1434

J. Opt. Soc. Am./Vol. 73, No. 11/November 1983

Michael Reed Teague

Deterministic phase retrieval: a Green's function solution

Michael Reed Teague
The Charles Stark Draper Loboratory, Inc., 555 Technology Square, Cambridge, Massachusetts 02139
Received November 29, 1982; revised manuscript received May 31, 1983
Equations for the propagation of phase and irradiance are derived, and a Green's function solution for the phase
in terms of irradiance and perimeter phase values is given. A measurement scheme is discussed, and the results
of a numerical simulation are given. Both circular and slit pupils are considered. An appendix discusses the local
validity of the parabolic-wave equation based on the factorized Helmholtz equation approach to the RayleighSommerfeld and Fresnel diffraction theories. Expressions for the diffracted-wave field in the near-field region
are given.

1.

INTRODUCTION

The optical phase-retrieval problem, which is to deduce optical phase from minimal irradiance measurements by using
noninterferometric techniques, has been studied extensively
in recent years. Its literature has been repeatedly surveyed
and is easily traceable.'- 3
Recently 4 a scheme was developed that allowed optical
phase to be determined from measured irradiance moments
in multiple measurement planes (in particular, two planes
sufficed in principle) if the pupils were sufficiently "soft," i.e.,
not sharp. The present paper, which is based on Green's
functions, evolved from a study of methods to determine
phase from two-plane irradiance measurements without the
auxiliary device of introducing irradiance moments with their
concomitant mathematical existence problems associated with
sharp pupils. (Which of these two schemes would actually
perform best in terms of dynamic range, bandwidth, and signal-to-noise ratio in a given measurement situation can be
determined only by further analysis, simulation, and experimental studies.) Both methods are deterministic phaseretrieval schemes in the sense that the phase is given directly
in terms of the measured irradiance data, and the phase
uniqueness question does not arise here as it does in methods
based on iterative algorithmsl- 3 to retrieve phase. Previous
studies based on two-defocus 5' 6 measurement methods and

analyticity properties 7- 9 of the phase have also concluded that

additional irradiance measurements (multiple transversemeasurement planes or measurements with different phase
filters inserted in one plane) ensure that the phase that is recovered is unique.
In Section 2 the propagation equations for phase and irradiance are derived, and one of the equations furnishes a basis
for phase retrieval. A Green's function solution for the phase
is given in Section 3. A new feature is that for finite pupils
(i.e., non-Gaussian-like beams) a perimeter phase measurement near the actual pupil boundary is also necessary, and
two-plane irradiance measurements alone are insufficient to
determine the phase across the area of the pupil. In fact,
however, for a uniformly illuminated pupil, perimeter phase
measurements and irradiance measurements in one plane a
distance e6z from the pupil plane suffice to determine k.
0030-3941/83/111434-08$01.00

Moreover, for circularly symmetric aberrations, only onephase irradiance measurements are needed for phase retrieval;
the perimeter phase measurements are unnecessary. Irradiance measurements near the pupil plane are generally inconvenient and difficult from the viewpoint of noise considerations. Section 4 indicates how to make the necessary irradiance measurements in a more favorable plane by using a
beam compressor. The penultimate section gives results of
a numerical simulation of the phase-determination scheme.
The Green's function method of phase retrieval applies
locally anywhere and not just at the pupil, as discussed in the
last paragraph. In general, the phase inside a region bounded
by some perimeter in a transverse plane is determined by the
longitudinal gradient of the irradiance in the given plane and
by perimeter phase measurements. As is discussed in Section
3, in some cases, e.g., in the image plane, it is permissible to
extend the perimeter effectively to infinity. Then imageplane phase is determined only by the longitudinal irradiance
gradient at the image plane, and pupil-plane phase is then
given by inverse Fresnel transformation.1
This paper is based on the local validity on the parabolicwave equations for describing the diffracted-wave field. As
is discussed in an extensive appendix, this is not the same as
assuming the validity of Fresnel diffraction theory since the
parabolic-wave equation remains approximately valid locally
even when the Fresnel diffraction integral gives an inaccurate
description of the diffracted wave (e.g., in the near-field region
of a sharp pupil). The somewhat unconventional approach
to diffraction theory presented in Appendix A is especially
suited to near-field-region considerations.
2. PROPAGATION EQUATIONS FOR PHASE
AND IRRADIANCE
Suppose that light propagates nominally in the +z direction,
and let the time-dependent wave amplitude be written as
exp(-i27rct/A)u,(r), where r = (x, y) is a two-dimensional
vector in the transverse direction. Then, as is discussed extensively in Appendix A, the amplitude u, (r) satisfies approximately the parabolic equation'0 [see Eq. (A18) of Appendix A]
1983 Optical Society of America

Vol. 73, No. 11/November 1983/J. Opt. Soc. Am.

Michael Reed Teague

2k + k u,(r) = ,

(i d+

(1)

where V2 = [(a2 /ax2) + (a2/ay2)] and k = 2r/X. The normalization of u, (r) is such that
(2)
I.. (r) = Iu-,(r) II
is the irradiance at point (x, y, z).
The wave amplitude may be expressed in terms of the irradiance I and the phase 0, which are real-valued quantities,
i.e.,
uz(r) = [Iz(r)]1/ 2 exp[i-P,(r)].

(3)

is slightly inside the actual physical perimeter of the round

pupil]. We choose, for convenience, to have G satisfy Dirichlet boundary conditions; i.e.,

2\

I= -v

Iv(,

az

(4)

(5)
= 2 Iv2I - - (VI) 2 - I 2(vO) 2 + kI 2.
X z
2
4
Thus the real quantities o and I satisfy a coupled set of nonlinear equations. If both 0 and I are unknown, it is obviously
easier to solve the single linear Eq. (1) for the complex-valued
amplitude u. However, if the irradiance is known (e.g., by
measurement), then Eq. (4) becomes a linear equation, which
may be solved to obtain the phase 0.
On introduction of an auxiliary function 4', which satisfies
v'=

IvO,

(6)

(9)

v 2 G(r, r') = 6(r - r')

(10)

for r, r' e R. By using elementary considerations (and the

method of images"1 ), it is straightforward to construct G
satisfying Eqs. (9) and (10); it is
G(r, r)

- ln I r2Ir2
a

(11)

+ a -2r * r'

For convenience, we also quote

whereas, if they are added and the sum is multiplied by I, one

gets
Ir

G(r, r') = 0,
for r' on the perimeter.
In addition,

Let Eq. (1) be multiplied on the left-hand side by u,* and

the complex conjugate of Eq. (1) be multiplied on the lefthand side by u,. If the two resulting equations are subtracted,
one gets (suppressing the z subscript)
27r

1435

a 2 - r2
27ra \a2 + r2-2r .r'(

aG(r, r')
an'

where in Eq. (12), but not in Eq. (11), r' = ar' and r' is a unit
vector.]
Finally, we now specialize to the case that the irradiance is
constant (I = Io) over the pupil plane (z = 0). Then Eq. (6)
implies that
=

Io(5 + constant).

(13)

Using Eqs. (8)-(13), we may express the phase (over the

interior of a constant-irradiance, round region) as

00(r)

SSR dr'G(r, r') [-2

za

Eq. (4) becomes

V24' = 2

I,

+ f

(7)

dz

i.e., the two-dimensional Poisson equation." Once 4 is

known, k is determined (to within an additive constant) by
integrating Eq. (6).

ds'po(r') d, (r, r').

[The constant in Eq. (13) drops out of Eq. (14) because of the
identity

r ds' an'a G(r, r')= 1

jp

3. GREEN'S FUNCTION SOLUTION FOR THE

PHASE
The solution of Eq. (7) inside a two-dimensional region R,
which is perpendicular to the z axis and bounded by perimeter
P,is given by"
(- 2

X fR

dz

ds' [G(r, r') ai'(r'-

az
(r') a G(r,r'

(8)

where a/an' = n' - v', n' is the outward-pointing unit vector

normal to the perimeter at r', and G (r, r') is a Green's function
that depends on the shape of the region R.
For now we consider the special case of a circular region R
of radius a, with xy-coordinate origin at the center of the circle
[e.g., round pupil centered on the optical axis. However, for
reasons discussed below, the perimeter to be used in Eq. (8)

(15)

for any r, as is straightforward' 2 to verify from Eq. (12).]

Equation (14) indicates that the phase 00(r) at the pupil
plane is determined by the value of 00(r) on the perimeter P
and by the value of the longitudinal gradient of the irradiance
over the area of the pupil. In fact, since the gradient must be
approximated by measured values

aI0 (r')

d r'G (r, r')-Iz (r')

(14)

an'

I& (r')

Io

6Z

and since I0 is a (known) constant, irradiance measurements

need to be made only in the single plane z = 6z. In subsequent sections we discuss how to make this measurement and
how to choose bz.
The perimeter phase measurements to be used in Eq. (14)
determine only the relative phase of points on the perimeter,
and the relative phase between perimeter points and interior
points can be indeterminate to within an overall unknown
constant. Because of Eq. (15) this affects 0 as determined by

1436

J. Opt. Soc. Am./Vol. 73, No. 11/November 1983

Michael Reed Teague

Eq. (14) only by the same unknown constant (and hence is a

trivial piston phase error).
In some cases it is permissible to extend the radius of the
perimeter to infinity (e.g., with an untruncated Gaussian
beam). Then the Green's function becomes G(r, r') = (1/
27r)ln(Ir - r'J/A), where A is a constant length and va is independent of the particular choice of A. Moreover, the perimeter integrals in Eq. (8) do not appear; thus two-plane irradiance measurements alone suffice to determine 0.
(For reference purposes we quote here the analogous results
for an infinite slit aperture over which the irradiance is uniform. Again +z is the propagation direction, y is a transverse
coordinate parallel to the two parallel slit edges, which are
separated by distance 2a, and x is a transverse coordinate
perpendicular to the slit edges. The coordinates' origin is at
the center of the slit, and it is assumed that the wave amplitude is independent of y.) Then, analogous to Eq. (14), we
have

I i

zUO
PUPIL
-PLANE

- F

OUTPUT
PLANE

Fig. 1. Beam compressor. Output plane is located at Z2 = Z1 + (1

+ f/F)f. The output-plane wave amplitude is u,2 (r) = (-1/m)uo
(-r/m), where m = f/F.

00(x)

dx'G (x, x')

+

iow)l

2a ) o(a)

-2a)o(-a),

(16)

Hereafter, let the xy-coordinate system at Z2 be rotated 180

about the z axis. Then
I. 2(r) =

where the Green's function is

2Io(r/m)

(21)

and

G(x, x') = Ix - xI- I |-I - al

(17)

and satisfies

,z2(r) = ko(r/m) + 7r,

and, using Eq. (4), which applies both to planes z = 0 and z =

Z2 ,

d2 G(x, x') = 6(x-x%

G(x, a) = G(x,-a) = 0.

4.

(22)

(19)

'Z2

MEASUREMENT SCHEME

In this paper we assume that there is no basic problem in

making the necessary perimeter phase measurements (e.g.,
by using a set of Hartmann13 sensors). We also assume that
the light is nominally collimated at the pupil. Then 00(r) is
the pupil-plane phase aberration. In general it is difficult to
measure dIo/az by subtracting irradiance values at points near
the pupil plane because [see (Eq. (4)] 61o/Io - (Xbz/a 2) X
(number of waves of aberration). For example, for visible
light and large pupils (a - 1 m), 6Io/'o - 10-6 ( 6z/m). The
longitudinal irradiance variation can be greatly increased by
using a beam compressor (see Fig. 1) with, e.g., two converging
lenses with focal lengths F and f, respectively. The magnification is m = f/F << 1 for our case. (In fact, the first lens will
usually already be part of the optical system, and 0o is the
phase aberration just before the lens plus the aberration introduced by the lens.)
We treat the beam compressor as a linear filter, assume
paraxial optics and thin, aberration-free lenses, and ignore
finite-aperture effects. If the output plane Z2 is chosen so
that
Z2= Z1 + (1 + m)f z, +f,
then one can show, by using standard techniques,' that
,

1 , ,
uo(-r/m).
m

Thus 3 1Z2/1z2

aI'

2 (r)

(20)

1 1OIo(r/m)

m2 Jo

Oz

(23)

Oz
2 2
(Xz/m a ) X (number of waves of aberration)

and, for a high-ratio compressor (e.g., m z 10-3), aiz2 /Iz2 GOz/1 meter), which is readily measurable. Equation (23) is
then used to obtain, for use in Eq. (14), (1/o)dIo/Oz from
measured data. (Incidentally, the beam-compressor technique makes feasible the two-plane irradiance-moment
method 4 of pupil-plane phase retrieval.)
5. NOISE CONSIDERATIONS AND CHOICE
OF 6z
Near the plane z

= Z2 , the

irradiance may be written as

I, 2+6z(r) = 1Z2[1 + bibz + b 2 (6Z) 2 J + AIl 2,

(24)

where
b= -2 (

V2 2

(25)

+ ofixyy + 0Y yyy + (OYY)2]z =z2

(26)

and
b2 = 4 (-)2 [(XX)2 + XXXXx

uZ2 (r) = - -

we have

(18)

In Eq. (26) oxx = (a /1x )0, etc. Equations (24)-(26) are

derived by using Eq. (A32) of Appendix A to second order in
6Z AIZ2 is a noise term. Assuming the most favorable case
of photon-noise-limited detection (Poisson-detection statistics) and unity quantum efficiency, one can show1 5 that AIZ 2

1437

Vol. 73, No. 11/November 1983/J. Opt. Soc. Am.

Michael Reed Teague

has zero mean and standard deviation:

0.04

(A1

)rms = 1z2

Z2/N

COMA

(27)
0

where Ad is the detector element area and N2 2 is the total

number of photons counted in area Ad in time At. [Since
IZ2+bz differs from IZ2 by only a few per cent, one can use Z 2
rather than Z2 + bZ for estimation purposes in Eq. (27).] The
measured quantity is
'Z2+6Z

Z2

6ZeZ2

= bi + b 2 6ez + AZ 22
5ZZ2

-e

Wo
Pn

_... :

-- ,

-,

-6.;!"' ,

.0

X --

-O.DI

(28)

-0.OZ~

5z must be small enough that the first term dominates the

second:

OO

6z << I bl/b2i,

-D.04

and large enough that the noise term is small:

(29)

6z.

lb

-0.0 j

[One can show that, for fixed b 2 and NZ2 , the optimal, in the

sense of minimizing the last two terms on the right-hand side

of Eq. (28), 6z is I b21- 112 NZ2-1/4.]
6.

Fig. 2. Phase error for one wave of coma. Both here and in Fig. 3
a plot of b0(x, 0) = kactual(X, 0) - 4kcalculated(X, 0) is given. The solid
curve has simulated noise added to the calculation. The dotted curve
is the noise-free calculation and is nonzero because the number of
points used in the numerical integrations was finite. The pupil radius
is 1.0.

NUMERICAL SIMULATION
in

For simulation purposes the following system parameters were

chosen arbitrarily: F = 3 m, a = 1 m, f = 0.93 mm (thus m =
3.1 X 10-4), X = 10-6 m, Ad = (20 /um) 2. Then the detector
array at Z2 has 31 X 31 elements. The demagnified round
pupil, with a diameter of 610 Am, then just fits as a circle inscribed inside the 610-ym X 610-Mm square detector array.
This array size was chosen empirically as this mesh size
guaranteed that the phase calculated by Eq. (14) agreed with
the known phase aberration to within 1/20 wave when there was
no added noise. The area integrals were performed by a 31X 31-point 2-d Simpson's rule. For the perimeter integrals
a 93-point Simpson's rule was used. The Green's function
and its normal derivative have integrable singularities that
slow the convergence of the numerical integration. An ad hoc
procedure of drawing a circle of radius 0.001 (relative to a pupil
radius of 1.0) around the singularities and setting the integrand to zero inside the small circle was adopted. A larger
keep-out radius speeds convergence but to an inaccurate answer.
Gaussian random noise of zero mean and 1/20-wave standard
deviation was added to the perimeter phase values. Gaussian
random noise of zero mean and standard deviation corresponding to AI/I = 1/1 (i.e., 100 photons/detector cell) was
added to the area integrations.
Figures 2 and 3 show results of the calculations for one
wave of coma and one wave of spherical aberration, respectively. The quantity plotted is
Ak0(X, 0) = O(X,

O)actual -

k(X,

O)calculated-

For these conditions, 3z - 2 mm allowed inequalities (25) and

(26) to be met.
The purpose of the numerical simulation was to verify the
correctness of Eqs. (11), (12), and (14) and their relative noise
immunity. No attempt was intended to optimize a system
design, e.g., in terms of minimum array size or maximum
tolerable noise.

SPHERICAL ABERRATION

-0.0

-0.0

10.0

0.4

-0

c0

00

0.4

0r.6

0.01

1.0

Fig. 3. Phase error for one wave of spherical aberration.

In all cases for which numerical calculations were made, the
area integral and the perimeter integral in Eq. (14) produced
results of the same order of magnitude; thus neither can be
neglected for a finite pupil {unless it is known that the aberration is circularly symmetric, in which case the phase is
constant on the line integral and only the area integral need
be calculated [see Eq. (15)]}.

7.

CONCLUDING COMMENTS

The approach to phase retrieval presented in this paper allows

optical phase to be determined uniquely from longitudinal
irradiance gradients and possibly also perimeter phase measurements. The solution presented involves Green's functions, and a numerical simulation showed that adequate accuracy is attained with a relatively small detector array.
Moreover, the algorithm is not particularly sensitive to noise.
If pupil-plane phase is desired, then irradiance measurements
must be made at a demagnified version of the pupil by using

1438

J. Opt. Soc. Am./Vol. 73, No. 11/November 1983

Michael Reed Teague

a beam compressor. This technique would not be necessary

if measurements were made at the image plane, but then the
scheme of this paper will give the optical phase at the image
plane.

whore k, = 27r/X, and the square-root operator is defined

ultimately (as discussed in detail in Section A.3) in terms of
the Fourier transform of Eq. (A6), i.e.,
U,(p) = exp[ikz(1 - X2 p2 )1/2 ]Uo(p),

(A7)

APPENDIX A. EXACT DIFFRACTION

THEORY, APPROXIMATE VALIDITY OF THE
PARABOLIC-WAVE EQUATION, AND NEARFIELD CONSIDERATIONS

where the two-dimensional Fourier transform is generally

defined by

In this Appendix it is shown in Section A.2 that, while the

Fresnel diffraction integral and the parabolic-wave equation
imply each other, the parabolic-wave equation is, locally,
approximately valid even when the global Fresnel diffraction
integral produces absurd results (e.g., in the near-field diffraction region of a sharp pupil). In Section A.3, expressions
describing the diffracted-wave field of aberrated pupils in the
near-field region are given. The results of Sections A.2 and
A.3 follow from scalar diffraction theory as summarized in
Section A.1 starting with the factorized wave equation, which
has long been known in the physics literature' 6 -18 and has
been used increasingly for treating problems in optics19 2 1and
acoustics. 2 2' 23 The reader is reminded that the methods of
the factorized wave equation (which leads to the diffraction
operator approach1 9 -21), the angular spectrum,' 4' 24 and the
Rayleigh-Sommerfeld diffraction theoryll""4' 25 are all
equivalent formulations of diffraction theory.

It is, however, known 26' 27 that the inverse Fourier-transform

relationship

A.1. Rayleigh-Sommerfeld versus Fresnel Diffraction

Theories
The time-independent wave equation in empty space
[az2 +

V2+ (27r 2]

(r) =

(Al)

[recall that V2 = ax 2 + ay2 and r = (x, y)] that describes waves

traveling in all directions may be written as
L+L-T'z(r) = 0,

(A2)

where

L = a
az

i22F [1 + (XV1211/2

\2irJ]

FT'ljexp[ikz(I - X2p 2 )1/2]j = -

dz R
2,7r az R

(A3)

(A4)

(A8)

(A9)

holds, where
R-

(Z 2 + r 2 )1/2.

Therefore, from Eqs. (A7) and (A9), we have

u,-,(r) =- 1- -a [u0(r)

eikR
* *-_I,
~ Rj

27r-az I

(A10)

which is just the usual Rayleigh-Sommerfeld diffraction

theory result,11"14' 25 which gives the wave amplitude in a
transverse plane z > 0 in terms of the wave amplitude in an
earlier plane z = 0. Notice that ** denotes two-dimensional
convolution, i.e., for two functions f and g:
f(r) ** g(r) --

dr'f(r')g(r - r').

(All)

The conventional Fresnel diffraction theory results when

the square roots in Eqs. (A6) and (A7) are expanded to lowest
order to get

uz,,(r) = exp(ikz)exp [(i

)] uo(r)

(A12)

(A13)

and
UZ,F(p) = exp(ikz)exp(-i7rXzp

The solutions of Eq. (Al) thus separate into two classes:

L+u (r) = 0

Uz(p) = f dr exp(-i27rp- r)u2 (r) - FT[uz(r)].

)Uo(p),

and, using the elementary (inverse) Fourier-transform relationship,

FT-1[exp(-ivrXzp 2)]

=exp(i-xr 2 /Xz)

(A14)

iXZ
L-.v (r) = 0.

(A5)

(Choose z = 0 to be the initial plane.) The solutions JuzI describe either oscillatory waves with a positive z component of
the wave vector 27r[p, (1/A)(I - X2p2)1/2] or evanescent waves
if 2p 2 > 1 and z > 0, whereas the solutions lv,} describe either
oscillatory waves with a negative z component of the wave
vector 2 7r[p, -(l/X)(1 - X2 p2 )1/2 ] or evanescent waves if X2 p2

> 1 and z < 0. In the absence of scattering by charged matter,

the two solution classes do not mix. Hereafter, we consider
only the solutions {u,}.
A formal solution of Eq. (A4) may be written as
u, W = exp

ikz (1 +

k-2] uo(r),

(A6)

we have from Eq. (A12)

uZ,F(r)

exp(ikz)

uo(r) **

exp(irr2 /X_11

(A15)

iZ z~]

which is the usual Fresnel diffraction theory result' 4 (with the

obliquity factor taken equal to unity).
One can see immediately that there are cases in which the
Rayleigh-Sommerfeld and Fresnel diffraction theories predict
essentially the same diffracted-wave field for any z ' 0. For
example, if the spectrum UO(p) of the initial plane-wave amplitude is essentially zero for IPI > E/X (where E is a small
number, say, '/1o), then Eq. (A7) is effectively the same as Eq.
(A13), and both theories predict essentially the same diffracted wave. This situation occurs characteristically when
the initial plane has a smooth wave-amplitude distribution

Michael Reed Teague

Vol. 73, No. 11/November 1983/J. Opt. Soc. Am.

with a transverse width >>X and explains why the Fresnel

diffraction theory accurately describes the propagation of a
Gaussian beam for any z > 0.
On the other hand, it is characteristic of the two theories
to predict significant differences in the near-field region when
the initial plane has a sharp pupil. This can be seen immediately for a round pupil of radius a, normally illuminated by
collimated, monochromatic light. For axial observation
points (r = 0), it is straightforward to evaluate analytically
both Eqs. (A10) and (A15) to get
2
+ a 2 )1/2
Uz (O)= exp (ikz) -zepi(Z
Uz()

=p(

(A16)

2 + a 2)(1/2

J,(2 rap)

(2rap)
and, for p as large as

U0

A.2. Local Validity of the Parabolic Equation

From Eq. (A15) it may be verified that the Fresnel result
satisfies exactly the parabolic equation' 0

ad+ 2k + k) uZF(r) =0

(A18)

[which is also Eq. (A4) with the square root expanded to lowegt,
nontrivial order]. In fact, Eq. (A18) has a greater region of
validity than the integral result of Eq. (A15) in the sense that
locally, away from sharp pupil edges, the correct (RayleighSommerfeld) wave field satisfies approximately the parabolic
equation. To see this, we write [using Eqs. (A7) and (A8)]
u,(r)

r dp exp(i2wrp *r)exp[ikz(1 - X p )1/ ]Uo(p)

2 2

and apply the operator [i(d/dz) + (v 2/hk)

of this equation to get

.a V2

(id + v

(A19)
+ k] to both sides

+ k) u, (r)

= Jdp exp(i27rp

X Uo(p)[-k(1

- r)exp[ikz(1 -

X2p2)1/2 -

X2p2)

2
rXp + k].

(A20)

In general, it is permissible to expand, in powers of p 2 , the

second expression inside square brackets in Eq.(A20), and
moreover this is not equivalent to neglecting evanescent waves
since the square root in the exponential exp[ikz (1 - X2p2)1/2],
to which the integral in Eq. (A20) is much more sensitive, is
not expanded. For example, for a round pupil of radius a,
normally illuminated, the initial plane spectrum is proportional to

()

(A21)

U(0)

()

assuming a >> X. To lowest nonvanishing order, a term proportional to p 4 remains inside the second set of square
brackets in Eq. (A20), and this may be brought outside the
integral as an operator proportional to V 4 if the point (r, z)
is not located at a pupil edge so that Eq. (A20) becomes

(A17)

[In arriving at Eq. (A16} it is best to use R' = (Z2 + r'2 )1/ 2
rather than r' as an integration variable in Eq. (A10).
Moreover, it appears generally true in evaluating Eq. (A10)
analytically that it is better to differentiate with respect to z
after the convolution is performed.] For z >> a, both theories
predict the same result. [A more precise condition for the
agreement of Eqs. (A16) and (A17) is that zia >> (a/X)1/ 3 be
satisfied, as discussed on p. 59 of Ref. 14.] Significant differences occur when z a, and for z << a Eq. (A17) predicts
longitudinal oscillations in the wave field over distances even
smaller than X, which is an unphysical result. Equation
(A16), however, correctly reproduces the initial field as z
0.

U0(p) has already fallen to the small

1/3 X,

value

and

Uz,,F() = exp(ikz) -exp Jik (z + Ia2-)1 -

1439

( dz

V2

8z

2k

8 k 1 u r ;-

(A22)

[We note that Eq. (A22) is formally the same as Eq. (A4) with
the square-root operator expanded to terms of order V 4 ;
however, a straightforward expansion of Eq. (A4) gives no
indication of the connection between the accuracy of a truncated expansion and the initial plane spectrum UO(p).]
We now consider, for several cases, the order of magnitude
of the ratio of the term involving V4 in Eq. (A22) to the term
involving v2 , i.e., we look at
1 V4U, (r)
2 u-

X742 12

4k V uZ(r)

as follows:
(1) Near the pupil, but away from pupil edges, for nominally collimated light
42

[2 X (number of waves of aberrationY2j,

where a is the pupil radius.

(2) Near the pupil, away from pupil edges, for light
nominally focused at z = f,
7742 t 0 I(f#)2J

wheref# =f/2a.
(3) In a focused system, near the image plane,
[r2/X21
0I.
;
fl42O

In these cases we see that 7742 is small and hence the parabolic
equation is approximately valid locally, away from pupil edges,
if a >> Xand, for focused system, if f # >> 1. These conclusions are consistent with results recently reported, 27 which
were based on direct numerical calculations comparing the
Fresnel and Rayleigh-Sommerfeld theories.
A.3. Diffraction Operators and Near-Field
Considerations
Equations (A6) and (A12) give, respectively, for the Rayleigh-Sommerfeld and Fresnel diffraction theories, a diffraction-operator solution of the wave equation. In this
subsection we consider how these operators are defined and
use the results to investigate the near-field behavior of the
diffracted-wave field.

1440

J. Opt. Soc. Am./Vol. 73, No. 11/November 1983

Michael Reed Teague

It is convenient to consider first the mathematically simpler

Fresnel theory. It is natural to attempt to define the exponential operator appearing in Eq. (A12) by its power-series
expansion, i.e.,
(1
exp (i "Z V2) uo(r) -

1(iXzV

X jfdp'Po(p')Fo(P- p')

\N

uo(r).

(A23)

at r and the series in Eq. (A23) converges. If uo(r) describes

a round sharp pupil of radius a, normally illuminated, Eq.
(A23) gives, for any z > 0, zero if IrI < a or IrI > a and an indeterminate (and infinite) result JrI = a. So Eq. (A23) produces an absurd result if the initial plane contains sharp pupils. The fact that a power-series representation of Eq. (A12)
generally fails to be valid is not surprising since any operator
involving effectively an infinite number of derivatives is actually a nonlocal operator and, in fact, one correct representation of Eq. (A12) is just the familiar Fresnel-diffraction
result of Eq. (A15).
Hereafter we shall always consider only initial wave distributions uo(r) that have a Fourier transform [see Eq. (A8)]
Uo(p). Then the Fresnel-diffraction operator can always be
defined by the first equality of the equation
2

Iuo(r) -

X (-

iwXzp

exp (i

47w

) fo(r - Xzp') = exp i - V12

2 EZ
/r
4(
/ M=O N=O
X f(-n)r)

(t, 1) and fo(mn)(r)

where p
finally,

uz (r) =

exp (ikz)

(A24)

m-o n=O

-,Z

)m-(_XZn) nI

= 8m/Ox man/aynfo(r).

-exp
I l ti
m!n!

(-

X{(-Zd)~

X.2
47r

(A27)
Thus,

mwn(r)]
fo

Azd_)n fdpPo(p')

(A24')

)NUo(p).

X exp(i2rp' *r)exp(-irXzp'2)J.

(A28)

Equation (A28) is an exact result of Fresnel diffraction theory,

valid for all z > 0. The expression in square brackets describes the effect of aberrations, and the expression in braces
involves only the aberration-free diffracted-wave field. In
the near field (lowest order in z), we have
uz(r) - pz,F(r)

[exp (i X v2)zf(r)z,

(A29)

r <a

where Pz,F(r) is the Fresnel diffracted-wave field of the aberration-free pupil.

In the case of the exact Rayleigh-Sommerfeld diffraction
theory, the effects of the sharp pupil and the smoothly varying
aberrations can be separated only approximately. If we start
with Eqs. (A6) and (A25) and follow steps analogous to Eq.
(A26) through Eq. (A29), problems arise in the third equality
of the equation analogous to Eq. (A26). One can proceed with
the derivation only if the spectra Po(p) and Fo(p) are such that
the approximation

otherwise

[1-

po(r)fo(r),

(A25)

where po(r) describes a sharp pupil and necessitates the use

of Eq. (A24) while fo(r) describes a smooth function and
ultimately follows Eq. (A23) to be used operating on fo(r).
For example, we might have
po(r) = lo

(A26)

where the third equality of Eq. (A26) results from the second
equality on introduction of dummy variable p" = p - p'. The
expression inside the braces may be written as

dp' exp[i2rp" - (r - Xzp')]

X exp(-i7rXzpi2)Fo(pii)},

Whereas Eq. (A24) is always valid for physically realizable

cases, e.g., sharp pupils or smooth pupils, Eq. (A24'), if valid,
requires each term in the series to be finite and also the serieE
to converge. This will occur if, e.g., as p
, Uo(p) - I
0[exp(-const. pfn)] for n = 1, 2, 3, . .. , and this implies that
uo(r) is a smooth, e.g., Gaussian-like, pupil. We do not wish I
to limit our considerations to such smooth pupils, however,I
and in the following paragraphs extend the results to moregeneral pupils.
We now consider the near-field behavior of the Fresnel
diffracted wave. We express the initial wave as
uo(r)

3'dp'PO(p')exp(i2rp' - r)
X If

fN dp exp(i27rp- r)
2

dp'Po(p') J dp exp(i2irp - r)

X exp(-iwrXzp' 2 )

dp exp(i27rp -r)
X exp(- iwrXZp 2)Uo(p)

X exp(-irXzp 2 )Fo(p - p')

Obviously this will work only if u 0 (r) is infinitely differentiable

expzi 4Zv

dp exp(i27rp - r)exp(-i7-Xzp 2 )

exp (i XV) uo(r) =

while

(p' + p") 2 ]1/ 2

(1p-2)1/2
- \

fo(r) = exp[ikw(r)]exp(-cr2 ),
where w(r) is the pupil-plane wave-front aberration and e is
a very small (ultimately set equal to zero) convergence parameter used to ensure that the Fourier transform of fo exists.
Then from Eq. (A24) and (A25) we have

p'.P
p" + (1 - X2 p"2 )1/2

(A30)

is valid. For small aberrations, Fo(p) is sharply peaked about

IpI = 0 while Po(p) has properties typical of Eq. (A21). Thus
Eq. (A30) is a reasonable approximation and will not lead to
the type of physical absurdity in the near-field region characteristic of Eq. (A17). With Eq. (A30) one gets as an approximation to the Rayleigh-Sommerfeld diffracted wave:

Michael Reed Teague

axX 1

_r E -

2 ro n-0

m'n!

XZ)m

X exp[ikz(1 -

(_

Vol. 73, No. 11/November 1983/J. Opt. Soc. Am.

exp[ikz(1 + v2/k2 )1/2]fo(mn)(r)}

A

d nbdPoPo(p')exp(i27rP. *r)

X2p2)1/2]}exp(ikz).

(A31)

The first expression in braces describes the effects of aberrations, and here the diffraction operator may be expanded
in a power series. The second factor in braces now involves
the Rayleigh-Sommerfeld diffracted wave from an aberration-free pupil. For the near-field region,

uz (r) - pz (r) 1exp (i X v 2)fo(r)j'

(A32)

where P (r) is now the wave predicted by Rayleigh-Sommerfeld theory for an aberration-free pupil.

ACKNOWLEDGMENT
The author expresses his gratitude to V. N. Mahajan for
helpful discussions regarding this paper.
A preliminary version of this paper was presented as a
postdeadline paper at the Topical Meeting on Signal Recovery
and Synthesis with Incomplete Information and Partial
Constraints, January 12-14, 1983, Incline Village, Nevada.

REFERENCES
1. J. R. Fienup, "Phase retrieval algorithms: a comparison," Appl.
Opt. 21, 2758-2769 (1982).
2. R. H. T. Bates and W. R. Fright, "Reconstructing images from
their Fourier intensities," in Signal and Image Reconstruction
from Incomplete Data. Theory &Experiment, Vol. I of Advances in Computer Vision and Image Processing, T. S.Huang,
ed. (JAI, Greenwich, Conn., 1983), Chap. 5.
3. H. P. Baltes, ed., Inverse Source Problems in Optics (SpringerVerlag, New York, 1978); Inverse Scattering Problems in Optics
(Springer-Verlag, New York, 1980).
4. M. R. Teague, "Irradiance moments: their propagation and use
for unique retrieval of phase," J. Opt. Soc. Am. 72, 1199-1209
(1982).

5. B.J. Hoenders, "On the solution of the phase retrieval problem,"

J. Math. Phys. 16, 1719-1725 (1975).

6. A. M. Huiser and H. A. Ferweda, "The problem of phase retrieval

1441

in light and electron microscopy of strong objects II. On the

uniqueness and stability of object reconstruction procedures using
two defocused images," Opt. Acta 23, 445-456 (1976).
7. J. G. Walker, "The phase retrieval problem: a solution based on
zero location by exponential apodization," Opt. Acta 28, 735-738
(1981).
8. J. M. Wood, M. A. Fiddy, and R. E. Burge, "Phase retrieval using
two intensity measurements in the complex plane," Opt. Lett.
6, 514-516 (1981).
9. P. Kiedron, "Phase retrieval methods using differential filters,"
Proc. Soc. Photo-Opt. Instrum. Eng. 413 (to be published).
10. A. Papoulis, Systems and Transforms with Applications in
Optics (McGraw-Hill, New York, 1918). The parabolic equation
is treated in problem 9.9.
11. J. D. Jackson, ClassicalElectrodynamics, 2nd ed. (Wiley, New
York, 1975), Chaps. I, II, IX.
12. I. S. Gradsteyn and I. M. Ryzhik, Tables of Integrals,Series and
Products (Academic, New York, 1965), formula (2.553), no. 3.
13. J. W. Hardy, "Active optics: a new technology for the control of
light," Proc. IEEE 66, 651-697 (1978).
14. J. W. Goodman, Introduction to FourierOptics (McGraw-Hill,
New York, 1968).
15. R. H. Kingston, Detection of Optical and Infrared Radiation
(Springer-Verlag, Berlin, 1978).
16. P. A. M. Dirac, "The quantum theory of the electron," Proc. R.
Soc. London Ser. A 117, 610-624 (1928).
17. N. N. Bogoliubov and D. V. Shirkov, Introduction to the Theory
of Quantized Fields (Wiley-Interscience, New York, 1959), Sec.
6-1.
18. J. D. Bjorken and S. D. Drell, Relativistic QuantumMechanics
(McGraw-Hill, New York, 1964), p. 5.
19. M. D. Feit and J. A. Fleck, "Light propagation in graded index
optical fibers," Appl. Opt. 17, 3990-3998 (1978).
20. M. J. Bastiaans, "Transport equations for the Wigner distribution
functions," Opt. Acta 26, 1265-1272 (1979).
21. A. W. Lohmann, J. Ojeda-Castaneda, and N. Streibl, "Symmetries
in coherent and partially coherent fields," Opt. Acta 30, 399-402
(1983).
22. F. D. Tappert, "The parabolic approximation method," in Wave
Propagationand Underwater Acoustics, J. Ehlers et al., eds.
(Springer-Verlag, Berlin, 1977), p. 272.
23. L. Fishman and J. J. McCoy, "Direct and inverse wave propagation theories and the factorized Helmholtz equation. Path integral representations," Proc. Soc. Photo-Opt. Instrum. Eng. (to
be published).
24. M. Born and E. Wolf, Principlesof Optics, 5th ed. (Pergamon,
Oxford, 1975).
25. R. K. Luneburg, Mathematical Theory of Optics (U. California
Press, Berkeley, Calif., 1964), Sec. 45.
26. E. Lalor, "Conditions for the validity of the angular spectrum of
plane waves," J. Opt. Soc. Am. 58, 1235-1237 (1968).
27. W. H. Southwell, "Validity of the Fresnel approximation in the
near field," J. Opt. Soc. Am. 71, 7-14 (1981).