-reduction algorithm were per- replace each value above the threshold with a sample
formed, again E 0 fell rapidly to a level consistent with of a random variable uniformly distributed between
the improved image quality that was present. This is zero and unity. The result is a random (unbiased)
the characteristic of the input-output algorithms that starting input having approximately the same size and
mandated the use of a few iterations of the error-re- shape as the original object.
duction algorithm to make a fair comparison. For the A curious phenomenon often occurs for the problem
input-output algorithms the image quality is often of phase retrieval from a single-intensity measurement.
better than what one would infer from the value of The phase-retrieval algorithm often stagnates at a local
Eo. minimum characterized by a pattern of stripes across
the image. 1 8 19 In most cases the stripes are barely
VIl. Diameter Constraint, Starting Input, and Stripes noticeable and are of low contrast, superimposed on an
A side issue pertinent to all algorithms is how one otherwise excellent reconstructed image. In some cases
defines the diameter of the object for the purpose of the stripes are of high enough contrast to be objection-
applying the diameter constraint in the object domain. able, although they still permit the object to be recog-
For the problem of phase retrieval from a single inten- nized. The cause of this phenomenon is not well un-
sity measurement the object is usually of finite extent derstood, but it is thought that it is an algorithm con-
and on a dark (zero value) background. Bounds on the vergence problem rather than a uniqueness problem' 9
support of the object (the set of points over which the (it is at a local, not a global, minimum of Eo). A method
object is nonzero) can then be determined from the of avoiding this phenomenon is presently being sought,
support of the autocorrelation, which, being the inverse although it fortunately is not much of a problem in most
Fourier transform of the square of the Fourier modulus, cases.
can be computed from the given data. As shown in Ref.
10, for extended objects the support of the autocorre- Vill. Image Reconstruction Example
lation usually does not uniquely define the support of An example of a computer experiment using the it-
the object. Nevertheless, reasonably tight bounds can erative reconstruction algorithm for the problem of
be made on the support of the object. Locator sets' 0 phase retrieval from a single intensity measurement is
can be defined that contain all possible solutions. One shown in Fig. 9. In this example a realistic stimulation
can, therefore, define a region (a mask) outside of which was performed to arrive at the kind of noisy Fourier

2766 APPLIED OPTICS / Vol. 21, No. 15 / 1 August 1982

modulus data that would be provided by stellar speckle 0.5
interferometry, 2 0 including the effects of atmospheric
turbulence and photon noise.18 0.4
An undegraded object, a digitized photograph of a
satellite shown in Fig. 9(a), was convolved with 156 w
0'
0.3
different point spread functions to produce 156 differ-
ent blurred images. Each point spread function rep- In 02
resented a different realization of the effects of the
turbulent atmosphere. The blurred images were then 0.1
subjected to a Poisson noise process to simulate the
effects of photon noise. For this example there were 0 20 40 60
-300,000 photons/degraded image (or of the order of ITERATION k
100 photons/pixel over the extent of the object), which Fig. 10. RMS error vs the number of iterations.
is realistic for objects of this type when imaged through
a telescope of 1.2-m diam. Two of the resulting 156
degraded images are shown in Figs. 9(b) and 9(c). The aperture. In practice the denominator of this expres-
degraded images were then processed by Labeyrie's 2 0 sion would be obtained by making measurements on a
method as modified by Goodman and Belsher.2 ' The reference star through an atmosphere having the same
estimate of the modulus of the Fourier transform of the statistics as that which blurred the images or by using
object is given by' 8 a model of the effects of atmospheric turbulence. W(u)
M 1/2 was included to restore the natural MTF due to the
Ip(u)I = W(u) = (45) telescope aperture which was removed by the denomi-
nator of the equation above. Figure 9(d) shows the
E ISm(U)12 resulting Fourier modulus estimate.
M=M+l
The object was reconstructed (or equivalently, the
where Im (u) is the Fourier transform of the mth de- Fourier phase was retrieved) using the hybrid input-
graded image, Np is the total number of photons de- output algorithm alternately with the error-reduction
tected (it is subtracted to compensate for a certain noise algorithm. The result, shown in Fig. 9(e), agrees very
bias term that arises in the power spectrum due to well with the original object shown in Fig. 9(a), despite
photon noise 2 l), S () is the Fourier transform of the the noise present in the Fourier modulus data. Good
mth point spread function (to provide compensation for reconstructed images were also obtained when only
the MTF of the speckle interferometry process), and the one-tenth as many photons were assumed to be avail-
weighting factor W(u) is the MTF due to the telescope able.' 8
Figure 10 shows E 0 vs the number of iterations for
this reconstruction. The starting input used was the
randomized demagnified autocorrelation described in
Sec. VII, using a threshold value of 0.004 times the peak
of the autocorrelation. For the first ten iterations the
error-reduction algorithm was used, and the mask de-
fining the diameter constraint was chosen to be the re-
gion over which the autocorrelation function, spatially
demagnified by a factor of 2, exceeded 0.004 of its
maximum value (providing a fairly tight diameter
constraint). For iterations 11-20 the error-reduction
algorithm was used, and the mask for these and the re-
maining iterations was chosen to be a square of length
64 pixels, which is larger than the actual object extent
of -60 X 40 pixels (imbedded in an array of 128 X 128
pixels). The error decreased suddenly at the tenth it-
eration since some positive-valued points that were
inside the second mask but outside the first mask were
no longer counted as contributing to E. By the
twentieth iteration the error-reduction algorithm was
converging very slowly. For iterations 21-60 the hybrid
input-output algorithm, with equal to one, was used.
At first E 0 increased sharply (although the output image
Fig. 9. Image reconstruction experiments. (a) Undegraded object; appeared no worse than at iteration 20) but then de-
(b), (c) examples of degraded images simulated to include the effects creased fairly rapidly until stagnating at E 0 o 0.05 at
of atmospheric turbulence and photon noise; (d) Fourier modulus about iteration 55. For iterations 61-70 the error-
estimate computed from the degraded images; (e)image reconstructed reduction algorithm was used, for which E0 dropped
using the iterative algorithm. suddenly from 0.05 to 0.02, although the visual ap-

1 August 1982 / Vol. 21, No. 15 / APPLIED OPTICS 2767

pearance of the reconstructed image remained the same troversy surrounding this question, both for the prob-
as for iteration number 60. lem of two intensity measurements 22 2 ' 2 3 and for the
This final value of Eo is comparable to 0.03, the nor- problem of a single intensity measurement.24-2 8 For
malized rms error of the Fourier modulus estimate it- the latter problem, the experimental reconstruction
self. 18 It is impossible to reduce E 0 to zero since the results shown here and elsewhere 4 5 10 9 suggest that for
noise in the Fourier modulus estimate results in an in- complicated 2-D objects the solution is usually unique
consistency of the non-negativity constraint and the in practice, even when the Fourier modulus data are
Fourier modulus estimate. This inconsistency can be corrupted by a considerable amount of noise.1 9
seen from the fact that the autocorrelation estimate
computed from the Fourier modulus estimate has areas This research was supported by the Air Force Office
of negative value. of Scientific Research under contract F49620-80-C-
Reconstruction experiments do not always proceed 0006.
as smoothly as the one described above. When stag- Appendix
nation occurs before E 0 is reduced to a level consistent
with the error in the Fourier modulus data, often the In this Appendix the relationship between the
best strategy is to restart the algorithm using a different steepest-descent method for the problem of two in-
set of random numbers as the initial input. One must tensity measurements and the Gerchberg-Saxton al-
also be careful in the way one chooses the mask as dis- gorithm is explored.
cussed in Sec. VII. Finally, it is always advisable to For the problem of two intensity measurements, the
reconstruct the image two or three times from the same error B is minimized with respect to the phase estimates
Fourier modulus data using different starting inputs 0(x). The partial derivative of B with respect to the
each time. If the same image is reconstructed each value of a given phase 0(x) is
time, one would have confidence that the solution is OB alG(u)I
a0B= =2N-2 E [IG(u)l-IF(u)] 00(x) (Al)
unique.
Since
IX. Comments and Conclusions
OG(u) 0
For the problem of phase retrieval from a single-in-
00(x) = 0()a E f(y)I exp[i0(y)] exp[-i27ru y/N]
tensity measurement it has been shown that the error-
reduction algorithm (the Gerchberg-Saxton algorithm = ilf(x)l exp[i0(x)] exp[-i27ru *x/NJ
applied to this problem) is equivalent to the steepest- = ig(x) exp[-i2iru *x/N], (A2)
descent method with a double-length step. Further- then
more, it was shown that the error-reduction algorithm a|G(u)l G(u)(-i)g*(x) exp[i2ru *x/N] +c.c.
(A3)
converges in the sense that the error monotonically 00(x) 21G(u)l
decreases. However, in practice the error-reduction where c.c. indicates the complex conjugate of what
algorithm converges very slowly; several other algo- precedes it. Using Eq. (27), Eq. (Al) becomes 2
rithms converge much faster, including other gradient
search algorithms which utilize the method of Fourier 09B = ig*(x)[g'(x) -g(x)] + c.c.
transforms for rapidly computing the gradient of the = ig*(x)g'(x) + c.c. = -2 Im[g*(x)g'(x)]
error and the input-output family of iterative Fourier = -21f(x)I Ig'(x)Isin[0'(x) - 0(x)], (A4)
transform algorithms. Of the algorithms investigated
so far, the hybrid input-output algorithm has converged where 0'(x) is the phase of g'(x). Therefore, the gra-
the fastest. Nevertheless, the gradient-search algo- dient can be easily computed using two Fourier trans-
rithms also converge reasonably fast, and they deserve forms (the first three steps of the Gerchberg-Saxton
further development. The performance of any given algorithm) plus evaluating Eq. (A4). Analogous to the
algorithm can vary depending on the stage of recon- linear approximation of Eq. (29) is
struction and on what is being reconstructed. Although B Bk + Y 3oBk[0(x) - Ok(X)I, (A5)
a practical phase-retrieval algorithm is in hand, there
is still an element of trial and error involved. where OoBk is 60B evaluated at 0(x) = Ok(X), the phase
For the problem of phase retrieval from two intensity of gk (x). This linear expansion of B is equal to zero
measurements, the steepest-descent method and the for
Gerchberg-Saxton algorithm were compared and were
found to share some similarities. When manipulating O( ) o ( ) Bkl)OBk (A6)
just the object's phase, the object-domain constraint, E ,ok)
that the object's modulus equals the measured modulus,
is automatically satisfied. Therefore, the gradient- which can be easily verified by inserting Eq. (A6) into
search methods seem inherently better suited to the Eq. (AS). By Eq. (A4),
problem of two intensity measurements than to the
problem of a single intensity measurement. Fi (aoBk) 2 = 4 Z If(y)121g(y) 2
Not considered here has been the question of the 2
uniqueness of solutions to either of the two phase-re- X sin [0(y) - k(Y)], (A7)
trieval problems. There has been considerable con- and so Eq. (A6) becomes

2768 APPLIED OPTICS / Vol. 21, No. 15 / 1 August 1982

 (X) - Ok(x) =- kf(x)g(x)I sin0'(x) - Ok(x)I That is, when the Gerchberg-Saxton algorithm is con-
2 2 (A8)
2 E f(y)1 1g(y)12 sin [,(y) - k(Y)] verging slowly, the steepest-descent method indicates
that a much larger step size be used. Note that if one
Now consider the situation after several iterations have uses the double-length step steepest-descent method,
been performed and the error is decreasing slowly. the factor of 1/2 in Eqs. (A8), (A12), and (A16) is replaced
Then for most values of x, by unity.

sin[0 x)- Ok(X)I 0ok(X) - Ok(X)- (A9)

Then from Eq. (A8) it is seen that 6(x) - (x), which
is the change of Ok (x) indicated by the steepest-descent
method, is approximately proportional to 6'k(x) - Ok (x). References
Recall from Eq. (9) that 'k(x) is the new phase that 1. R. W. Gerchberg and W. 0. Saxton, Optik 35, 237 (1972).
would be given by the Gerchberg-Saxton algorithm. 2. W. 0. Saxton, Computer Techniques for Image Processingin
Electron Microscopy (Academic, New York, 1978).
Thus, for the problem of two intensity measurements 3. R. A. Gonsalves, J. Opt. Soc. Am. 66, 961 (1976).
the steepest-descent method gives a new phase that is 4. J. R. Fienup, Opt. Lett. 3, 27 (1978).
going in the same general direction as the Gerchberg- 5. J. R. Fienup, Opt. Eng. 18, 529 (1979).
Saxton algorithm but uses a different step size. For the 6. J. R. Fienup, Opt. Eng. 19, 297 (1980).
steepest-descent method the direction differs somewhat 7. J. R. Fienup, "Reconstruction and Synthesis Applications of an
from 6'k(x) - k (x) due to the weighting factor Iterative Algorithm," in Transformations in Optical Signal
If(x)gk(x) , which varies with x. Processing, W. T. Rhodes, J. R. Fienup, and B. E. A. Saleh, Eds.
To obtain some notion of the step size for the steep- (Society of Photo-Optical Instrumentation Engineers, Belling-
est-descent method, continue to consider the situation ham, Wash., 1982), to be published.
where the error is decreasing slowly and Eq. (A9) 8. P. M. Hirsch, J. A. Jordan, Jr., and L. B. Lesem, "Method of
applies. Then, using Eq. (16) one can approximate Making an Object-Dependent Diffuser," U.S. Patent 3,619,022
(9 Nov., 1971).
9. N. C. Gallagher and B. Liu, Appl. Opt. 12, 2328 (1973).
Bk = E | If(x)l exp[iOk(x) - Ig'(x)l expliO,(x)1I 2 10. J. R. Fienup, T. R. Crimmins, and W. Holsztynski, J. Opt. Soc.
x
Am. 72, 610 (1982).
= E 11f(X)1 2 + g'(X)12 - 2f(x)g(x)l CoS[0k() - Ok(X)]} 11. R. N. Bracewell, The FourierTransform and Its Applications
(McGraw-Hill, New York, 1965).
12. R. H. Boucher, Proc. Soc. Photo-Opt. Instrum. Eng. 231, 130
Z [If(X)l - lgk(X)1]2 + f(x)g(x)l (1980).
x x
13. M. T. Manry and J. K. Aggarwal, IEEE Trans. Circuits Syst.
X sin 2 [0'(x) - Ok(x)]. (A10) CAS-23, 185 (1976).
14. D. P. Feder, Appl. Opt. 2, 1209 (1963).
To further simplify things, consider the special case for 15. D. R. Buchele, Appl. Opt. 7, 2433 (1968).
which f (X) = fa, a constant independent of x. Fur- 16. B. R. Frieden and D. G. Currie, J. Opt. Soc. Am. 66, 1111 (1976)
thermore, if B is small, for most values of x, (Abstract).
17. J. R. Fienup, "Improved Synthesis and Computational Methods
I f(x)I - igk(x)iiI If(x)I. (All) for Computer-Generated Holograms," Ph.D. Thesis, Stanford
Under these circumstances, inserting Eqs. (A9) and University, May, 1975 (University Microfilms No. 75-25523),
Chap. 5.
(A10) into (A8) yields 18. G. B. Feldkamp and J. R. Fienup, Proc. Soc. Photo-Opt. Instrum.
Eng. 231, 84 (1980).
0(x-^x
W W 2-k + [0(X) (X)], (A12) 19. J. R. Fienup, "Fourier Modulus Image Construction," Report
RADC-TR-81-63 (1981).
where 20. A. Labeyrie, Astron. Astrophys. 6, 85 (1970); D. Y. Gezari, A.
Labeyrie, and R. V. Stachnik, Astrophys. J. Lett. 173, Li
Bk= [lf(X) - gk(X)]2, (A13) (1972).
x
21. J. W. Goodman and J. F. Belsher, Proc. Soc. Photo-Opt. Instrum.
j jg'(x)[0(x)-Ok(x)12 Eng. 75, 141 (1976).
k= (A14)
22. A.M. J. Huiser, A. J. J. Drenth, and H. A. Ferwerda, Optik 45,303
are the respective radial and tangential components of (1976); A. M. J. Huiser and H. A. Ferwerda, Optik 46, 407
the error Bk B + Bk. The radial component Br is (1976).
the object-domain squared error of the modulus of 23. A. J. Devaney and R. Chidlaw, J. Opt. Soc. Am. 68, 1352
(1978).
lg (x) I and is equal to E2 according to Eq. (14). When 24. Yu. M. Bruck and L. G. Sodin, Opt. Commun. 30, 304 (1979).
the error is decreasing slowly, 25. W. Lawton, Proc. Soc. Photo-Opt. Instrum. Eng. 231, 94
Bk Bk -Bk= -E <<E =B. (A15) (1980).
26. A. M. J. Huiser and P. Van Toorn, Opt. Lett. 5, 499 (1980).
Therefore, under these circumstances the coefficient 27. T. R. Crimmins and J. R. Fienup, J. Opt. Soc. Am. 71, 1026
of Eq. (A12), (1981).
28. J. R. Fienup, "Image Reconstruction for Stellar Interferometry,"
+ Bk 1 in CurrentTrends in Optics, F. T. Arecchi and F. R. Aussenegg,
Eds. (Taylor & Francis, London, 1981), pp. 95-102.

1 August 1982 / Vol. 21, No. 15 APPLIED OPTICS 2769