Introduction to Fourier optics

Textbook: Goodman (chapters 4-5)

Overview:

Fourier transform Properties of lenses.

Imaging properties of lenses.
Frequency analysis of imaging systems
Coherent and incoherent imaging
Lenses as phase transformations

Let us now introduce what is perhaps the most important optical element into the
framework of Fourier optics – the thin lens. For simplicity, let's just consider a plano-
concave lens with radius of curvature R. The thickness of the lens is then:

 x2  y2 
 x y
2 2
x, y    0  R1  1 
 R2  0
2R
 

Where the last term is in the paraxial (small angle) approximation. Since the optical
pathlength difference between passing through the lens and passing in the vacuum
outside is proportional to (n-1), we immediately get that the phase transformation
imposed by the lens is:

 
t l x, y   exp  i
k 2
x  y2   
 2f 

Where f is defined as the focal length of the lens by the "lensmaker's formula":

 1 1 
 n  1 
1

f  R1 R2 

Let us now consider the combination of a lens with free-space propagation, having an
object illuminated by a plane wave, followed immediately by a lens of focal length f.

Let us derive again the Fresnel approximation. We start with:

e ikr01
U  x, y    
1
 U
i aperture
 ,
r01
cos dd

Which we approximate as:

e ikz  x 2  y  2 
ik
U x, y    U  , e 2z
dd
iz aperture

Taking the x2,y2 phase terms out of the expression, we get:

e ikz 2 z x 2  y 2   2  2   2 z x y 

ik ik ik
U x, y   e  U  , e 2z
e dd
iz aperture

Where we have approximated r01 as z in the denominator, neglected cos, and

approximated the phase term to lowest order.
Clearly, upon addition of the "correct" amount of quadratic phase at the aperture
plane, the Fresnel formula would be transformed into something equivalent to the
Fraunhoffer one.

Now, let us consider a lens of focal length f with an input placed immediately before
it.

After the lens, we get:

 
ik 2 2
  
U l  ,   At a  , P , e 2f

where P is the pupil aperture of the lens (a circ function).

Inserting this to the Fresnel formula we get:

  
 ik ik 
Ae ikz 2 z x 2  y 2     2
ik ik
x y 
2

U x, y   e  a
t  , P  , e  2z 2 f 
e 2z
dd
iz

which at z=f is just:

Ae ikf 2 f x 
ik ik
2
 y2  x y 
U x, y   e  t  , P , e
2f
dd
if
a

which is just the Fraunhoffer diffraction pattern – a Fourier transform multiplied by a

phase prefactor.

The integral in the above expression is simply, from the ASPW approach, the ASPW
 x y 
of the object ta in these parameters: F0  , 
 f  f 

Let us now consider the case where the object of interest is placed a distance d before
the lens. In this case we only have to propagate the object diffraction pattern a
distance z before applying the previous result.

The ASPW of the input, assuming a plane wave of amplitude A is:

F0  f x, f y   F At a 

Which after propagating a distance d to the lens becomes:


Fl  f x, f y   F0  f x, f y exp  id f x2  f y2  
From the previous result we now get:

Ae ikf 2 f x   x y 
ik 2
 y2
U  x, y   e Fl  , 
if   f f 
but substituting the two above equations we get:

Ae ikf 2 f  1 f  x   x y 

ik  d 2
 y2
U  x, y   e Fo  , 
if  f f 

Choosing d=f completely eliminates also the phase prefactor, leaving us with a perfect
Fourier transform.

This implies also that the following 4f system is a perfect 1:1 imager (also in the
phase sense – not just in intensity sense). This system will be used for a variety of
applications when we discuss image processing.

Relation to Basics of spectroscopy

In a previous talk we write the Fraunhofer solution for diffraction by an amplitude

grating. There, we say that the first order diffraction pattern is directed at an angle f0
to the incoming beam. Let us now position a grating in the back focal plane and
illuminate it with polychromatic light such that its central frequency is directed
towards the lens. For other values of l, this will (for a "narrow" enough spectrum) fan
out such that =f0. Thus the frequency spectrum is directly mapped onto the
angular spectrum, which is mapped onto the focal plane intensity distribution. Thus,
the Fourier transformation in space is mapped into a frequency domain spectrum. This
will be further discussed in the next tutorial. This is also of relevance to Fourier
domain pulse shaping which will be discussed in relation to all optical processing.

Magnification in Fourier transformation

More flexibility in the Fourier transform can be achieved if the object in placed
between the lens and the focal plane. In this case (see full derivation in Goodman) it
can be shown that the Fourier transform is magnified by a factor f/d.

Image formation by a lens

Let us write the electric field at the output of a lens using an impulse response:

U i u, v     hu, v;  , U o  , dd

For this to be an image we must require that the image of a point be a point, and that
distances be preserved up to magnifications:

hu, v;  ,   K u  M , v  M 

To find the impulse response let us use Fresnel formulae for propagation up to the
lens, include the quadratic phase of the lens, and propagate from the lens.

For a pointlike object:

 1 2 zoi  x    y   
ik 2 2

U l  x, y   e
iz o

To which we have to add the quadratic phase of the lens:

 
ik 2 2
x y 
U l x, y   U l x, y Px, y e 2f

Then we propagate a distance zi from the lens:

ik

u  x 2 v  y 2 
hu, v;  ,     U lx, y e
1 2 zi
dxdy
iz i

Overall this gives us:

  u  
ik
u 2
v 2  ik
 2  2  ik  1 1 1
   
 2 2
 x y  ik   
  v
 x  
 z z
 y

hu, v;  ,   2   Px, y e
1 2  zo zi f 
   zo zi   o i  
e 2 zi e 2 zoi
e dxdy
 zi zo

There are a lot of phase terms in this expression. Let us now see what is required in
order to have approximately a delta function in this expression.

The term outside the integration with the u,v coordinates is just a phase term at the
output, and vanishes if we look at the output intensity.

The quadratic term inside the integration is only eliminated if the prefactor is zero,
that is:

1 1 1
   0
zo zi f

Which is just the lens law as we derived in the ray optics picture.

Now we are left with a quadratic phase term dependent on the object coordinates.
This term is not easily eliminated. For the special case of an object illuminated by a
sonverging spherical wave, this can be exactly compensated for. Otherwise, the
requirement is such that the object dimensions are small enough so that this phase
term is small and can be neglected. In practice, this requires that the object be no
larger than ~1/4 of the size of the lens.

Assuming all these are satisfied, we are now left with the following expression for the
impulse response:

  u   v  
ik     x  
 z z
 y

hu, v;  ,   2   Px, y e
1   zo zi   o i  
dxdy
 zi zo

Defining the magnification as:

 zi
M 
zo

we finally get:


ik
u  M  x v  M  y 
hu, v;  ,     Px, y e
1 zi
dxdy
 zi zo
2

which is the result we actually wanted. Clearly the expression on the right is not a
delta function but is rather convolved with the pupil of the lens. This sets the basis for
the diffraction limit of an optical system.

For the rest of this discussion, it may be convenient to redefine the impulse response
in such a manner so as to remove the effect of magnification. To do so, we can
redefine the coordinates in the object space such that:

  M ;  M

Thus, we have:

ik
 
u  x  v   y 
hu , v;  ,   hu   , v     2


  Px, y e
A zi
dxdy
 zi zo

If we define the perfect image:

  
U g  ,  
1
U o  , 
M M M 

We then get:

U i u, v     hu   , v   U g  , 

with:


ik
ux vy 
hu, v   2   Px, y e
A zi
dxdy
 zi zo

So that for a diffraction limited optical system the image is a convolution of the
Gaussian (perfect) image predicted by geometrical optics with an impulse response
which is the Fraunhoffer diffraction pattern of the pupil.
Coherent and incoherent illumination

When analyzing the frequency response of an optical system we have to also consider
the illumination conditions. Until now we have generally assumed plane wave
illumination of the object, i.e. that the object is illuminated by a coherent wave. In
most cases, however, the object of interest is illuminated by spatially incoherent light.
In the context of optical microscopy this is also the case when observing incoherent
scattering excited by a coherent source (such as fluorescence). The basic result from
that is that scatterings from two different points in the object do not interfere with
each other, but only with themselves. We will now proceed to analyze the properties
of imaging systems through the use of transfer functions (applicable for the case of
invariant linear systems, which according to the above derivation applies for imaging
scaled up to magnification)

Coherent case

We have just seen, for the coherent case, that


ik
u  M  x  v  M  y 
hu, v;  ,     Px, y e
1 zi
dxdy
z i

When considering a small object, on the lens axis, we can assign a very clear meaning
to the integration coordinates x and y – relating them to spatial frequencies of the
object and of the image, such that:

x x
o  ; i 
zo zi

In this description it is also clear why the pupil function of the lens serves as a cutoff
in the spatial frequency space, with

d
fc 
2z i

for typical parameters (d=2cm,z=10cm, =0.5m the cutoff frequency is about

200/mm).

The transformation to the magnified coordinates is necessary to make the

transformation from object to image spatially invariant. Under these conditions, we
can write:

U i u, v     hu   , v   U g  , 

Corresponding, in Fourier space, to:

U i  f x , f y   H  f x , f y U o  f x , f y 
Where the transfer function is simply the pupil in the normalized coordinates:

ik
 
hu, v     Pz i f x , z i f y e df x df y  F Pz i f x , z i f y 
 f xu  f y v
zi

Which due to the symmetry of P can be written as (rather than the minus signs on the
argument):

H  f x , f y   F F Pzi f x , zi f y   Pzi f x , zi f y 

And has only values of either unity or zero.

Incoherent case

The incoherent case is a bit more complicated. We are dealing only with intensity
patterns, and thus we have to introduce an intensity transfer function.

We get:

I i u, v     hu   , v    I g  , d d

2

For intensity distributions, it is convenient to describe the Fourier conjugate in

normalized coordinates. This is since that for positive definite real functions, the
maximal value of the FT is at (0,0).

In Fourier domain, this corresponds to:

 
F I i   F h F I o 
2

rather than the coherent version:

F I i   F Ai   F hF Ao 

2 2

Again, working in normalized coordinates, and using the identities of FT we can

write:


H inc 
F h  2


  H  p, q H  p  f , q  f dpdq
*
x y

  hu, v  dudv   H  p, q  dpdq

2 2

Thus, the incoherent OTF is the normalized autocorrelation of the amplitude transfer
function.

Some properties immediately arise from the above definition:


H inc 0,0  1
   *
H inc  f x , f y   H inc  f x , f y 
 
H inc  f x , f y   H inc 0,0

The last property follows from Schwartz inequality with H,H*

2 2 2

  XYdpdq    X dpdq  Y
2 2
dpdq

In particular, since we saw that H, the (coherent) amplitude transfer function is just
the pupil function:

H  f x , f y   Pzi f x , zi f y 

We get that (since P is real):

 z i f x z i f y   z i f x z i f y 
   
P  x 
2
, y  
2  
P  x 
2
, y 
2
dxdy

H inc 
 P x, y dxdy
2

Geometrically, this means that we have to calculate the area of overlap of two shifted
pupils relative to the pupil area. This is very simple to calculate for a square aperture,
where the OTF has a triangular cross section.

From the above definition it also follows that for an aberration free system, the OTF is
positive definite too.

Aberrations and apodization

The above consideration was done for an aberration-free system, and assuming a
unity value of the transmission at the pupil. Both of these do not necessarily have to
be fulfilled.

Let us consider first apodization, which is a simpler effect. We saw that, for example,
using a circular aperture we get an Airy pattern, which has significant sidelobes. If the
pupil is amplitude modulated, we can trade off the bandwidth of the central lobe with
the intensity of the sidelobes. Let us compare two cases: a circular aperture, and one
apodized (multiplied) by a Gaussian. The latter has a broader FWHM but smaller
sidelobes. In the limit where the Gaussian width is much smaller than the circular
aperture, there are, in fact, no sidelobes, since the FT of a Gaussian is a Gaussian.

Alternatively, we can inverse apodize to enhance the higher frequency components,

by allowing more light transmission at the edges

The discussion of aberrations closely follows the one we used in the geometrical
optics case. There, we took the derivative of the phase shift at the output pupil in
order to find the geometrical shift of the imaged point as a function of the position it
passed the pupil. Here, we simply have to introduce this phase term into the frequency
response H to get the exact effect of aberrations. Thus, the pupil function P has now
an addition space-dependent phase term W(x,y)

H  f x , f y   Pzi f x , zi f y e

ikW zi f x ,zi f y 

Let us consider now the effects of aberrations on the OTF. For unapodized pupils, we
simply have, for eaxh frequency component, to integrate over the area of overlap of
the two shifted pupils. In this notation, for an aberrated image, we get:

  z f z i f y   z f z f 
ik W  x  i x , y   W  x  i x , y  i y
 


 e   2 2   2 2  
dxdy
 A f x , f y 
H inc  f x , f y  
 dxdy
A0 , 0 

From the Scwartz inequality, using the two exponents, it is clear that this reduces the
absolute value of the OTF. In particular, the OTF can become negative with
aberrations.

One can immediately note that:

1. aberrations do not affect the (0,0) component
2. Near the cutoff (where A is amall) aberrations do not affect the OTF
3. The effect of aberrations is negligible for W<<. In practice, the system is nearly
aberration free for W<~/4.

A simple example is a focusing error, which corresponds to a quadratic phase shift.

For a square aperture of width w, we can write this as:

W x, y  
1 z 2 2  1 w2  x 2  y 2

    
x2  y2    
x y  2  z Wm
2 zi2  2 zi  w
2
w2

This has a solution of the form:

   8W  fx    8W  fy  
H inc  f x , f y   H unaberrated  f x , f y sin c  m
f f
 1  x  sin c  m  1  y 
   
   2 fo  2 fo     2 fo  2 fo 

From this, we can readily see the above qualities:

Fow Wm<<l, the OTF is unchanged. The cutoff and the (0,0) values are unchanged.
In practice, however, the OTF has a low absolute value beyond the first zero of the
sinc, which occurs for fx such that:

8Wm f x  4
   fx  fo
2f o Wm
Alternatively, we could say that the image is undistorted as long as:

 1 w2    z i2
 2
z    z 
 2 zi  4 2 w2

For a focused Gaussian beam, this corresponds to the Rayleigh range

f2
R.r.   2
d

It should be noted that from this analysis it is clear that the high frequency
components are always depth resolved. We will get back to this point when we
discuss optical microscopy, in particular confocal microscopy.

Comparison of coherent and incoherent transfer functions

Frequency response:

As we saw, in the incoherent case the cutoff frequency was twice that of the coherent
case. The amplitude of the high frequency components, however, is smaller for the
incoherent case. Thus, the resolving power of either of these two cases can differ for
different samples. Let us consider, for example, two objects with the same amplitude
transmittance, but with a different phase transmittance:

t A  ,   cos 2f
~

t A  ,   cos 2f
~

~
The frequency spectrum of the first is two lobes at  f 2 , while for the second the
frequency spectrum it corresponds to a central lobe at zero frequency with two lobes
~
at  f . The first is transmitted in full by a coherent system, but we get reduced
contrast for an incoherent system. For the second, we get no contrast for the coherent
case, but some contrast for the incoherent one.

Resolution:

Assume the Rayleigh criterion (two points separated by 1.22). For the incoherent
case we get a small dip in the intensity response. For the coherent case this depends
on the phase of the two scatterings, and can go from full constructive interference,
where the peak appears at the center, to full destructive interference, where we get a
zero at the center. Clearly, the very definition of resolution in the coherent case is
more complicated.