MATH 262/CME 372: Applied Fourier Analysis and Winter 2021

Elements of Modern Signal Processing

Lecture 18 — March 16, 2021

Prof. Emmanuel Candes
Scribe: Carlos A. Sing-Long, Edited by E. Bates

1 Outline

Agenda: Lenses

1. Thin lenses
2. Fourier transform properties of thin lenses
3. Examples

Last Time: Using Rayleigh-Sommerfeld’s diffraction formula, we studied two regimes of diffrac-
tion. Fresnel diffraction occurs when the aperture plane and the screen are far apart relative to the
wavelength, but we are still able to observe near-field effects. In this case, the diffraction pattern
is, modulo a factor, the Fourier transform of the wave field on the aperture, after being modulated
by a quadratic phase term. On the other hand, Fraunhofer diffraction occurs when the aperture
plane and the screen are much further apart, and only far-field effects are observed. In this case
the quadratic phase term can be neglected, and the interference pattern is, modulo a modulation
by a quadratic phase term, the Fourier transform of just the wave field. To distinguish these two
regimes, we introduced the Fresnel number F , which separates Fresnel and Fraunhofer regimes by
F  1 and F  1, repsectively.

2 Thin lenses

Lenses are optical devices that transmit and refract light. In the most basic sense, the refraction
they induce is needed for selective transmission (see Fig. 1). The refraction is due to the change in
speed of propagation. For example,
n = 1 in vacuum (and ≈ 1 air), and n = 1.5 in glass.
In wave optics, the change in the speed of propagation through a glass lens introduces a phase-shift
to the wave field. This phase-shift will be proportional to the length of the path traversed by light.
In addition, if we have a thin lens the effects of propagation within the lens can be neglected.
Therefore, the wave field is only multiplied by a phase factor. This will be our working assumption.
To study this effect, we can enclose the lens on a box of width ∆0 , the thickness of the lens. If we
denote1 by ∆ = ∆(x) the thickness of the lens at x (see Fig. 2a), then the phase delay experienced
1
See Figure 1 in Lecture 16 for the coordinate system used.

1
Figure 1: If a sensing screen were placed away from the tree, the contributing optical wave fields would
come from all directions. To isolate just the light waves from the part of the environment containing
the tree, an optical system of lenses is placed an appropriate distance away, and the screen behind it. In
this lecture, we will discuss what the correct distance is, as well as why the image detected is rotated.

by the wave after traversing this box is

φ(x) = kn∆(x) + k(∆0 − ∆(x)) . (1)

| {z } | {z }
glass vacuum

Consequently, the difference between the incoming and outgoing wave fields uin and uout , respec-
tively, is
uout (x) = T (x)uin (x), where T (x) = eik∆0 eik(n−1)∆(x)
is known as the transmission function.
Of course, to understand the concrete effect of this we need to determine the thickness function
∆ of the lens. Here we will focus on circular lenses. In this case, each one of its surfaces is a
circular sector of a sphere. We will call the line joining the two centers the lens axis. We follow
the conventions that

1. If a ray encounters a convex surface (as in Fig. 2b), then radius of the corresponding circle is
taken to be positive.

2. If a ray encounters a concave surface (as in Fig. 2c), then the radius of the corresponding
circle is taken to be negative.

A lens such as the one shown in Fig. 2a is a converging, or doubly convex, lens. In this case,

1. The thickness of the left component of the lens is

s !
kxk2
∆1 (x) = ∆0,1 − R1 1− 1− ,
R12

where R1 > 0, and ∆0,1 > 0 is the maximum thickness of this component.

2
 incident incident
ray ray
R2
(x) x
R1

ez ez ez

(a) (b) (c)

Figure 2: (a) Coordinate system used to describe the thickness of the lens. (b) Convex surface with
R1 > 0. (c) Concave surface with R2 < 0.

2. The thickness of the right component of the lens is

s !
kxk2
∆2 (x) = ∆0,2 + R2 1− 1− ,
R22

where R2 < 0, and ∆0,2 > 0 is the maximum thickness of this component.

Consequently, the thickness function is given by

s ! s !
kxk2 kxk2
∆(x) = ∆1 (x) + ∆2 (x) = ∆0 − R1 1− 1− + R2 1− 1− .
R12 R22

As we are interested in the paraxial case (i.e. the light field near the lens axis), we may think of
kxk as small relative to the radii R1 and R2 . In this case, the Taylor expression
p x2

1 − x2 = 1 − + o x4
2
suggests the following paraxial approximations:
s s
kxk2 kxk2 kxk2 kxk2
1− ≈ 1 − and 1 − ≈ 1 − ,
R12 2R12 R22 2R22

These simplifications are accurate whenever

kxk kxk
1 and  1,
R1 R2
Using our approximations, we write the lens thickness as
 
1 1 1
∆(x) = ∆0 − − kxk2 .
2 R1 R2

3
The physical properties of the lens can then be encoded by a single number f , called the lens focal
length:  
1 1 1
= (n − 1) − .
f R1 R2
Under the paraxial approximation, the phase-shift (1) induced by the lens is now written
φ(x) = kn∆(x) + k(∆0 − ∆(x))
= k∆0 + k(n − 1)∆(x)
 
1
= k∆0 + k(n − 1) ∆0 − kxk2
2f (n − 1)
k
= kn∆0 − kxk2 .
2f
The transmission function, which introduces this phase delay, is then
− ik kxk2
T (x) = eikn∆0 |e 2f{z } .
important part

For instance, if the incident signal uin is a unit amplitude plane wave, then the outgoing wave, on
the other side of the lens, is
ik
− 2f kxk2
uout (x) ∝ e , (2)
which is a quadratic approximation to a spherical wave.
The sign of f indicates the position of the lens relative to the source of this wave, and so (2)
corresponds to the time-reversal of an expanding spherical wave with source at the origin, distance
f away. That is, the converging lens will focus a plane wave as an incoming spherical wave
converging in front of the lens, on the lens axis and at the focal distance f > 0. A similar result
holds for a diverging, or doubly concave lens, for which the focal length is negative. In that
case, the lens will focus a plane wave as an outgoing spherical wave with source behind the lens,
again on the lens axis, at the focal length f < 0. Fig. 3 illustrates the difference between these
cases.

3 Fourier transform properties of thin lenses

Suppose we place a converging lens on the aperture. In our previous study we did not account
for the fact that the lens is finite. So here we define the pupil function P , which is simply the
indicator function of the shape of the lens. If we denote the wave field just behind the lens aperture
as uin , then the wave field immediately in front of the lens aperture is
uout (x) = P (x)T (x)uin (x).

After traversing the lens, the wave field experiences diffraction. If we are interested in near-field
effects, Fresnel’s diffraction integral yields
Z
1 ikz ik kx0 k2 ik 2 ik
u(x0 ) = e e 2z uout (x)e 2z kxk e− z hx, x0 i dx
iλz
ZA
1 ikz ik kx0 k2 ik 2 ik
= e e 2z uin (x)P (x)T (x)e 2z kxk e− z hx, x0 i dx.
iλz

4
 | {z } | {z }
f f

(a) converging lens (b) diverging lens

Figure 3: Plane wave encountering a converging (doubly convex) or diverging (doubly concave) lens.
The focal point (blue dot) is along the lens axis (dashed line). The wave front (orange) is focused
inward in (a) and outward in (b), in either case as a spherical wave with source at the focal point. The
thin lens approximation has been made, and the finite spatial extension of the lens has been ignored.

Using the paraxial approximation yields

ik
ik 2 ik − 2f kxk2 ik 2 ik
T (x)e 2z kxk e− z hx, x0 i = eikn∆0 e e 2z kxk e− z hx, x0 i
= eikn∆0 e
ik
− 2f (1− fz )kxk2 − ikz hx, x0 i
e .

Therefore, if the screen is placed at the focal distance z = f , exponential terms exactly cancel, and
the diffraction pattern becomes
Z
1 ik(f +n∆0 ) 2f
ik
kx0 k2 − ik hx, x0 i
u(x0 ) = e e P (x)uin (x)e f dx.
iλf

Therefore, if the incident wave field is localized in the lens (that is, the physical extent of the input
is much smaller than the lens aperture, such as in a microscope or a telescope examining stars),
then we will observe its Fourier transform, up to modulation by a quadratic phase factor. Note
that this modulation is experimentally unimportant since we see only intensity and not phase.
Otherwise, we will observe the Fourier transform of the incoming wave convolved with the Fourier
transform of the pupil function. Recall that multiplication in the spatial domain is equivalent to
convolution in the frequency domain. This convolution kernel, the Fourier transform of the pupil
function, is called the point-spread function (PSF). In the case of a circular lens, this is the Airy
disk. The key feature of thin lenses is that they allow us to see the effects of diffraction at much
shorter distances.
We can consider more general arrays including lenses. Suppose we have an array as the one shown
in Fig. 4a with z1 and z2 assumed to be arbitrary at first. Once again, we will neglect the finite
extent of the apertures so that the pupil function is not needed. If the incoming wave field is u0 ,
then the wave field just before the lens is
Z
1 ikz1 ik
kx−x0 k2
uin (x0 ) = e u0 (x)e 2z1 dx,
iλz1

5
 incoming screen
wave field lens

u0 / û0

z1 = f z2 = f

(a) 2-f optical system

incoming / û0 screen

wave field lens lens

u0 / ũ0

f f f f

(b) 4-f optical system

Figure 4: Example of optical systems. (a) This system allows us to observe the Fourier transform of
the incoming wave field up to a multiplicative factor. (b) By replicating the 2-f system, we can obtain
the incoming wave field up to a multiplicative factor and a rotation about the origin. Here ũ0 represents
a rotation of u0 with respect to the origin.

and the wave field just after the lens is

uout (x) = T (x)uin (x).

And then the wave field observed on the screen is

Z
1 ikz2 ik
kx−x0 k2
u(x0 ) = e uout (x)e 2z2 dx
iλz2
ZZ
1 ik
− ik kxk2 2z ik
kx−x0 k2 2z kx−x0 k2
=− 2 eik(z1 +z2 +n∆0 ) u0 (x0 )e 2f e 1 e 2 dxdx0 .
λ z1 z 2

Notice that the phase term in the integral can be written as

     
k 1 1 1 2 1 0 1 1 0 2 1 2
− − − kxk − 2 x, x + x0 + kx k + kx0 k .
2 f z 1 z2 z1 z2 z1 z2

For simplicity, suppose z1 = z2 = f . (One can deduce this, but this assumption greatly simplifies
the computations.) Then we can write the above phase term as
k   k  
kxk2 − 2hx, x0 + x0 i + kx0 k2 + kx0 k2 = kx − x0 − x0 k2 − 2hx0 , x0 i ,
2f 2f

6
and therefore we observe
Z Z 
1 ik
− 2f kx−x0 −x0 k2 − ik hx0 , x0 i
u(x0 ) = − 2 2 eik(2f +n∆0 ) e dx u0 (x0 )e f dx0 .
λ f

However, since we are ignoring the finite extent of the apertures, we can perform a change of
variables to obtain
Z Z r
ik
− 2f kx−x0 −x0 k2 ik
− 2f kxk2 − iπ πf
e dx = e dx = e 4 ,
k
and deduce
r Z
πf 1 − iπ ik(2f +n∆0 ) − ik hx0 , x0 i
u(x0 ) = − 2 2
e 4 e u0 (x0 )e f dx0
k λ f
r  
πf 1 − iπ ik(2f +n∆0 ) k
=− e 4 e û0 x ,
k λ2 f 2 f 0

which is yet another Fourier transform!

Unfortunately, in practice we have only the physical means of measuring the intensity of a wave
field; phase information is lost. This is not a problem if we are observing the desired, final image
(since our eyes only detect intensity), but thus far we have only shown how to observe its Fourier
transform. Our rescue is the fact the Fourier transform is nearly its own inverse. Indeed, in
two dimensions the Fourier transform of fˆ is the scaled rotation (2π)2 f (−x). So we consider
the extended 4-f system, which replicates the setup of the 2-f system to take a second Fourier
transform. The resulting output is a scaled and rotated version of the input. In the next lecture
we will study simpler ways of achieving the same result.