1748 Vol. 37, No.

11 / November 2020 / Journal of the Optical Society of America A Research Article

Analysis of numerical diffraction calculation

methods: from the perspective of phase space
optics and the sampling theorem
Wenhui Zhang,1 Hao Zhang,1, * Colin J. R. Sheppard,2,3 AND Guofan Jin1
1
State Key Laboratory of Precision Measurement Technology and Instruments, Department of Precision Instrument, Tsinghua University,
Beijing 100084, China
2
School of Chemistry, University of Wollongong, Wollongong, NSW 2522, Australia
3
Nanoscopy and NIC@IIT, Istituto Italiano di Tecnologia, Via Enrico Melen, 83 Edificio B, 16152 Genova, Italy
*Corresponding author: [email protected]

Received 2 July 2020; revised 21 September 2020; accepted 21 September 2020; posted 22 September 2020 (Doc. ID 401908);
published 13 October 2020

Diffraction calculations are widely used in applications that require numerical simulation of optical wave propa-
gation. Different numerical diffraction calculation methods have their own transform and sampling properties.
In this study, we provide a unified analysis where five popular fast diffraction calculation methods are analyzed
from the perspective of phase space optics and the sampling theorem: single fast Fourier transform-based Fresnel
transform, Fresnel transfer function approach, Fresnel impulse response approach, angular spectrum method, and
Rayleigh–Sommerfeld convolution. The evolutions of an input signal’s space-bandwidth product (SBP) during
wave propagation are illustrated with the help of a phase space diagram (PSD) and an ABCD matrix. It is demon-
strated that all of the above methods cannot make full use of the SBP of the input signal after diffraction; and some
transform properties have been ignored. Each method has its own restrictions and applicable range. The reason
why different methods have different applicable ranges is explained with physical models. After comprehensively
studying and comparing the effect on the SBP and sampling properties of these methods, suggestions are given for
choosing the proper method for different applications and overcoming the restrictions of corresponding methods.
The PSD and ABCD matrix are used to illustrate the properties of these methods intuitively. Numerical results
are presented to verify the analysis, and potential ways to develop new diffraction calculation methods are also
discussed. © 2020 Optical Society of America
https://doi.org/10.1364/JOSAA.401908

1. INTRODUCTION Recently, with the rise of computational imaging, which

Diffraction is an important feature of optical waves. It is chal- relies directly on numerical diffraction calculation to recon-
lenging and essential to understand the diffraction phenomenon struct images [6–9], to understand and utilize the numerical
in physical optics [1,2]. In the past two centuries, optical diffrac- diffraction calculation well becomes more and more important.
tion has been studied by a large number of scientists, and In addition, numerical diffraction calculation is also the key
many theories have been developed to help people understand process for digital holography [10–12], computer holography
and calculate the diffraction fields [1–5]. Based on either the [13,14], beam shaping [15,16], and diffraction tomography
Rayleigh–Sommerfeld theory or angular spectrum theory, both [17,18], etc. Due to the importance of this topic, numerical
strict solutions to the free-space wave equation, the diffraction diffraction calculation has already been extensively studied and
field is an integral transform of the input aperture [1]. However, discussed in different ways. Voelz et al. analyzed scalar optical
only a few simple apertures have their diffraction fields in an diffraction from the view of chirp function sampling [7,19].
analytical form. Therefore, in most cases, the diffraction field An ideal sampling distance zc was defined; and it was suggested
can be obtained only by means of discrete computation. In such that the impulse response method was better when propaga-
cases, it is essential to correctly sample the input field and the tion distance z > zc , and the transfer function method was
transform kernel to avoid aliasing errors. Meanwhile, it should better when z < zc . Kelly has also concluded similar remarks
be guaranteed that the discrete operation does not change the for the numerical calculation analysis of the Fresnel transform
true diffraction field predicted by analytical theories. [20]. Goodman added a new chapter titled “Computational

1084-7529/20/111748-19 Journal © 2020 Optical Society of America

 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1749

diffraction and propagation” to discuss the sampling issue of to express the structure of Fresnel diffraction. Based on a phase
different numerical diffraction methods in the fourth edition of space analysis, a generalized sampling method was proposed
Introduction to Fourier Optics [2]. [38]. Such a method could realize a resolution enhancement
The task of numerical diffraction calculation is to quickly digital holography at the cost of large detector size [39], which is
obtain the accurate diffraction field. Besides, because of the use similar to the sampling analysis in [40] when the object is space
of fast Fourier transform (FFT) algorithm, the sampling pitch of limited. All these mentioned methods have successfully solved
the diffraction field is fixed by the calculation parameters. Thus, one problem in numerical diffraction calculation and pushed
to flexibly adjust the sampling pitch is also a demand. Therefore, the research forward. However, there lacks a general analysis of
high calculation accuracy, high calculation efficiency, and the existing commonly used diffraction calculation methods to
high calculation flexibility are our goals. Unfortunately, these answer the questions about the properties of different methods,
three goals are contradictory. Generally, massive computing
how to choose one proper method for different applications,
resources are required to get a high calculation accuracy, which
and how to realize a generalized method which can efficiently,
would sacrifice the calculation efficiency. The ordinary FFT
accurately, and flexibly calculate the diffraction field. This uni-
cannot be directly used if high calculation flexibility is desired,
which would also decrease the calculation efficiency. During fied analysis is essential and would help establish a generalized
the last two decades, to get a high calculation accuracy or flexi- framework for understanding the diffraction calculation in
bility without decreasing the efficiency, some techniques have the numerical diffraction calculation community and related
been proposed for different numerical diffraction calculation research areas.
methods. In this work, we have comparatively analyzed five popular
Aiming to overcome the low calculation accuracy of the diffraction calculation methods: single FFT-based Fresnel trans-
angular spectrum method (ASM) in the far field, Matsushima form, Fresnel transfer function (Fres-TF) approach, Fresnel
et al. proposed band-limited ASM [21]; Kim et al. proposed impulse response (Fres-IR) approach, ASM, and Rayleigh–
wide-range ASM [22], and we proposed band-extended ASM Sommerfeld convolution (RSC). The effect on the signal’s
[23]. Both of the first two methods use only the nonaliased SBP and the sampling properties of each method have been
transfer function of ASM to calculate the diffraction field, and studied in detail through the phase space diagram (PSD) and
the last method extends the effective bandwidth of the nona- the sampling theorem. Compared with the analysis in single-
liased transfer function. By these methods, calculation errors space domain or single-frequency domain, joint analysis in
caused by an aliased transfer function can be avoided to get a space-frequency domain with PSD is a more effective method
high accuracy. Aiming to overcome the fixed sampling pitch because it gives a complete signal status. Such a unified analysis
in ASM, Shimobaba et al. proposed scaled ASM to adjust the provides a new perspective: to study whether the mathematics
sampling pitch of the diffraction field by nonuniform FFT [24]. of diffraction calculation methods (examined by the sampling
Inspired by the single FFT-based Fresnel transform, Tomasz theorem) is consistent with the physics (represented in phase
et al. proposed a compact space-bandwidth product (SBP)
space). With the results obtained, we provide solutions to over-
method to make the sampling pitch of the diffraction field vary
come the restrictions of each method, and give suggestions on
with propagation distance. In this way the calculation accuracy
how to choose the proper method, or adapt one, for different
of ASM is high in the large propagation distance range, while
the flexibility of adjusting the sampling pitch is limited because applications. Finally, we suggest an outlook on development of
it is mainly determined by the propagation distance [25]. Note new diffraction methods.
that high accuracy in [25] is determined only for the diffraction This paper is organized as following: Section 1 is the intro-
amplitude, and the phase is not considered. Aiming to overcome duction and motivation of this work; Section 2 introduces the
the fixed sampling pitch in the single FFT-based Fresnel trans- PSD and ABCD matrix that are the tools used in our analysis;
form, which is determined by propagation distance, Zhang et al. Section 3 analyzes five methods in detail; Section 4 provides
proposed two-step Fresnel propagation to adjust the sampling solutions to overcome the restrictions of each method and gives
pitch according to the ratio between these two propagation dis- suggestions on how to choose or adapt one method for specific
tances [26]. Scaled Fresnel diffractions were used in computer applications; and Section 5 summarizes this work and presents
holography and digital holography to implement scale opera- an outlook for new diffraction calculation methods.
tions on the diffraction field based on chirp-z transform [27,28].
Due to aliasing errors in the scaled Fresnel method, the calcula-
tion accuracy would decrease in the near field. For this problem, 2. PSD AND ABCD MATRIX
Shimobaba et al. used a band-limited function to reduce A. PSD
the aliasing errors for high accuracy [29]. Other concerns in
numerical diffraction calculation, such as oblique illumination An optical signal can be viewed in the space domain (x ) or the
[30], shifted observation window [31], propagation between tilt spatial frequency domain ( f x ). If it is viewed simultaneously
planes [32], and high-order dispersion caused by phase aliasing from these two domains (x , f x ), it is represented in a phase
[33], have also been studied. In addition, sampling in the frac- space. Analytically, the PSD of a signal (which we refer as the
tional Fourier domain [34–36] and phase space [37] has been SBP distribution in phase space) can be given by the Wigner
studied. By observing the Fresnel diffraction field at a reference distribution function (WDF).
spherical surface with fractional Fourier transform, a natural For a one-dimensional signal u(x ), its WDF can be expressed
sampling grid, both transverse and longitudinal, was defined as [41,42]
 1750 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 1. (a) PSD of a discretized signal u(n1x ); (b) PSD after paraxial Fresnel propagation; (c) PSD after chirp modulation; (d) PSD after Fourier
transform.

+∞
x0 ∗ x0
Z    
written the lens-system diffraction integral under the paraxial
Wu (x , f x ) = u x+ u x−
−∞ 2 2 approximation in terms of matrix optics [45]. And Tan et al.
have proven that the ABCD matrix method can be extended
× exp(−i2π f x x 0 )dx 0 , (1) into the nonparaxial regime [46]. Therefore, the PSD and the
ABCD matrix can be perfectly combined to analyze how dif-
where ∗ is a complex conjugation. Because WDF doubles ferent calculation methods affect the signal’s SBP and how to
the number of the dimensions of the input signal, a one- satisfy their sampling requirements.
dimensional signal has a two-dimensional WDF, and a Diffraction propagation can be regarded as an operator P on
two-dimensional signal has a four-dimensional WDF that the signal. Take paraxial Fresnel diffraction as example and we
cannot be illustrated intuitively in three-dimensional space. can get the diffraction field,
Therefore, we take a one-dimensional signal for our example
to conduct the analysis, it being easy to extend to the two- exp(i2π z/λ)
dimensional case. WDF is a powerful analytical tool; however, it ũ(X ) = P {u(x )} = √
iλz
is not intuitive. Therefore, we use PSD to represent the signal’s
SBP distribution in phase space, as shown in Fig. 1. Note that
 
iπ
Z
PSD used here is a joint space-spatial frequency diagram to × u(x ) exp (X − x ) dx ,
2
(2)
λz
intuitively show how the SBP distributes, which is more like
a representation in the physical scene. As shown in Fig. 1, the where λ is the wavelength and z is the propagation distance. The
SBP distribution of a signal is a specific area in the space-spatial effect of the operator P on WDF is
frequency domain (x , f x ), which is defined by the location
(x ) and by the range of spatial frequency ( f x ) within which the W(x , f x ) → W(a x + bfx , c x + dfx ) = W(X , f X ), (3)
signal is nonzero [41]. where
In numerical diffraction calculations, the signal u(x ) is
uniformly sampled as u(n1x ) where n = −N/2, −N/2 +
 
a b
M= (4)
1, ... 0, ... N/2 − 1, with N being the sampling number c d
(assumed to be even) and 1x being the sampling pitch. The
discrete sampling operation should obey the sampling theorem is the ABCD matrix. Equation (3) can be rewritten as a matrix
that the sampling rate should be no less than twice the signal’s multiplication,
highest frequency. In this way, the signal is bounded within a     
X a b x
finite effective region in both the space domain and the spatial = . (5)
fX c d fx
frequency domain. Therefore, the PSD of the discretized signal
u(n1x ) is a rectangle, as shown in Fig. 1(a). The side length Equation (5) tells us the new position (X , f X ) of the point in
is L x = N1x and B f = 1/1x along the x axis and f x axis, the phase space after propagation from a point that is originally
respectively. The enclosed area represents the maximum SBP located at (x , f x ). The parameters in matrix M are related to
value of the signal, which is L x B f = N. Signal energy is negli- the transform property of P . For paraxial Fresnel diffraction
gible outside this region. Note that the PSD of discrete signals [42,45],
is periodic [37], but here we only consider one period located at
1 λz
 
the origin. As long as there is no overlap between neighboring
M= . (6)
copies, this operation is valid. The request of no overlapping is 0 1
guaranteed by the sampling theorem, which is used to analyze
the sampling request of different methods. The PSD of the signal after paraxial Fresnel propagation is
shown in Fig. 1(b). As we can see, a horizontal shearing of the
PSD along the x axis is caused in this case. For the nonparaxial
B. ABCD Matrix case, the PSD behaves quite differently. A spherical aberration is
As introduced above, viewing the signal in phase space is con- introduced, and therefore, the shearing of the rectangle is not a
venient. However, it cannot tell us how the signal changes after simple shear, but the sides of the figure become curved [47–49].
diffraction. Fortunately, the ABCD matrix is a mathematical Furthermore, we deal with four different opera-
form that can be used to analyze the evolution properties of tions in Section 3: (1) chirp modulation of the signal,
the signal during paraxial propagation [43,44]. Collins has u(x )exp(iπ x 2 /λz); (2) fast Fourier transformation of the
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1751

Changing f to z, we can get the ABCD matrix of the optical

Fourier transform in free-space diffraction, as given in Eq. (7).
The corresponding PSDs of operation (1) and operation (2) are
shown in Figs. 1(c) and 1(d), respectively. It is shown that chirp
modulation of the signal u(x ) causes a vertical shearing of the
PSD along the f x axis, and FFT causes a rotation of PSD by π/2
rad [41]. Note that these PSDs are determined in two steps: find
the corner points’ coordinates after transformation, and then
connect them in order by straight lines.
Fig. 2. Diagram of the optical Fourier transform.

3. NUMERICAL DIFFRACTION CALCULATION

ANALYSIS
signal, FFT{u(x )}; (3) chirp modulation of the Fourier-
transformed signal, U ( f x )exp(−iπ λz f x2 ); and (4) optical In this section, we analyze five popular numerical diffraction
Fourier transformation of the signal, OFT{u(x )}. The cor- methods: three paraxial Fresnel diffraction methods, which are
responding ABCD matrices of these operations are given single FFT-based Fresnel transform, Fres-TF, Fres-IR, and two
below: rigorous diffraction methods without paraxial approximation,
    which are ASM and RSC. As introduced above, the goal of all
M(1) =
1 0
, M(2) =
0 1
, the methods is to quickly get the accurate diffraction field ũ(X )
1/λz 1 −1 0 in the destination plane of the input signal u(x ), as shown in
Fig. 3. The calculation flexibility will be discussed in Section 4.
λz
   
1 0 0 Before getting into the calculation, we would like to introduce a
M(3) = , M(4) = . (7)
−λz 1 −1/λz 0 little about the diffraction theory used. The first three methods
are based on the Fresnel approximation, where a spherical wave
There are two points requiring attention. First, FFT is a fast is approximated by a parabolic wave. According to how the
algorithm for implementing Fourier transform and should approximation is applied, single FFT-based Fresnel transform is
be applied to discrete variables. For simplicity, we use u(x ) to most suitable for calculation of a diffracted field over a spherical
represent its discrete version u(n1x ), and FFT is applied to the (or parabolic) surface [50]. RSC is based on the first Rayleigh–
dimensionless array u n at the interval of 1x . This convention Sommerfeld integral (RSI) theory, and ASM is its corresponding
also applies to other variables when FFT is used in the paper. Fourier form [2]. Please note that some methods have been pro-
Second, the so-called optical Fourier transform is different posed to improve the accuracy of the Fresnel approximation
from FFT but implemented by FFT. For a signal u(x ), we refer [51], and to model the Rayleigh–Sommerfeld diffraction in k
to ∫ u(x )exp[−i2π X x /(λz)]dx as its optical Fourier trans- space [52], but here we mainly focus on the commonly used
form [42]. Therefore, operation (2) and operation (4) have calculation methods.
different ABCD matrices. FFT is a purely mathematical or Given the input signal u(x ) with the sampling number N
algorithmic transformation, and it has no relationship with the and sampling pitch 1x , it is necessary to calculate the spatial
physical propagation distance z. In contrast, the optical Fourier range L X of the diffraction field ũ(X ). Only inside the range
transform has physical meaning, as explained and associated [−L X /2, L X /2], is it possible to get an accurate result; other-
with Fig. 2. wise there is no physical meaning because L X denotes the reach
Two rays indicated by the red line and the blue line have of diffracted light. According to the simple geometry shown in
the same spatial location x 1 = x 2 = x at the front focal plane Fig. 3, it is easy to calculate such a range as L X = L x + 2z tanα
of the lens. Their spatial frequencies are different: the red with α = arcsin[λ/(21x )]. From Fig. 3, we can see that the sig-
line is parallel to the optical axis having f x 1 = 0, and there is nal u(x ) would spread over a larger spatial range after free-space
an angle θ between the blue line and the optical axis giving diffraction, which can be also observed by comparing Figs. 1(a)
f x 2 = sin θ/λ ≈ tan θ/λ = x 20 /λ f . The approximation is valid and 1(b). However, the methods we introduce below cannot
in paraxial optics. Based on geometrical optics, these two rays make full use of the spreading SBP without any extra cost—for
would be parallel after passing through the lens and have the example, computational complexity.
same spatial frequency f x0 1 = f x0 2 ≈ tan θ 0 /λ = −x /λ f . Their
spatial locations at the rear focal plane of the lens can be easily
obtained as x 10 = λ f · f x 1 = 0 and x 20 = λ f · f x 2 . The above
analysis can be intuitively viewed by the change of PSD shown
in Fig. 2. It is evident that the optical Fourier transform not only
rotates the PSD but scales the rotated PSD with a factor λ f .
Therefore, the optical Fourier transform can be regarded as the
matrix multiplication of a FFT and a scaling matrix,

λf λf
    
0 0 0 1
= . (8) Fig. 3. Diffraction scenario between two planes in the free space.
−1/λ f 0 0 1/λ f −1 0
 1752 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

In single FFT-based Fresnel transform, the observation

window changes with the propagation distance because the
sampling pitch of the diffraction field changes. The other four
methods are all convolution-based, and the observation window
is the same size as the input window. The difference among the
convolution-based methods is that the transfer functions are
directly sampled in the spatial frequency domain in Fres-TF and
ASM, while in Fres-IR and RSC, the spatial impulse response
functions are first sampled and then Fourier transformed to get
the transfer function. In analytical form, the transfer function
and the impulse response function are equivalent. But in discrete
form, they have different sampling requirements, as analyzed
below. Since sampling is essential during discretization, here we
would like to mention a little about the sampling theorem in the
numerical diffraction calculation. There are integral kernels in Fig. 4. PSD of the signal before (a) and after (b)–(d) trans-
the diffraction calculations, and these kernels should be sam- formations. PSD after z = zc diffraction propagation is shown
pled correctly during discretization. The principle is that the in (e).
sampling rate should be no less than twice the maximum local
spatial frequency of the kernels’ phases. For different methods,
the kernels are different, and we analyze them one by one in the where ϕ = π x 2 /λz is the phase of the inner chirp function.
From inequality (11) we can get
following parts of this paper.
Let us start with the single FFT-based Fresnel transform, since 21x
it is also called the “direct method” [19]. z≥ |x |max . (12)
λ

A. Single FFT-Based Fresnel Transform Substituting |x |max = L x /2 = N1x /2, we get

Equation (2), describing paraxial Fresnel propagation, can be N12x
rewritten as z≥ . (13)
λ
 Z   
exp(i2π z/λ) iπ 2 iπ 2 This means that the inner chirp function can be only sam-
ũ(X ) = √ exp X u (x ) exp x
iλz λz λz pled correctly within this distance range, and the propagation
  distance should be as large as possible to satisfy its sampling
X requirement. This conclusion is also consistent with the trend
× exp −i2π x dx .
λz (9) of signal bandwidth after diffraction: the larger the distance
z, the smaller the increment of bandwidth L x /λz, the smaller
Ignoring the constant term, we can see that there are three the sampling SBP, and the easier to satisfy its sampling require-
steps to get the diffraction field ũ(X ) from the input signal u(x ): ment. In conventional methods, it is thought that as long as the
(1) chirp modulation of u(x ); (2) (optically) Fourier transform inner chirp function is sampled without aliasing, the optical
the chirp-modulated signal; and (3) chirp modulation of the Fourier transform can be operated correctly because it satisfies
Fourier transformed signal. Step (2) is implemented by a FFT in the sampling SBP. However, it is not rigorous but, nevertheless,
practice by regarding f X = X /λz, but it has physical meaning. a correct estimate. As we stated above, the chirp modulation
Please note that we call the chirp function in (1) the inner chirp (multiplication with a chirp function) would cause a bandwidth
function and that in (3) the outer chirp function. If we use the
increase, as shown in Fig. 4(b). That is to say, the product has
ABCD matrix to model these steps, we can get
a larger bandwidth. Therefore, the sampling theorem should
   
λz
   be applied to the product rather than only the chirp function.
X
= 1 0 0 1 0 x
. For small-bandwidth signal and a small-bandwidth increase,
fX 1 λz 1 −1 λz 0 1 λz 1 fx
(10) to consider the chirp function separately turns out to provide a
correct estimate.
PSDs after each step are shown in Fig. 4 to help find which The other operation needing attention is the outer chirp
step needs attention in sampling to avoid aliasing errors. The modulation to the Fourier transformed signal. Similarly, we find
first step is the chirp modulation of the signal u(x ). As we can the sampling requirement by solving
see, the bandwidth becomes larger and the spatial length holds  1 ∂8 
 
1
constant, which leads to a larger sampling SBP. By the sampling  2π ∂ X  ≤ 21 , (14)
 
SBP, we refer to the area of the circumscribed rectangle of the max X

parallelogram. To make sure that the chirp-modulated signal is where 8 = π X 2 /λz is the phase of the outer chirp func-
correct, the chirp function has to be sampled correctly, which is tion. Before solving the inequality for z, we need to calculate
 1 ∂ϕ 
 
1 the sampling pitch 1 X of the diffraction field. As shown
 2π ∂ x  ≤ 21 , (11) in Fig. 4(d), the spatial length of the diffraction field is
 
max x
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1753

L X = L x + λz B f = N1x + λz/1x . Therefore, we have Table 1. Parameters Used for Single FFT-Based
1 X = L X /N = 1x + λz/N1x and |X |max = L X /2 = Fresnel Transform
N1x /2 + λz/21x . Substituting the expressions of 1 X and Parameters Values
|X |max into inequality (14), we have
Sampling number of the signal N = 1000
λ2 2 Sampling pitch of the signal 1x = 5 µm
z + λz + N12x ≤ 0. (15) Wavelength λ = 0.5 µm
N12x
There is no meaningful solution to inequality (15) because phase distribution of the outer chirp function. The results with
λ2 − 4 · (λ2 /N12x ) · N12x < 0, which means the sampling z2 = 100 mm are shown in Figs. 5(d)–5(f ) in the same order. In
requirement of the outer chirp function over the whole diffrac- addition, the RSI results are also presented as reference. RSI is
tion field cannot be satisfied. Fortunately, due to the property of calculated by the point-by-point integral without fast calcula-
FFT, the actual spatial length of ũ(X ) is L X = λz B f = λz/1x tion algorithms. Therefore, it can be used as a reference as long as
[7]. In this situation, the new sampling pitch of the diffrac- the sampling of the input signal and diffraction field are correct.
tion field is 1 X = L X /N = λz/N1x . Substituting the new From Fig. 5, we can see that when the propagation distance
expressions of 1 X and |X |max into inequality (14), we have z1 = zc , the amplitude and phase results given by the single
N12x FFT-based Fresnel transform are correct. The little difference
z≤ . (16) between the results calculated by single FFT-based Fresnel
λ
transform and RSI is because there is no bandwidth limitation in
This means that the outer chirp function can be only sam- RSI, while there is in the other. However, when the propagation
pled correctly within this distance range, and the propagation distance z2 > zc , the phase of the diffraction field is no longer
distance should be as small as possible to satisfy its sampling correct, while the amplitude is still correct, which is consistent
requirement. This conclusion is also consistent with the trend with our analysis. By further analyzing the phase profile, we
of spatial length of the signal after diffraction: the smaller the notice that it is correct for small X values, as indicated by the
distance z, the smaller the increment of spatial length λz B f , the red rectangle in Fig. 5(g). That is to say, the sampling of the
smaller the sampling SBP, and the easier to satisfy its sampling outer chirp in these parts is correct, which is consistent with
requirement. the observation from Fig. 5(h), where the inner part is correct
Combining inequality (13) and (16), only one propa- and the outer parts are aliased. We could calculate this effective
gation distance zc = N12x /λ can simultaneously satisfy region by solving inequality (14) but regarding X as the variable,
the sampling requirements of these two chirp functions,
N1x
which severely limits the application of the single FFT- |X effective | ≤ . (17)
based Fresnel transform. Defining the Fresnel number 2
NF = (N1x /2)2 /λz, z = zc is equivalent to NF = 0.25. This means that the calculated diffraction field is correct
Therefore, NF can be also used to express the distance range. within this region. The small difference with that calculated by
In this paper, we directly give the distance range, and it can the RSI is due to the limited bandwidth of the single FFT-based
be translated to NF if needed. At z = zc , we can calculate Fresnel transform and approximation in the sampling process.
the spatial length of the diffraction field and its bandwidth, Of course, one can do zero padding to the outer chirp to make
which are L X (z = zc ) = λ · (N12x /λ)/1x = N1x and the whole diffraction field correct, which will be discussed in
B f (z = zc ) = 1/1 X = N1x /(λ · N12x /λ) = 1/1x . The Section 4.
PSD of the diffraction field in this case is shown in Fig. 4(e). In the single FFT-based Fresnel transform, the size of the
The yellow rectangle surrounded by the dashed lines is the cal- diffraction field is automatically scaled with the propagation dis-
culated SBP inside the range L X (z = zc ) · B f (z = zc ), while tance in the form of L X = λz/1x , not as indicated in Fig. 4(d)
the effective SBP after propagation is inside the parallelogram. L X = N1x + λz/1x , because of the property of FFT. Besides,
Therefore, the effective SBP is not fully obtained with the this method does not require zero padding to the input signal,
original N sampling points. which is needed in the convolution-based methods, to avoid
By revisiting these two chirp functions, it is easy to notice that circular convolution errors. However, it is only applicable at
the outer chirp function only affects the phase distribution of the one specific propagation distance to ensure the correctness of
diffraction field; therefore, if only amplitude (or intensity) is of the calculated diffraction field, if the phase is required, and this
concern, one only needs to satisfy inequality (13). correctness depends on the sampling. Actually, zero padding
To verify the above analysis, the diffraction field of a rectangle can make more spatial frequency components transferred in this
aperture, u(x ) = rect(x /l ), where l = 2.5 mm is the width of method because f X = X /λz would increase with X .
the aperture, is calculated by single FFT-based Fresnel trans-
form under different propagation distances. The parameters
B. Fres-TF and Fres-IR
of this calculation are listed in Table 1. With these parameters,
zc = N12x /λ = 50 mm, and we set two propagation distances There is another understanding of Eq. (2): that the diffraction
as z1 = 50 mm and z2 = 100 mm. field ũ(X ) is the linear convolution of the input signal and the
Figure 5 shows the results. Figures 5(a) and 5(b) show convolution kernel,
the diffraction amplitude and phase distributions with exp(i2π z/λ)

iπ X 2

z1 = 50 mm, respectively. Figure 5(c) shows the unwrapped h(X ) = √ exp , (18)
iλz λz
 1754 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 5. Diffraction fields calculated by the single FFT-based Fresnel transform and RSI with (a)–(e) z1 = 50 mm; and (f )–(j) z2 = 100 mm .
The unwrapped phases of the outer chirp function [(c) and (h)] are also shown in these two cases to illustrate its correct part.

which is modeled as convolution-based methods, x and X are located at the same

Z positions because 1 X = 1x . To get fast calculation of Eq. (20),
ũ(X ) = u(x )h(X − x )dx . (19) based on the convolution theorem, multiplication of the Fourier
transform of u(X ) and the Fourier transform of h(X ) is used,
Usually, Eq. (19) is represented by a convolution symbol as
rather the convolution because the computational complexity
ũ(X ) = u(X ) ⊗ h(X ), (20) of multiplication is much cheaper than that of convolution. We
where ⊗ denotes linear convolution. Please note that the input can obtain the analytic expression for the Fourier transform of
signal is expressed as u(X ) here due to the mathematical con- h(X ) which is
ventions on the convolution formula. Actually, x and X are
just coordinate symbols, used to indicate the spatial locations exp(i2π z/λ)
of the signal before and after diffraction propagation; in the H( f X ) = √ exp(−iπ λz f X2 ), (21)
iλz
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1755

Fig. 6. PSDs of the signal (a) before and (b)–(d) after

transformations.

where f X is the spatial frequency. Mathematically, h(X ) and Fig. 7. Two strategies to avoid circular convolution errors. (a) The
H( f X ) are equivalent to model the diffraction. However, standard process, which does zero padding to both u(X ) and h(X ),
they have totally different sampling properties during discrete and (b) the popular process, which only does zero padding to u(X ) and
calculation. Based on how the transfer function is modeled, models h(X ) with 2 N points to get h̄(X ).
the convolution-based paraxial Fresnel diffraction can be
divided into two approaches: Fres-TF and Fres-IR. In Fres-TF,
the transfer function is directly modeled by H( f X ), while N-zero padded u(X ) and h(X ) are û(X ) and ĥ(X ), respec-
in Fres-IR, the transfer function is obtained as the Fourier tively: the circular convolution of û(X ) and ĥ(X ) is equivalent
transform of h(X ). In these two approaches, there are also to the linear convolution of u(X ) and h(X ) [53].
three steps from u(X ) to ũ(X ): (1) Fourier transform u(X ) to However, in practice, no one uses the above strategy. The
get U ( f X ) by FFT; (2) obtain A( f X ) = U ( f X ) · H( f X ) or reason is that since we have to use 2 N sampling points, why
A( f X ) = U ( f X ) · FFT{h(X )}; (3) inverse Fourier transform not fully use them to sample the impulse response function
A( f X ) to get the diffraction field ũ(X ) by IFFT, but only use N points surrounded by N zeros? The popular
process is just that: (1) zero padding the N point sampled u(X )
ũ(X ) = IFFT{A( f X )}. (22) to û(X ); (2) using 2 N points to sample the impulse response
If we use the ABCD matrix to model these steps, we can get function or the transfer function to get h̄(X ) or H̄( f X ), respec-
tively; (3) inverse Fourier transform the product of their Fourier
transforms IFFT{ Ā( f X )} = IFFT{FFT{û(X )} · FFT{h̄(X )}}
      
X 0 −1 1 0 0 1 x
= . (23) or IFFT{ Ā( f X )} = IFFT{FFT{û(X )} · H̄( f X )} to get the
fX 1 0 −λz 1 −1 0 fx
diffraction field. The difference between this strategy and the
Similarly, we draw the PSDs after each step. first strategy is shown in Fig. 7.
Note that in Fig. 6, L x and B f are only values without units. Figure 7(a) shows the standard process to transform linear
The units are given by the domains (space or spatial frequency) convolution to circular convolution by zero padding. All the 2 N
where they are drawn. Because FFT does not change the sam- points of ũ(X ) are correct. However, the high spatial frequencies
pling SBP, there is only one step needing attention to ensure of ĥ(X ) are lost because they are sampled by zeros. In contrast,
correct sampling: the sampling of H( f X ) in Fres-TF and h(X ) Fig. 7(b) shows the popular process in practice, where all the 2 N
in Fres-IR. points are employed to sample the impulse response function
Before discussing the sampling of H( f X ) or h(X ), it is nec- h̄(X ). In this way, the high spatial frequencies of h̄(X ) are fully
essary to discuss the difference between circular convolution used. The cost of this process is that only N points of ũ(X ) are
and linear convolution, since the convolution theorem for dis- correct because h̄(X ) is not obtained by zero padding. Note that
crete signals is based on circular convolution. The convolution Fig. 7(b) shows the Fres-IR case; for Fres-TF, H̄( f X ) is directly
theorem states that the convolution of two signals is equiva- sampled by 2 N points. The process shown in Fig. 7(b) can be
lent to the inverse Fourier transform of the product of their regarded as that: to fast calculate the linear convolution of N
Fourier transforms. In discrete calculation, the convolution point sampled u(X ) and 2 N point sampled h̄(X ), N zeros
is circular convolution. However, Eq. (19) is a linear convo- are padded to u(X ) for the 2 N point sampled û(X ), and the
lution. Therefore, to make the convolution theorem suitable circular convolution of û(X ) and h̄(X ) is carried out, which
for Eq. (19), zero padding to u(X ) and h(X ) is required [53] can be accelerated by FFT. Therefore, only N points are correct
and the number of padded zeros is N − 1, at least. Suppose the [53]. In this way, the spatial length of the correct diffraction
 1756 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 8. PSDs with (a) z1 < ẑ, (b) z2 = ẑ, and (c) z3 > ẑ.

field û(X ) is the same as that of the input signal u(X ), which is z2 = 100 mm, z3 = 300 mm, and the critical distance in this
N1x . The padded zeros are used only to guarantee the correct case is ẑ = 2N12x /λ = 100 mm. RSI results are also given as
sampling in the spatial frequency domain. reference.
Now, let us analyze the sampling requirement of H̄( f X ) From Fig. 9, we can conclude these points: (1) with z1 < ẑ,
or h̄(X ), since they are used in practice. They have the same Fres-TF gives correct results, while aliasing errors appear in
analytical expressions as H( f X ) or h(X ), respectively, and the Fres-IR results; (2) with z2 = ẑ, both results given by Fres-TF
difference is just that they are sampled by 2 N points not N and Fres-IR are correct, and they are identical; (3) with z3 > ẑ,
points. √ Fres-IR gives correct results, while high-frequency noises appear
For H̄( f X ), the first term, exp(i2π z/λ)/ iλz, is a con- in Fres-TF results. One more observation is that with z increas-
stant that does not induce sampling aliasing; the second ing, the diffraction field inside the observation window, which
one, exp(−iπ λz f X2 ), does, because it varies with the spatial is [−2.5 mm, 2.5 mm], is becoming sparse. “Sparse” means
frequency f X . According to the sampling theorem, the diffraction field is spreading over a larger area, and the part
 1 ∂ϕ( f X ) 
  inside the observation window becomes smooth. This is also
  ≤ 1 , (24) consistent with the above analysis: that with larger z values, the
 2π ∂ f  21 f
X max SBP utilization rate would decrease because the covered effective
SBP becomes smaller.
where ϕ( f X ) = −π λz f X2 is the phase of H̄( f X ) and
The reason why Fres-IR and Fres-TF have opposite and
1 f = B f /2N = 1/2N1x is the sampling pitch of H̄( f X ). complementary applicable distance ranges can be explained
From inequality (24), we have physically. A spherical wave model is used in Fres-IR and a plane
2N12x wave model is used in Fres-TF, as shown in Fig. 10. Because
z≤ , (25) the input signal is sampled by pitch 1x , the highest spatial
λ
frequency 1/21x determines the nonaliased biggest diffrac-
which means that H̄( f X ) can be correctly sampled only within tion angle, which is αmax = arcsin(λ/21x ). This case occurs
this distance range. √ precisely when z = ẑ both in Fres-IR and Fres-TF, as shown in
Similarly for h̄(X ), the first term, exp(i2π z/λ)/ iλz, is a Figs. 10(a) and 10(b). However, when z < ẑ, the biggest diffrac-
constant that does not induce sampling aliasing; the second one, tion angle in Fres-IR is bigger than αmax , as line ® indicates
exp(iπ X 2 /λz), does, because it varies with X . According to the in Fig. 10(a), which would lead to aliasing error, while when
sampling theorem, z > ẑ, the biggest diffraction angle in Fres-IR is smaller than
αmax , as line ¬ indicates in Fig. 10(a), which would give correct
 1 ∂ϕ(X ) 
 
1
 2π ∂ X  ≤ 21 ,
  (26) results. As for Fres-TF, there is no worry about the diffraction
max X angle exceeding αmax because the highest spatial frequency is
always set to be 1/21x . However, the problem is whether the
where ϕ(X ) = π X 2 /λz is the phase of h̄(X ) and 1 X = 1x is
observation window could receive such high spatial frequencies.
the sampling pitch of h̄(X ). From inequality (26), we have
As shown in Fig. 10(b), when z < ẑ, all the spatial frequencies
2N12x can be received by the observation window, providing correct
z≥ , (27) results, while when z > ẑ, some high spatial frequencies exceed
λ
the observation window, as line ® indicates, which would lead
which means that h̄(X ) can be correctly sampled only within to high-frequency noises in the diffraction field. In addition, we
this distance range. can see another property of the plane-wave-model methods. For
By comparing inequalities (25) and (27), we can see that both small propagation distances, the diffracted light from any point
Fres-TF and Fres-IR can be used when z = ẑ = 2N12x /λ. PSDs inside the source window cannot cover the full observation win-
with z1 < ẑ, z2 = ẑ and z3 > ẑ are shown in Fig. 8. The yellow dow. Therefore, the calculated results may be slightly different
rectangles surrounded by dashed line are the calculated SBPs, from that calculated by RSI, which can achieve a full-bandwidth
which do not fully cover the effective SBP inside the parallelo- calculation, as shown in Fig. 9.
gram. It is evident that with small z values, the utilization rate
of the effective SBP, which is defined as the ratio of the covered
C. ASM and RSC
effective SBP and the whole effective SBP, is high.
To verify the above analysis, we carried out numerical simula- The above three methods are all based on the Fresnel approxima-
tions with parameters listed in Table 1, and the results are shown tion, which is applicable in paraxial optics. Rigorous numerical
in Fig. 9. The propagation distances are chosen as z1 = 30 mm, diffraction calculation methods are often needed, such as in the
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1757

Fig. 9. Diffraction fields calculated by Fres-TF, Fres-IR, and RSI with z1 = 30 mm, z2 = 100 mm, and z3 = 300 mm.

case of high numerical aperture. ASM and RSC are strict solu-
exp(i2πr /λ)
 
tions to the diffraction theory without paraxial approximation, 1 z 1 i2π
h RSC (X , Y ) = − , (29)
which can be used in such cases. Like Fres-TF and Fres-IR, ASM 2π r r λ r
and RSC also model diffraction with plane waves and spherical √
waves, respectively. That is, ASM models diffraction by the where r = X 2 + Y 2 + z 2 . Mathematically, HASM ( f X , f Y )
transfer function, and RSC models diffraction by the impulse and h RSC (X , Y ) are a Fourier transform pair [2]. However,
response function, which are given as below, again, they have opposite applicable distance ranges when
  discretely sampled. The calculation process is similar to that
2π
q
HASM ( f X , f Y ) = exp i z 1 − (λ f X )2 − (λ f Y )2 , (28) introduced above, except the nonapproximated functions, and
λ
 1758 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 10. Diagrams to physically explain why Fres-IR and Fres-TF have opposite and complementary applicable distance ranges.

Fig. 11. (a) PSD shown in x , f x coordinate system; pseudo-PSD shown in x , fˆx coordinate system (b) before and (c) after diffraction.

give ũ(X , Y ) = IFFT{FFT{û(X , Y )} · HASM ( f X , f Y )} or Table 2. Parameters Used in the Simulation to Show
ũ(X , Y ) = IFFT{FFT{û(X , Y )} · FFT{h RSC (X , Y )}}. the Relationship of Pseudo-PSDs in the x, f̂x Coordinate
When it comes to obtaining the PSD through the ABCD System and PSDs in the x, fx Coordinate System
matrix, we should be careful because the ABCD matrix is Parameters Values Units
mainly used in paraxial optics. Although it has been proved that
Length of input signal L x = 10 mm
the ABCD matrix can be extended to nonparaxial optics,
Bandwidth of input signal B f = 3 and 1 µm−1
that would lead to a very complicated form [46]. This is
Wavelength λ = 0.5 µm
because sinα ≈ tanα no longer holds in this case; and the Propagation distance z = 10 mm
polynomial expansion of trigonometric functions is complex.
Physically, X = x + z tanα, where α is the direction angle and
f X = f x = sinα/λ. When α is small, i.e., f x is small, we have the pseudo-PSD after diffraction is always a parallelogram, no
X = x + z tanα ≈ x + z sinα = x + λz f x ; in this way we can matter whether in paraxial or nonparaxial optics. This is because
get a concise ABCD matrix. However, p when α is large, i.e., f x is fˆx = tanα/λ and X = x + z tanα describe the real physical
large, tanα = sinα/cosα = λ f x / 1 − (λ f x )2 6 = sinα; we do scene. However, PSD is described in the phase space where the
not have a concise ABCD matrix anymore if the vector [x , f x ] is spatial frequency f x is used. Therefore, the pseudo-PSDs shown
still used [54]. Therefore, in order to continue using the concise in Figs. 11(b) and 11(c) are not real PSDs but the sketched
ABCD matrix form to describe the diffraction phenomenon, version because fˆx is obtainedpby sketching the corresponding
we definep a new parameter named the pseudo-spatial frequency f x , and the sketching rate 1/ 1 − (λ f x )2 is related with f x .
fˆx = f x / 1 − (λ f x )2 and use the ABCD matrix to find the Mathematically, after considering the pseudo-PSD in the x , fˆx
relationship of the vector [x , fˆx ] before and after diffraction. In coordinate system, the PSD in the x , f x coordinate system
this way, we still have the concise form X = x + λz fˆx , can
p be obtained by sketching the pseudo-PSD with the rate
1 − (λ f x )2 . It can be expected that the difference between
1 λz
    
X x
= . (30) pseudo-PSD and PSD would be large in the nonparaxial cases
fˆX 1 1 fˆx
and would be become small in the paraxial cases. We simulate
Please note that this is just for convenience and the actual these two cases with the parameters shown in Table 2.
spatial frequency is always f x rather fˆx . Therefore, the SBP When
p f x is small, i.e., the bandwidth is small, fˆx =
distribution in such a domain (x , fˆx ) is not a PSD anymore, f x / 1 − (λ f x )2 ≈ f x , the pseudo-PSDs in x , fˆx coordi-
but a pseudo one, which we define as a pseudo-PSD. Because nate system would tend to PSDs in the x , f x coordinate system.
we use the pseudo-spatial frequency fˆx , it is complicated to As can be seen from Fig. 12, with decreasing bandwidth B f ,
draw the PSDs after each step, the difference between fˆx and f x becomes smaller and PSDs
 especially the multiplication with
in the x , f x coordinate system are becoming parallelograms.
p 
the transfer function exp i λ z 1 − (λ f x ) . Therefore, we
2π 2
It is also evident that even with large B f , the central parts of
directly show the pseudo-PSD after diffraction and skip the
pseudo-PSDs in the x , fˆx coordinate system and the central
intermediate steps in the x , fˆx coordinate, as shown in Fig. 11. parts of the PSDs in the x , f x coordinate system are nearly
In the x , fˆx coordinate system,
p the bandwidth B̂ f is larger coincident, which is consistent with the fact that sinα ≈ tanα
than B f because fˆx = f x / 1 − (λ f x )2 > f x . By using fˆx , when f x is small.
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1759

Fig. 12. PSDs in the x , f x coordinate system and the pseudo-PSDs in the x , fˆx coordinate system of large-bandwidth signal and small-bandwidth
signal. (a) PSD of a large-bandwidth (3 µm−1 ) signal before diffraction; (b) pseudo-PSD of the signal after diffraction; (c) PSD of the signal after
diffraction; (d) PSD of a small-bandwidth (1 µm−1 ) signal before diffraction; (e) pseudo-PSD of the signal after diffraction; (f ) PSD of the signal after
diffraction.

Figure 12 shows the PSDs/pseudo-PSDs given by ASM and Table 3. Parameters Used for ASM and RSC
RSC for both ASM and RSC. With these two methods, the Parameters Values
observation window has the same size as the spatial length of
Sampling number of the signal N = 1000
the input signal, which is 10 mm in this case. This is because the
Sampling pitch of the signal 1x = 1 µm
sampling pitch in the destination plane is the same as that of Wavelength λ = 0.5 µm
the input plane. That is to say, even with these two nonparaxial
methods, the SBP cannot be fully used, either, because the
diffraction field spreads on a larger spatial range. which means that h RSC (X ) can be correctly sampled only within
Next, let us analyze the sampling properties of HASM ( f X ) and this distance range.
h RSC (X ). By comparing inequalities (32), (34) and inequalities (25),
In ASM, the key to getting the correct diffraction field is (27), we can find that the relationship between ASM and
to sample a nonaliased transfer function HASM ( f X ), which is RSC is very similar to that between p Fres-TF and Fres-IR.
guaranteed by The only difference is the factor 1 − (λ/21x )2 . When
1x  λ, ASM and RSC will pdegenerate into Fres-TF and
 1 ∂ϕ( f X )  1 − (λ/21x )2 ≈ 1 in that
 
 ≤ 1 , (31)
Fres-IR, respectively; since
case. From another perspective, 1x  λ means the band-

 2π ∂ f  21 f
X max
width of the input signalp 1/1x is much smaller compared
to 1/λ, wherefore 2π z 1 − (λ f x )2 ≈ 2π z − π λz f x 2 and
p
where ϕ( f X ) = 2π
λ
z 1 − (λ f X )2 is the phase of HASM ( f X ) √ λ λ
and 1 f = B f /2N = 1/2N1x is the sampling pitch of 2π
λ
X 2 + z 2 ≈ 2π
λ
z + π x 2 /λz. The phases of ASM’s transfer
HASM ( f X ). From inequality (31), we get function and RSC’s impulse response function degenerate
s into that of Fres-TF’s transfer function and Fres-IR’s impulse
2N12x λ
 2 response function, respectively, while for large numerical aper-
z≤ 1− , (32) ture cases, ASM or RSC cannot be approximated as Fres-TF or
λ 21x Fres-IR.
To verify the above analysis, we carry out numerical simula-
which means that HASM ( f X ) can be correctly sampled only tions with the parameters listed in Table 3; the results are shown
within this distance range. Similarly, we can get the distance in Fig. 13. The propagation distances are chosen as z1 = 1 mm,
range where h RSC (X ) is applicable by calculating z2 = 3.87 mm, z3 = 10 mm, and p the critical distance with these
parameters is ẑ = 2N12x /λ · 1 − (λ/21x )2 = 3.87 mm.
 1 ∂ϕ (X ) 
 
1
 2π ∂ X  ≤ 21 ,
  (33) RSI results are also given as reference. What we can conclude
max X from Fig. 13 for ASM and RSC is similar to that concluded from
√ Fig. 9 for Fres-TF and Fres-IR. One more thing to notice is that
where ϕ(X ) = 2πλ
X 2 + z 2 is the phase of h RSC (X ) and Fres-TF can be correctly used in the applicable range of ASM,
1 X = 1x is the sampling pitch of h RSC (X ). From inequality and Fres-IR can be correctly used in the most applicable range
(33), we get of RSC. Here, “correctly” is in the sense of sampling. However,
s this does not mean that in the situations where ASM or RSC is
2N12x

λ
2 available, Fres-TF or Fres-IR is also available. The criterion of
z≥ 1− , (34) judgment is the ratio between the bandwidth of the signal 1/1x
λ 21x
 1760 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 13. Diffraction fields calculated by the ASM, RSC, and RSI with z1 = 1 mm, z2 = 3.87 mm, and z3 = 10 mm.

p
and 1/λ: if it is so small that 1 − (λ/21x )2 ≈ 1, Fres-TF and 4. METHODS COMPARISON, AND
Fres-IR can be used. IMPROVEMENT AND SELECTION
Up until now, all the five methods have been analyzed from SUGGESTIONS
the perspective of PSD and the sampling theorem. In Section 4, Among these five methods, single FFT-based Fresnel transform,
we would like to compare these methods and give some sug- Fres-TF, Fres-IR, ASM, and RSC, the last four methods are all
gestions on how to choose one proper method for different convolution-based methods, which regard the diffraction inte-
applications, or adapt one of these methods for a specific gral as a convolution. Fres-TF and ASM model the diffraction
purpose. in the spatial frequency domain by the transfer function, and
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1761

Fres-IR and RSC model the diffraction in the space domain by use FFT 3 times; and single FFT-based Fresnel transform uses
the impulse response function. There is, in each case, a func- FFT once. If any method can be used for an application and
tion to describe the response of the system during free-space high calculation efficiency is desired, especially in the iterative
diffraction propagation. All these four methods are based on optimizations, single FFT-based Fresnel transform, ASM, and
the FFT algorithm. The input signal is transformed to the Fres-TF should be prioritized.
spatial frequency domain by FFT and then multiplied with Although these methods enable us to calculate the diffrac-
the transfer function or the Fourier transform of the impulse tion field, there are still some functions these methods cannot
response function; then the product to the space domain is achieve, for example, scaling the sampling pitch in the destina-
transformed by inverse FFT to get the diffraction field. Due to tion plane based on needs, obtaining the region of interest rather
the property of the FFT algorithm, the observation window the whole diffraction field, etc. We call these two functions scal-
has the same size as the spatial length of the input signal. This ing and zooming, respectively. There are also some restrictions
is because the sampling pitch in the destination plane is the of each method, and here we provide some solutions to alleviate
same as that in the input plane, 1 X = 1x . On the contrary, the these restrictions and achieve these two functions.
first method, the single FFT-based Fresnel transform, does not Let us first alleviate the restrictions of each method.
obey 1 X = 1x , but 1 X = λz/N1x , which changes with the For single FFT-based Fresnel transform, one restriction is
propagation distance. This is because single FFT-based Fresnel that it works only at a specific propagation distance when both
transform is not a convolution-based method and what the amplitude and phase are of concern, as analyzed in Section 3.A.
FFT algorithm achieves is actually an optical Fourier transform, This is because the two chirp functions have opposite applicable
which can be seen from Eq. (9). distance ranges. By comparing inequalities (13) and (16), we
Whether the sampling pitch in the destination plane changes notice that we can use different sampling numbers to sample
or not is an important criterion to choose among diffraction these two chirp functions based on needs, i.e., N for the inner
calculation methods. In applications where the calculated chirp function and N̂ for the outer chirp function. As long as
diffraction field is as large as the input signal, such as recon- N̂ > N, the applicable distance range of these two chirp func-
structing an image in digital holographic microscopy [12], the tions would overlap, and the overlap region depends on how
convolution-based methods are preferable, while in applications much is N̂ bigger than N. In this way, we get the applicable
where the calculated diffraction field is larger than the input distance range of these two chirp functions,
signal, such as holographic projection from computer-generated
holograms [55,56], the single FFT-based Fresnel transform N12x
z≥ , (35)
is preferable. As for how to choose one convolution-based λ
method, ASM and RSC can be used in their applicable distance
ranges, no matter whether they are paraxial or nonparaxial; N̂12x
Fres-TF and Fres-IR can be only used in their applicable dis- z≤ . (36)
λ
tance ranges in paraxial optics. Therefore, ASM and RSC may
be better. On the other hand, because the SBP utilization rate Therefore, single FFT-based Fresnel transform can be used at
would decrease with the propagation distance, it is better to an extended distance range. Sampling number N̂ can be calcu-
design the application case where ASM is applicable for higher lated by inequality (36) as N̂ > [λz/12x ], where [·] represents a
SBP utilization, for example, to put the object as near to the rounding operation and z is the propagation distance in practi-
sensor as possible in digital holography. A diagrammatic com- cal applications.
parison among these four convolution-based methods is shown The corresponding algorithm can be divided into three steps:
in Fig. 14. (1) Multiply the N point sampled signal u(x ) with the inner
Another dimension of comparison is the calculation effi- chirp function exp(iπ x 2 /λz) that is also N point sampled;
ciency. ASM and Fres-TF use FFT twice; RSC and Fres-IR (2) apply N̂ point FFT to this product u(x )exp(iπ x 2 /λz);

Fig. 14. Diagrammatic comparison among these four convolution-based methods.

 1762 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 15. Diffraction fields calculated by the adapted single FFT-based Fresnel transform and RSI with z2 = 100 mm.

and (3) multiply N̂ point sampled outer chirp function For ASM, many methods have been proposed to make it
exp(iπ X 2 /λz) with the Fourier
√transform obtained in (2).
applicable for large propagation distances. Here, a brief review
The constant term exp i 2π λ
z / iλz can be simultaneously associated with some analysis is given.
p Recall the phase of ASM’s
multiplied in step (3). transfer function, ϕ( f X ) = λ z 1 − (λ f X )2 . This time, we do
2π

In Section 3.A, we made numerical simulations to demon- not calculate the distance range z but the spatial frequency f X .
strate that the phase distribution given by single FFT-based Based on inequality (31), we can get
Fresnel transform would be aliased if the propagation distance 1
is larger than N12x /λz. Here, we repeat this simulation while | f X | ≤ fc = p , (37)
using the adapted version when z2 = 100 mm; the results are λ 1 + z 2 /N 2 12x
shown in Fig. 15. In this simulation, N̂ = 2000, and other which means that the transfer function only inside [− f c , f c ]
parameters are the same as shown in Table 1. can be sampled without aliasing. From inequality (37), it is clear
It is clear that the outer chirp function has been sampled that f c would decrease with z and increase with N. The high-
without aliasing and the phase distribution of the calculated est spatial frequency associated with the sampling pitch 1x is
diffraction field by the adapted single FFT-based Fresnel 1/21x . Within the range given by inequality (32), f c = 1/21x
transform is correct. From the perspective of PSD, when the and f c would decrease with z beyond that range. Therefore,
propagation distance becomes large, the required sampling there are two typical solutions to sample the ASM’s transfer
SBP is also becoming large, which means more sampling points function without aliasing in large distances: (1) increasing
should be used for correct calculation. This is the reason why sampling number N; and (2) forcing the components of ASM’s
transfer function beyond [− f c , f c ] to be zero. The first solution
increasing sampling points ( N̂ > N) could give correct results.
makes f c = 1/21x always hold by increasing N, that is the zero-
For the convolution-based methods, ASM and Fres-TF can
padding method [57]. The number of padded zeros increases
be used only with a small propagation distance; and RSC and dramatically with large propagation distance. The second solu-
Fres-IR can be used only with a large propagation distance, as tion does not increase sampling number but limits the effective
shown in Fig. 14. The reason for these restrictions is because bandwidth of the transfer function. Since not all the bandwidth
aliasing errors would appear if they are used beyond their appli- [−1/21x , 1/21x ] can be correctly sampled, the aliased part is
cable distance ranges. Therefore, to make ASM and Fres-TF forced to be zero to avoid aliasing errors, which is the key point
applicable for large propagation distances or RSC and Fres-IR in this method, called band-limited ASM [21]. Inspired by
applicable for small propagation distances, the key is to avoid the this method, we have proposed band-extended ASM, which
aliasing errors. significantly extends the effective bandwidth of ASM’s transfer
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1763

function without increasing sampling number. In our method, ( f 0 + m1 f )(x 0 + n1x ) = x 0 (m1 f + f 0 )
band extension is achieved by rearranging the sampling points
in the spatial frequency domain. Recalling inequality (31), m2 n2 (m − n)2
+ 1x 1 f + n f 0 1x + 1x 1 f − 1x 1 f .
we use a new sampling pitch 1 f _new = 2 f x _extend /2N rather 2 2 2
than 1 f = 1/2N1x to solve for the effective bandwidth (40)
[− f X _extend , f X _extend ]. After some simplifications, we get
Substituting Eq. (40) into Eq. (39), we get
r
N  
m2

f X _extend = . (38) U ( f x ) = U (m1 f ) = exp −i2π x 0 (m1 f + f 0 ) + 1x 1 f
2λz 2
It is easy to verify that f X _extend is much larger than f c with N−1
n2
  
large z values. Thanks to the band extension, band-extended X
× u(n1x ) exp −i2π n f 0 1x + 1x 1 f
ASM can be used over a much wider distance range compared n=0
2
with ordinary ASM or band-limited ASM. Because we changed
the sampling pitch in the spatial frequency domain, ASM’s (m − n)2
 
transfer function HASM ( f X ) is sampled at these spatial frequen- × exp i2π 1x 1 f , m = 0, 1, ...M − 1.
2
cies n1 f _new , where n = −N, −(N − 1) ... 0 ... N − 1. (41)
However, the input signal û(X ) is sampled at the spatial loca-
tions of n1 X . FFT cannot transform û(X ) to û( f X ), where Equation (41) is precisely a convolution of the two sequences
f X is the same as that of HASM ( f X ), because 1 f _new 6 = 1 f . a n and b n defined by
Therefore, we employed nonuniform FFT to achieve this oper-
n2
  
ation. For details of this method, please refer to [23]; the link a n = u(n1x ) exp −i2π n f 0 1x + 1x 1 f , (42)
for the MATLAB code can be found there. Fres-TF can be pro- 2
cessed in the same way. As for the RSC and Fres-IR, the methods
n2
 
of limiting or extending the effective bandwidth of the impulse b n = exp i2π 1x 1 f , (43)
response function can be similarly used. 2
Next, we focus on how to achieve scaling and zooming. with the output of the convolution multiplied by a phase
Scaling means observing the diffraction field with adjustable term b m ,
sampling pitch; and zooming means observing a region of inter-
m2
  
est, not the whole diffraction field. Also, there are some methods
proposed to realize these two functions, such as [58]. Here, we b m = exp −i2π x 0 (m1 f + f 0 ) + 1x 1 f . (44)
2
take ASM as an example to illustrate how to achieve these two
functions. In ASM, there are two key steps, which are Fourier That is,
transformation of the signal û(X ), and inverse Fourier trans- N−1
formation of the product Û ( f X )HASM ( f X ). Conventionally,
X
U ( f x ) = U (m1 f ) = b m a n b m−n , m = 0, 1, ...M − 1.
the Fourier transform and inverse Fourier transform are imple- n=0
mented by the FFT algorithm, which fixes the relationship (45)
between the sampling pitches in the space domain and the spa- Equation (45) can be calculated efficiently by FFT because
tial frequency domain as 1 f = 1/2N1x . Therefore, we need to it is in the form of a convolution. In the same way, we can cal-
find another tool to achieve scaling and zooming. Here, we use culate the inverse chirp-z transform. Therefore, we can choose
a scaled FFT, usually called the chirp-z transform [59,60]. We the coordinates of the original points, the number of sampling
introduce how to use the chirp-z transform to realize scaling and points and sampling pitches. Benefitting from these flexibil-
zooming in principle, and the algorithm implementation can ities, scaling and zooming can be achieved in the diffraction
then be done easily. calculation. The only cost of this method is that the computa-
Suppose (N, 1x ) and (M, 1 f ) are the sampling parameters tional complexity may increase because the Fourier transform is
in the space domain and spatial frequency domain, respectively. implemented using convolution. According to the convolution
Please note that they are independent, and the relationship theorem, three FFTs are required to calculate Eq. (45). There are
1 f = 1/2N1x no longer holds. We write the discrete Fourier also other mathematical tools, such as nonuniform FFT, that
transform of the input signal u(x ) in the following form: can be used for scaling and zooming [61,62].

N−1
X h
U ( f x ) = U (m1 f ) = u(n1x ) exp − i2π 5. SUMMARY AND OUTLOOK
n=0 To summarize, we have analyzed five popular fast diffraction
i calculation methods: the single FFT-based Fresnel transform,
× ( f 0 + m1 f )(x 0 + n1x ) , m = 0, 1, ...M − 1, the Fres-TF approach, the Fres-IR approach, the ASM, and
(39) the RSC, from the perspective of PSD and the sampling theo-
rem. The sampling property and the effect on the signal’s SBP
where f 0 and x 0 are the coordinates of the beginning points. for each method have been studied in detail and solutions to
( f 0 + m1 f )(x 0 + n1x ) can be expressed by the identity improve these methods are given.
 1764 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

Fig. 16. (a)–(c) The sampling SBP can be gradually reduced down to the signal’s SBP by subdivision along the f x axis.

If the final phase is required, single FFT-based Fresnel trans- of the parallelogram. A large amount of the sampling SBP is
form is only applicable at one specific propagation distance, due wasted, as indicated by the blank area.
to the opposite applicable distance ranges of two chirp func- Since we can control the spatial frequency range of the Fourier
tions. To overcome this restriction, we propose to apply FFT to transform, as introduced in Section 4, it is possible to perform
the product of the object and the inner chirp with more sam- subdivision along the f x axis, as shown in Fig. 16(b). That is
pling numbers, to satisfy the sampling requirement of the outer to say, we do not calculate the whole diffraction field at once,
chirp function. In addition, the bandwidth increase caused by but in three sections. First, we calculate the lower part, then the
chirp modulation is ignored, which would cause aliasing with middle part, and, finally, the upper part. In this way, the required
large-bandwidth signals. This problem should be quantitatively sampling SBP, the sum of three rectangular areas, can be greatly
studied for a practical solution. As convolution-based methods, reduced. Going one step further, if more subdivisions are made,
Fres-TF, Fres-IR, ASM, and RSC all need zero padding to avoid the required sampling SBP would trend to the signal’s SBP, as
circular convolution errors during the calculation. Aiming to shown in Fig. 16(c). Another potential way might be to model
overcome this problem, we have proposed an adaptive-sampling the diffraction in other geometric spaces, such as in Riemann
ASM to automatically satisfy the sampling requirement to avoid space. In the Euclidean space, there is horizontal shearing
circular convolution without zero padding [63]. Both ASM between the PSDs before and after Fresnel diffraction, which
and Fres-TF model diffraction with a transfer function in the increases the area of the circumscribed rectangle. The reason
spatial frequency domain and are applicable for small propa- why we have to use a rectangle as the measure of SBP rather other
gation distances. Both RSC and Fres-IR model the diffraction shapes in the phase space is because we are using a Cartesian
with impulse response function in the space domain and are coordinate system in Euclidean space. Therefore, it is inevitable
applicable for large propagation distances. The physical reason that we increase the sampling SBP to embrace the signal’s SBP
why they have different applicable distance ranges is given in after diffraction, and some of it is wasted. It may be possible to
Section 3.B, with Fig. 10. Besides, approaches to make these make the sampling SBP and the signal’s SBP always the same by
four methods applicable for a larger distance range have been using other geometric measures in a proper space; related work
analyzed and discussed in Section 4. Finally, we introduced how in Ref. [34] may be a good point. If so, the PSD should be in a
to use chirp-z transform, a kind of scaled FFT, to achieve scaling different form in such a situation.
and zooming in the numerical diffraction calculation. Overall, the task of numerical diffraction calculation is to
Furthermore, all these five methods are used to calculate efficiently and flexibly obtain the accurate diffraction field
Fresnel diffraction in free space. From the PSD shown in according to demand. Current methods can achieve one or two
this paper, we can see that the required sampling SBP for the of these three functions. A generalized diffraction calculation
diffraction field is larger than the signal’s SBP. Even though the
method that can simultaneously achieve these three functions
signal’s SBP does not change after diffraction, it spreads over a
needs to be developed in the future. Also, new diffraction theory
larger space, which requires a larger sampling SBP to describe
in different geometric spaces could be explored.
the diffraction field. Thus, not all the diffraction field can be
obtained by these methods when the sampling SBP is set to
a number equivalent to the signal’s SBP. In single FFT-based Funding. National Natural Science Foundation of
Fresnel transform, the spatial range of the calculated diffraction China (62035003, 61875105); National Key Research and
field is λz/1x , while the whole diffraction field spreads over Development Program of China (2017YFF0106400).
a range of N1x + λz/1x , where N1x is the spatial length
of the input signal. Considering it can be only used when Acknowledgment. The authors thank the anonymous
z = N12x /λ, the calculated spatial range is actually N1x . In the reviewers.
four convolution-based methods, the calculated spatial range is
always N1x , the same as that of the input signal. Besides, when Disclosures. The authors declare no conflicts of interest.
the propagation distance is relatively large, the bandwidth of the
calculated diffraction field is also smaller than that of the input
signal. To obtain the whole diffraction field and maintain the REFERENCES
bandwidth, the only solution, currently, is to increase the sam- 1. M. Born and E. Wolf, Principles of Optics: Electromagnetic Theory of
Propagation, Interference and Diffraction of Light (Elsevier, 2013).
pling SBP, which would increase the computational complexity,
2. J. W. Goodman, Introduction to Fourier Optics (W. H. Freeman,
as shown in Fig. 16(a). The signal’s SBP is indicated by the blue Macmillan Learning, 2017).
parallelogram, while the required sampling SBP is indicated by 3. A. Sommerfeld, “Mathematische theorie der diffraction,” Math. Ann.
the dashed-line rectangle, whose area is much larger than that 47, 317–374 (1896).
 Research Article Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A 1765

4. J. B. Keller, “Geometrical theory of diffraction,” J. Opt. Soc. Am. 52, 29. T. Shimobaba, T. Kakue, N. Okada, M. Oikawa, Y. Yamaguchi, and T.
116–130 (1962). Ito, “Aliasing-reduced Fresnel diffraction with scale and shift opera-
5. A. Rubinowicz, “V the Miyamoto-Wolf diffraction wave,” in Progress tions,” J. Opt. 15, 075405 (2013).
in Optics (Elsevier, 1965), pp. 199–240. 30. C. S. Guo, Y. Y. Xie, and B. Sha, “Diffraction algorithm suitable for
6. Y. Park, C. Depeursinge, and G. Popescu, “Quantitative phase imag- both near and far field with shifted destination window and oblique
ing in biomedicine,” Nat. Photonics 12, 578–589 (2018). illumination,” Opt. Lett. 39, 2338–2341 (2014).
7. D. G. Voelz, Computational Fourier Optics: A MATLAB Tutorial (SPIE, 31. K. Matsushima, “Shifted angular spectrum method for off-axis
2011). numerical propagation,” Opt. Express 18, 18453–18463 (2010).
8. Y. Z. Liu, F. A. South, Y. Xu, P. S. Carney, and S. A. Boppart, 32. K. Matsushima, H. Schimmel, and F. Wyrowski, “Fast calculation
“Computational optical coherence tomography [Invited],” Biomed. method for optical diffraction on tilted planes by use of the angular
Opt. Express 8, 1549–1574 (2017). spectrum of plane waves,” J. Opt. Soc. Am. A 20, 1755–1762 (2003).
9. W. Zhang, L. Cao, D. J. Brady, H. Zhang, J. Cang, H. Zhang, and G. 33. L. Onural, “Exact analysis of the effects of sampling of the scalar
Jin, “Twin-image-free holography: a compressive sensing approach,” diffraction field,” J. Opt. Soc. Am. A 24, 359–367 (2007).
Phys. Rev. Lett. 121, 093902 (2018). 34. H. M. Ozaktas, S. Ö. Arık, and T. Coşkun, “Fundamental structure of
10. W. Zhang, L. Cao, G. Jin, and D. Brady, “Full field-of-view digital lens- Fresnel diffraction: natural sampling grid and the fractional Fourier
free holography for weak-scattering objects based on grating modu- transform,” Opt. Lett. 36, 2524–2526 (2011).
lation,” Appl. Opt. 57, A164–A171 (2018). 35. H. M. Ozaktas, S. Ö. Arık, and T. Coşkun, “Fundamental structure of
11. W. Zhang, L. Cao, R. Li, H. Zhang, H. Zhang, Q. Jiang, and G. Jin, Fresnel diffraction: longitudinal uniformity with respect to fractional
“Wavefront division digital holography,” AIP Adv. 8, 055304 (2018). Fourier order,” Opt. Lett. 37, 103–105 (2012).
12. P. Marquet, B. Rappaz, P. J. Magistretti, E. Cuche, Y. Emery, T. 36. S. Ö. Arık and H. M. Ozaktas, “Optimal representation and process-
Colomb, and C. Depeursinge, “Digital holographic microscopy: a ing of optical signals in quadratic-phase systems,” Opt. Commun.
noninvasive contrast imaging technique allowing quantitative visu- 366, 17–21 (2016).
alization of living cells with subwavelength axial accuracy,” Opt. Lett. 37. B. M. Hennelly, J. J. Healy, and J. T. Sheridan, “Sampling in phase
30, 468–470 (2005). space,” in Phase-Space Optics Fundamentals and Applications,
13. H. Zhang, L. Cao, and G. Jin, “Three-dimensional computer- M. E. Testorf, B. M. Hennelly, and J. Ojeda-Castaneda, eds.
generated hologram with Fourier domain segmentation,” Opt. (McGraw-Hill, 2010).
Express 27, 11689–11697 (2019). 38. A. Stern and B. Javidi, “Sampling in the light of Wigner distribution,”
14. Y. Zhao, L. Cao, H. Zhang, D. Kong, and G. Jin, “Accurate calcu- J. Opt. Soc. Am. A 21, 360–366 (2004).
lation of computer-generated holograms using angular-spectrum 39. A. Stern and B. Javidi, “Improved-resolution digital holography using
layer-oriented method,” Opt. Express 23, 25440–25449 (2015). the generalized sampling theorem for locally band-limited fields,”
15. A. Jesacher, S. Bernet, and M. Ritsch-Marte, “Combined holographic J. Opt. Soc. Am. A 23, 1227–1235 (2006).
optical trapping and optical image processing using a single diffrac- 40. L. Onural, “Sampling of the diffraction field,” Appl. Opt. 39,
tive pattern displayed on a spatial light modulator,” Opt. Lett. 39, 5929–5935 (2000).
5337–5340 (2014). 41. A. W. Lohmann, R. G. Dorsch, D. Mendlovic, Z. Zalevsky, and C.
16. M. J. Padgett, “Orbital angular momentum 25 years on [Invited],” Ferreira, “Space–bandwidth product of optical signals and systems,”
Opt. Express 25, 11265–11274 (2017). J. Opt. Soc. Am. A 13, 470–473 (1996).
17. S. S. Kou and C. J. R. Sheppard, “Image formation in holographic 42. B. M. Hennelly and J. T. Sheridan, “Generalizing, optimizing, and
tomography: high-aperture imaging conditions,” Appl. Opt. 48, inventing numerical algorithms for the fractional Fourier, Fresnel, and
H168–H175 (2009). linear canonical transforms,” J. Opt. Soc. Am. A 22, 917–927 (2005).
18. Y. Sung, W. Choi, C. Fang-Yen, K. Badizadegan, R. R. Dasari, and M. 43. J. Serna, R. Martínez-Herrero, and P. M. Mejías, “Parametric charac-
S. Feld, “Optical diffraction tomography for high resolution live cell terization of general partially coherent beams propagating through
imaging,” Opt. Express 17, 266–277 (2009). ABCD optical systems,” J. Opt. Soc. Am. A 8, 1094–1098 (1991).
19. D. G. Voelz and M. C. Roggemann, “Digital simulation of scalar 44. P. A. Bélanger, “Beam propagation and the ABCD ray matrices,” Opt.
optical diffraction: revisiting chirp function sampling criteria and Lett. 16, 196–198 (1991).
consequences,” Appl. Opt. 48, 6132–6142 (2009). 45. S. A. Collins, “Lens-system diffraction integral written in terms of
20. D. P. Kelly, “Numerical calculation of the Fresnel transform,” J. Opt. matrix optics,” J. Opt. Soc. Am. 60, 1168–1177 (1970).
Soc. Am. A 31, 755–764 (2014). 46. T. Weihan, “Generality of optical ABCD theorem, its physical meaning
21. K. Matsushima and T. Shimobaba, “Band-limited angular spectrum and applications,” Chin. J. Quantum Electron. 21, 149–162 (2004).
method for numerical simulation of free-space propagation in far and 47. C. J. Sheppard and K. G. Larkin, “Wigner function for highly conver-
near fields,” Opt. Express 17, 19662–19673 (2009). gent three-dimensional wave fields,” Opt. Lett. 26, 968–970 (2001).
22. Y.-H. Kim, C.-W. Byun, H. Oh, J. Lee, J.-E. Pi, G. Heon Kim, M.-L. 48. C. J. Sheppard and K. G. Larkin, “Wigner function for nonparaxial
Lee, H. Ryu, H.-Y. Chu, and C.-S. Hwang, “Non-uniform sampling wave fields,” J. Opt. Soc. Am. A 18, 2486–2490 (2001).
and wide range angular spectrum method,” J. Opt. 16, 125710 49. C. Sheppard and K. Larkin, “Wigner function and ambiguity function
(2014). for non-paraxial three-dimensional wavefields,” in Optical Processing
23. W. Zhang, H. Zhang, and G. Jin, “Band-extended angular spectrum and Computing, 17-18 April 2001, Orlando, USA: A Tribute to Adolf
method for accurate diffraction calculation in a wide propagation Lohmann (SPIE, 2002), pp. 37–56.
range,” Opt. Lett. 45, 1543–1546 (2020). 50. C. J. Sheppard, “Diffraction optics,” in Handbook of Biomedical
24. T. Shimobaba, K. Matsushima, T. Kakue, N. Masuda, and T. Ito, Optics, D. Boas, C. Pitris, and N. Ramanujam, eds. (CRC Press,
“Scaled angular spectrum method,” Opt. Lett. 37, 4128–4130 (2012). 2011), pp. 11–31.
25. T. Kozacki and K. Falaggis, “Angular spectrum-based wave- 51. C. Sheppard and M. Hrynevych, “Diffraction by a circular aperture:
propagation method with compact space bandwidth for large a generalization of Fresnel diffraction theory,” J. Opt. Soc. Am. A 9,
propagation distances,” Opt. Lett. 40, 3420–3423 (2015). 274–281 (1992).
26. F. Zhang, I. Yamaguchi, and L. P. Yaroslavsky, “Algorithm for recon- 52. C. Sheppard, J. Lin, and S. Kou, “Rayleigh–Sommerfeld diffraction
struction of digital holograms with adjustable magnification,” Opt. formula in k space,” J. Opt. Soc. Am. A 30, 1180–1183 (2013).
Lett. 29, 1668–1670 (2004). 53. V. K. Ingle and J. G. Proakis, Digital Signal Processing Using Matlab: A
27. R. P. Muffoletto, J. M. Tyler, and J. E. Tohline, “Shifted Fresnel diffrac- Problem Solving Companion (Cengage Learning, 2016).
tion for computational holography,” Opt. Express 15, 5631–5640 54. J. B. Almeida and V. Lakshminarayanan, “Wide angle near-field
(2007). diffraction and Wigner distribution,” Optik 114, 333–335 (2003).
28. J. F. Restrepo and J. Garcia-Sucerquia, “Magnified reconstruction of 55. H. Zhang, Y. Zhao, L. Cao, and G. Jin, “Layered holographic
digitally recorded holograms by Fresnel–Bluestein transform,” Appl. stereogram based on inverse Fresnel diffraction,” Appl. Opt. 55,
Opt. 49, 6430–6435 (2010). A154–A159 (2016).
 1766 Vol. 37, No. 11 / November 2020 / Journal of the Optical Society of America A Research Article

56. T. Shimobaba and T. Ito, “Random phase-free computer-generated 60. L. Bluestein, “A linear filtering approach to the computation of
hologram,” Opt. Express 23, 9549–9554 (2015). discrete Fourier transform,” IEEE Trans. Audio Electroacoust. 18,
57. X. Yu, T. Xiahui, Q. Y. Xiong, P. Hao, and W. Wei, “Wide-window angu- 451–455 (1970).
lar spectrum method for diffraction propagation in far and near field,” 61. J.-Y. Lee and L. Greengard, “The type 3 nonuniform FFT and its appli-
Opt. Lett. 37, 4943–4945 (2012). cations,” J. Comput. Phys. 206, 1–5 (2005).
58. B. M. Hennelly, D. P. Kelly, D. S. Monaghan, and N. Pandey, 62. L. Greengard and J.-Y. Lee, “Accelerating the nonuniform fast Fourier
“Zoom algorithms for digital holography,” in Information Optics transform,” SIAM Rev. 46, 443–454 (2004).
and Photonics (Springer, 2010), pp. 187–204. 63. W. Zhang, H. Zhang, and G. Jin, “Adaptive-sampling angular spec-
59. L. Rabiner, R. Schafer, and C. Rader, “The chirp z-transform trum method with full utilization of space-bandwidth product,” Opt.
algorithm,” IEEE Trans. Audio Electroacoust. 17, 86–92 (1969). Lett. 45, 4416–4419 (2020).