Chap8 PDF

Chapter 8
n-dimensional Fourier Transform

8.1
Space, the Final Frontier
To quote Ron Bracewell from p. 119 of his book Two-Dimensional Imaging, In two dimensions phenomena
are richer than in one dimension. True enough, working in two dimensions offers many new and rich
possibilities. Contemporary applications of the Fourier transform are just as likely to come from problems in
two, three, and even higher dimensions as they are in one imaging is one obvious and important example.
To capitalize on the work weve already done, however, as well as to highlight differences between the onedimensional case and higher dimensions, we want to mimic the one-dimensional setting and arguments
as much as possible. It is a measure of the naturalness of the fundamental concepts that the extension
to higher dimensions of the basic ideas and the mathematical definitions that weve used so far proceeds
almost automatically. However much well be able to do in class and in these notes, you should be able to
read more on your own with some assurance that you wont be reading anything too much different from
what youve already read.
Notation The higher dimensional case looks most like the one-dimensional case when we use vector
notation. For the sheer thrill of it, Ill give many of the definitions in n dimensions, but to raise the comfort
level well usually look at the special case of two dimensions in more detail; two and three dimensions are
where most of our examples will come from.
Well write a point in Rn as an n-tuple, say
x = (x1 , x2 , . . . , xn ) .
Note that were going back to the usual indexing from 1 to n. (And no more periodic extensions of the
n-tuples either!) Well be taking Fourier transforms and may want to assign a physical meaning to our
variables, so we often think of the xi s as coordinates in space, with the dimension of length, and x as
the spatial variable. Well then also need an n-tuple of frequencies, and without saying yet what
frequency means, well (typically) write
= (1 , 2 , . . . , n )
for those variables dual to x. Recall that the dot product of vectors in Rn is given by
x = x1 1 + x2 2 + + xn n .
The geometry of Rn is governed by the dot product, and using it will greatly help our understanding as
well as streamline our notation.
336
8.1.1
Chapter 8 n-dimensional Fourier Transform
The Fourier transform
We started this course with Fourier series and periodic phenomena and went on from there to define the
Fourier transform. Theres a place for Fourier series in higher dimensions, but, carrying all our hard won
experience with us, well proceed directly to the higher dimensional Fourier transform. Ill save Fourier
series for a later section that includes a really interesting application to random walks.
How shall we define the Fourier transform? We consider real- or complex-valued functions f defined on
Rn , and write f (x) or f (x1 , . . . , xn ), whichever is more convenient in context. The Fourier transform of
f (x) is the function Ff (), or f(), defined by
Z
e2ix f (x) dx .
Ff () =
Rn
The inverse Fourier transform of a function g() is

Z
F 1 g(x) =
e2ix g() d .
The Fourier transform, or the inverse transform, of a real-valued function is (in general) complex valued.
The exponential now features the dot product of the vectors x and ; this is the key to extending the
definitions from one dimension to higher dimensions and making it look like one dimension. The integral
is over all of Rn , and as an n-fold multiple integral all the xj s (or j s for F 1 ) go from to . Realize
that because the dot product of two vectors is a number, were integrating a scalar function, not a vector
function. Overall, the shape of the definitions of the Fourier transform and the inverse transform are the
same as before.
The kinds of functions to consider and how they enter into the discussion Schwartz functions, L1 , L2 , etc.
is entirely analogous to the one-dimensional case, and so are the definitions of these types of functions.
Because of that we dont have to redo distributions et al. (good news), and Ill seldom point out when this
aspect of the general theory is (or must be) invoked.
Written out in coordinates, the definition of the Fourier transform reads:
Z
e2i(x1 1 ++xn n ) f (x1 , . . . , xn ) dx1 . . . dxn ,
Ff (1 , 2 , . . . , n ) =
Rn
so for two dimensions,

Ff (1 , 2 ) =
e2i(x1 1 +x2 2 ) f (x1 , x2 ) dx1 dx2 .
The coordinate expression is manageable in the two-dimensional case, but I hope to convince you that its
almost always much better to use the vector notation in writing formulas, deriving results, and so on.
Arithmetic with vectors, including the dot product, is pretty much just like arithmetic with numbers.
Consequently, all of the familiar algebraic properties of the Fourier transform are present in the higher
dimensional setting. We wont go through them all, but, for example,
Z
Z
e2ix f (x) dx = F 1 f () ,
e2ix() f (x) dx =
Ff () =
Rn
Rn
which is one way of stating the duality between the Fourier and inverse Fourier transforms. Here, recall
that if = (1 , . . . , n ) then
= (1 , . . . , n ) .
8.1
337
To be neater, we again use the notation

f () = f () ,
and with this definition the duality results read exactly as in the one-dimensional case:
Ff = (Ff ) ,
(Ff ) = F 1 f
In connection with these formulas, I have to point out that changing variables, one of our prized techniques
in one dimension, can be more complicated for multiple integrals. Well approach this on a need to know
basis.
Its still the case that the complex conjugate of the integral is the integral of the complex conjugate, so
when f (x) is real valued,
Ff () = Ff () .
Finally, evenness and oddness are defined exactly as in the one-dimensional case. That is:
f (x) is even if f (x) = f (x), or without writing the variables, if f = f .
f (x) is odd f () = f (), or f = f .
Of course, we no longer have quite the easy geometric interpretations of evenness and oddness in terms of a
graph in the higher dimensional case as we have in the one-dimensional case. But as algebraic properties of
a function, these conditions do have the familiar consequences for the higher dimensional Fourier transform,
e.g., if f (x) is even then Ff () is even, if f (x) is real and even then Ff () is real and even, etc. You could
write them all out. I wont.
Soon enough well calculate the Fourier transform of some model functions, but first lets look a little bit
more at the complex exponentials in the definition and get a better sense of what the spectrum means
in higher dimensions.
Harmonics, periodicity, and spatial frequencies The complex exponentials are again the building
blocks the harmonics for the Fourier transform and its inverse in higher dimensions. Now that they
involve a dot product, is there anything special we need to know?
As mentioned just above, we tend to view x = (x1 , . . . , xn ) as a spatial variable and = (1 , . . . , n )
as a frequency variable. Its not hard to imagine problems where one would want to specify n spatial
dimensions each with the unit of distance, but its not so clear what an n-tuple of frequencies should mean.
One thing we can say is that if the spatial variables (x1 , . . . , xn ) do have the dimension of distance then
the corresponding frequency variables (1 , . . . , n ) have the dimension 1/distance. For then
x = x1 1 + + xn n
is dimensionless and exp(2ix ) makes sense. This corresponds to dimensions of time and 1/time in
the one-dimensional time domain and frequency domain picture.
For some further insight lets look at the two-dimensional case. Consider
exp(2ix ) = exp(2i(x1 1 + x2 2 )) .
338
(It doesnt matter for the following discussion whether we take + or in the exponent.) The exponent
equals 1 whenever x is an integer, that is, when
1 x1 + 2 x2 = n,
n an integer .
With = (1 , 2 ) fixed this is a condition on (x1 , x2 ), and one says that the complex exponential has zero
phase whenever 1 x1 + 2 x2 is an integer. This terminology comes from optics.
Theres a natural geometric interpretation of the zero phase condition thats very helpful in understanding
the most important properties of the complex exponential. For a fixed the equations
1 x1 + 2 x2 = n
determine a family of parallel lines in the (x1 , x2 )-plane (or in the spatial domain if you prefer that phrase).
Take n = 0. Then the condition on x1 and x2 is
1 x1 + 2 x2 = 0
and we recognize this as the equation of a line through the origin with (1 , 2 ) as a normal vector to the
line.1 (Remember your vectors!) Then (1 , 2 ) is a normal to each of the parallel lines in the family. One
could also describe the geometry of the situation by saying that the lines each make an angle with the
x1 -axis satisfying
2
,
tan =
1
but I think its much better to think in terms of normal vectors to specify the direction the vector point
of view generalizes readily to higher dimensions, as well discuss.
Furthermore, the family of lines 1 x1 + 2 x2 = n are evenly spaced as n varies; in fact, the distance between
the line 1 x1 + 2 x2 = n and the line 1 x1 + 2 x2 = n + 1 is
distance =
1
1
.
=p 2
kk
1 + 22
Ill let you derive that. This is our first hint, in two dimensions, of a reciprocal relationship between the
spatial and frequency variables:
The spacing of adjacent lines of zero phase is the reciprocal of the length of the frequency vector.
Drawing the family of parallel lines with a fixed normal also gives us some sense of the periodic nature
of the harmonics exp(2i x ). The frequency vector = (1 , 2 ), as a normal to the lines, determines
p
how the harmonic is oriented, so to speak, and the magnitude of , or rather its reciprocal, 1/ 12 + 22
determines the period of the harmonic. To be precise, start at any point (a, b) and move in the direction
of the unit normal, /kk. That is, move from (a, b) along the line
x(t) = (x1 (t), x2 (t)) = (a, b) + t
kk
or
x1 (t) = a + t
1
2
, x2 (t) = b + t
kk
kk
at unit speed. The dot product of x(t) and is

x(t) = (x1 (t), x2 (t)) (1 , 2 ) = a1 + b2 + t
12 + 22
= a1 + b2 + tkk ,
kk
Note that (1 , 2 ) isnt assumed to be a unit vector, so its not the unit normal.
8.1
339
and the complex exponential is a function of t along the line:

exp(2i x ) = exp(2i(a1 + b2 )) exp(2itkk) .
The factor exp(2i(a1 + b2 )) doesnt depend on t and the factor exp(2itkk) is periodic with period
1/kk, the spacing between the lines of zero phase. Now, if 1 or 2 is large, then the spacing of the lines is
close and, by the same token, if 1 and 2 are small then the lines are far apart. Thus although frequency
is now a vector quantity we still tend to speak in terms of a high frequency harmonic, when the lines
of zero phase are spaced close together and a low frequency harmonic when the lines of zero phase are
spaced far apart (high and low are relatively speaking, of course). Half way between the lines of zero
phase, when t = 1/2kk, were on lines where the exponential is 1, so 180 out of phase with the lines of
zero phase.
One often sees pictures like the following.
340
Heres what youre looking at: The function e2ix is complex valued, but consider its real part

Re e2ix = 21 e2ix + e2ix
= cos 2ix = cos 2(1 x1 + 2 x2 )
which has the same periodicity and same lines of zero phase as the complex exponential. Put down white
stripes where cos 2(1 x1 + 2 x2 ) 0 and black stripes where cos 2(1 x1 + 2 x2 ) < 0, or, if you want to
get fancy, use a gray scale to go from pure white on the lines of zero phase, where the cosine is 1, down to
pure black on the lines 180 out of phase, where the cosine is 1, and back up again. This gives a sense
of a periodically varying intensity, and the slowness or rapidity of the changes in intensity indicate low or
high spatial frequencies.
The spectrum The Fourier transform of a function f (x1 , x2 ) finds the spatial frequencies (1 , 2 ). The
set of all spatial frequencies is called the spectrum, just as before. The inverse transform recovers the
function from its spectrum, adding together the corresponding spatial harmonics, each contributing an
amount Ff (1 , 2 ). As mentioned above, when f (x1 , x2 ) is real we have
Ff (1 , 2 ) = Ff (1 , 2 ) ,
so that if a particular Ff (1 , 2 ) is not zero then there is also a contribution from the negative frequency
(1 , 2 ). Thus for a real signal, the spectrum, as a set of points in the (1 , 2 )-plane, is symmetric about
the origin.2 If we think of the exponentials of corresponding positive and negative frequency vectors adding
up to give the signal then were adding up (integrating) a bunch of cosines and the signal really does seem
to be made of a bunch of a stripes with different spacings, different orientations, and different intensities
2
N.b.: Its not the values Ff (1 , 2 ) that are symmetric, just the set of points (1 , 2 ) of contributing frequencies.
8.1
341
(the magnitudes |Ff (1 , 2 )|). It may be hard to imagine that an image, for example, is such a sum of
stripes, but, then again, why is music the sum of a bunch of sine curves?
In the one-dimensional case we are used to drawing a picture of the magnitude of the Fourier transform
to get some sense of how the energy is distributed among the different frequencies. We can do a similar
thing in the two-dimensional case, putting a bright (or colored) dot at each point (1 , 2 ) that is in the
spectrum, with a brightness proportional to the magnitude |Ff (1 , 2 )|. This, the energy spectrum or the
power spectrum, is symmetric about the origin because |Ff (1 , 2 )| = |Ff (1 , 2 )|.
Here are pictures of the spatial harmonics we showed before and their respective spectra.
Which is which? The stripes have an orientation (and a spacing) determined by = (1 , 2 ) which is normal
to the stripes. The horizontal stripes have a normal of the form (0, 2 ) and they are of lower frequency so
2 is small. The vertical stripes have a normal of the form (1 , 0) and are of a higher frequency so 1 is
large, and the oblique stripes have a normal of the form (, ) with a spacing about the same as for the
vertical stripes
Heres a more interesting example.3
For the picture of the woman, what is the function we are taking the Fourier transform of ? The function
f (x1 , x2 ) is the intensity of light at each point (x1 , x2 ) thats what a black-and-white image is for the
purposes of Fourier analysis. Incidentally, because the dynamic range (the range of intensities) can be so
large in images its common to light up the pixels in the spectral picture according to the logarithm of the
intensity.
Heres a natural application of filtering in the frequency domain for an image.
The first picture shows periodic noise that appears quite distinctly in the frequency spectrum. We eliminate
those frequencies and take the inverse transform to show the plane more clearly.4
Finally, there are reasons to add things to the spectrum as well as take them away. An important and
relatively new application of the Fourier transform in imaging is digital watermarking. Watermarking is an
old technique to authenticate printed documents. Within the paper an image is imprinted (somehow I
dont know how this is done!) that only becomes visible if held up to a light or dampened by water. The
3
I showed this picture to the class a few years ago and someone yelled : Thats Natalie!
All of these examples are taken from the book Digital Image Processing by G. Baxes.
342
idea is that someone trying to counterfeit the document will not know of or cannot replicate the watermark,
but that someone who knows where to look can easily verify its existence and hence the authenticity of the
8.1
343
document. The newer US currency now uses watermarks, as well as other anticounterfeiting techniques.
For electronic documents a digital watermark is added by adding to the spectrum. Insert a few extra
harmonics here and there and keep track of what you added. This is done in a way to make the changes in
the image undetectable (you hope) and so that no one else could possibly tell what belongs in the spectrum
and what you put there (you hope). If the receivers of the document know where to look in the spectrum
they can find your mark and verify that the document is legitimate.
Higher dimensions In higher dimensions the words to describe the harmonics and the spectrum are
pretty much the same, though we cant draw the pictures5 . The harmonics are the complex exponentials
e2ix and we have n spatial frequencies, = (1 , 2 , . . . , n ). Again we single out where the complex
exponentials are equal to 1 (zero phase), which is when x is an integer. In three-dimensions a given
(1 , 2 , 3 ) defines a family x = integer of parallel planes (of zero phase) in (x1 , x2 , x3 )-space. The
normal to any of the planes is the vector = (1 , 2 , 3 ) and adjacent planes are a distance 1/kk apart.
The exponential is periodic in the direction with period 1/kk. In a similar fashion, in n dimensions
we have families of parallel hyperplanes ((n 1)-dimensional planes) with normals = (1 , . . . , n ), and
distance 1/kk apart.
8.1.2
Finding a few Fourier transforms: separable functions
There are times when a function f (x1 , . . . , xn ) of n variables can be written as a product of n functions of
one-variable, as in
f (x1 , . . . , xn ) = f1 (x1 )f2 (x2 ) fn (xn ) .
Attempting to do this is a standard technique in finding special solutions of partial differential equations
there its called the method of separation of variables. When a function can be factored in this way, its
Fourier transform can be calculated as the product of the Fourier transform of the factors. Take n = 2 as
a representative case:
Z
e2ix f (x) dx
Ff (1 , 2 ) =
Rn
Z Z
e2i(x1 1 +x2 2 ) f (x1 , x2 ) dx1 dx2
=

Z Z
e2i1 x1 e2i2 x2 f1 (x1 )f2 (x2 ) dx1 dx2
=

Z Z
2i1 x1
=
e
f1 (x) dx1 e2i2 x2 f2 (x2 ) dx2
Z
e2i2 x2 f2 (x2 ) dx2
= Ff1 (1 )
= Ff1 (1 ) Ff2 (2 )
In general, if f (x1 , x2 , . . . , xn ) = f1 (x1 )f2 (x2 ) fn (xn ) then
Ff (1 , x2 , . . . n ) = Ff1 (1 )Ff2 (2 ) Ffn (n ) .
If you really want to impress your friends and confound your enemies, you can invoke tensor products in
this context. In mathematical parlance the separable signal f is the tensor product of the functions fi and
5
Any computer graphics experts out there care to add color and 3D-rendering to try to draw the spectrum?
344
one writes
f = f1 f2 fn ,
and the formula for the Fourier transform as
F(f1 f2 fn ) = Ff1 Ff2 Ffn .
People run in terror from the symbol. Cool.
Higher dimensional rect functions The simplest, useful example of a function that fits this description
is a version of the rect function in higher dimensions. In two dimensions, for example, we want the function
that has the value 1 on the square of side length 1 centered at the origin, and has the value 0 outside this
square. That is,
(
1 12 < x1 < 21 , 21 < x2 < 12
(x1 , x2 ) =
0 otherwise
You can fight it out how you want to define things on the edges. Heres a graph.
We can factor (x1 , x2 ) as the product of two one-dimensional rect functions:

(x1 , x2 ) = (x1 )(x2 ) .
(Im using the same notation for the rect function in one or more dimensions because, in this case, theres
little chance of confusion.) The reason that we can write (x1 , x2 ) this way is because it is identically
1 if all the coordinates are between 1/2 and 1/2 and it is zero otherwise so its zero if any of the
coordinates is outside this range. Thats exactly what happens for the product (x1 )(x2 ).
8.1
345
For the Fourier transform of the 2-dimensional we then have

F(1 , 2 ) = sinc 1 sinc 2 .
Heres what the graph looks like.
A helpful feature of factoring the rect function this way is the ability, easily, to change the widths in the
different coordinate directions. For example, the function which is 1 in the rectangle a1 /2 < x1 < a1 /2,
a2 /2 < x2 < a2 /2 and zero outside that rectangle is (in appropriate notation)
a1 a2 (x1 , x2 ) = a1 (x1 )a2 (x2 ) .
The Fourier transform of this is
Fa1 a2 (1 , 2 ) = (a1 sinc a1 1 )(a2 sinc a2 2 ) .
Heres a plot of (2 sinc 21 )(4 sinc 42 ). You can see how the shape has changed from what we had before.
346
The direct generalization of the (basic) rect function to n dimensions is

(
1 21 < xk < 12 , k = 1, . . . , n
(x1 , x2 , . . . , xn ) =
0 otherwise
which factors as
(x1 , x2 , . . . , xn ) = (x1 )(x2 ) (xn ) .
For the Fourier transform of the n-dimensional we then have
F(1 , 2 , . . . , n ) = sinc 1 sinc 2 sinc n .
Its obvious how to modify higher-dimensional to have different widths on different axes.
Gaussians Another good example of a separable function one that often comes up in practice is
a Gaussian. By analogy to the one-dimensional case, the most natural Gaussian to use in connection with
Fourier transforms is
2
2
2
2
g(x) = e|x| = e(x1 +x2 ++xn ) .
This factors as a product of n one-variable Gaussians:
2
g(x1 , . . . , xn ) = e(x1 +x2 ++xn ) = ex1 ex2 exn = h(x1 )h(x2 ) h(xn ) ,
where
h(xk ) = exk .
Taking the Fourier transform and applying the one-dimensional result (and reversing the algebra that we
did above) gets us
2
Fg() = e1 e2 en = e(1 +2 ++n ) = e|| .
8.2
Getting to Know Your Higher Dimensional Fourier Transform
347
As for one dimension, we see that g is its own Fourier transform.

Heres a plot of the two-dimensional Gaussian.
8.2
You already know a lot about the higher dimensional Fourier transform because you already know a lot
about the one-dimensional Fourier transform thats the whole point. Still, its useful to collect a few of
the basic facts. If some result corresponding to the one-dimensional case isnt mentioned here, that doesnt
mean it doesnt hold, or isnt worth mentioning it only means that the following is a very quick and
very partial survey. Sometimes well work in Rn , for any n, and sometimes just in R2 ; nothing should be
read into this for or against n = 2.
8.2.1
Linearity
Linearity is obvious:
F(f + g)() = Ff () + Fg() .
8.2.2
Shifts
In one dimension a shift in time corresponds to a phase change in frequency. The statement of this is the
shift theorem:
If f (x) F (s) then f (x b) e2isb F (s).
It looks a little slicker (to me) if we use the delay operator (b f )(x) = f (x b), for then we can write
F(b f )(s) = e2isb Ff (s) .
348
(Remember, b involves b.) Each to their own taste.

The shift theorem in higher dimensions can be made to look just like it does in the one-dimensional
case. Suppose that a point x = (x1 , x2 , . . . , xn ) is shifted by a displacement b = (b1 , b2 , . . . , bn ) to
x + b = (x1 + b1 , x2 + b2 , . . . , xn + bn ). Then the effect on the Fourier transform is:
The Shift Theorem
If f (x) F () then f (x b) e2ibF ().
Vectors replace scalars and the dot product replaces multiplication, but the formulas look much the same.
Again we can introduce the delay operator, this time delaying by a vector:
b f (x) = f (x b) ,
and the shift theorem then takes the form
F( b f )() = e2ibFf () .
(Remember, b involves a b.) Each to their own taste, again.
If youre more comfortable writing things out in coordinates, the result, in two dimensions, would read:
Ff (x1 b1 , x2 b2 ) = e2i(1 b1 2 b2 ) Ff (1 , 2 ) .
The only advantage in writing it out this way (and you certainly wouldnt do so for any dimension higher
than two) is a more visible reminder that in shifting (x1 , x2 ) to (x1 b1 , x2 b2 ) we shift the variables
independently, so to speak. This independence is also (more) visible in the Fourier transform if we break
up the dot product and multiply the exponentials:
Ff (x1 b1 , x2 b2 ) = e2i1 b1 e2i2 b2 Ff (1 , 2 ) .
The derivation of the shift theorem is pretty much as in the one-dimensional case, but let me show you
how the change of variable works. Well do this for n = 2, and, yes, well write it out in coordinates. Lets
just take the case when were adding b1 and b2 . First off
Z Z
F(f (x1 + b2 , x2 + b2 )) =
e2i(x1 1 +x2 2 ) f (x1 + b1 , x2 + b2 ) dx1 dx2
We want to make a change of variable, turning f (x1 +b1 , x2 +b2 ) into f (u, v) by the substitutions u = x1 +b1
and v = x2 + b2 (or equivalently x1 = u b1 and x2 = v b2 ). You have two choices at this point. The
general change of variables formula for a multiple integral (stay with it for just a moment) immediately
produces.
Z Z
e2i(x1 1 +x2 2 ) f (x1 + b1 , x2 + b2 ) dx1 dx2

Z Z
e2i((ub1 )1 +(vb2 )2 ) f (u, v) du dv
=
Z
Z
=
e2ib1 1 e2ib2 2 e2i(u2 +v2 ) f (u, v) du dv

Z Z
2i(b1 1 +b2 2 )
=e
e2i(u2 +v2 ) f (u, v) du dv
2i(b1 1 +b2 2 )
=e
Ff (1 , 2 ) ,
8.2
349
and theres our formula.

If you know the general change of variables formula then the shift formula and its derivation really are just
like the one-dimensional case, but this doesnt do you much good if you dont know the change of variables
formula for a multiple integral. So, for completeness, let me show you an alternative derivation that works
because the change of variables u = x1 + b1 , v = x2 + b2 changes x1 and x2 separately.
Z Z
Ff (x1 + b2 , x2 + b2 ) =
e2i(x1 1 +x2 2 ) f (x1 + b1 , x2 + b2 ) dx1 dx2

Z

Z
2ix1 1
2ix2 2
e
e
f (x1 + b1 , x2 + b2 ) dx2 dx1
=

Z
Z
2ix1 1
2i(vb2 )2
e
e
f (x1 + b1 , v) dv dx1
=
(substituting v = x2 + b2 )

Z
Z
2ib2 2
2iv2
2ix1 1
=e
e
e
f (x1 + b1 , v) dv dx1

Z
Z
2ib2 2
2iv2
2ix1 1
=e
e
e
f (x1 + b1 , v) dx1 dv

Z
Z
e2i(ub1 )1 f (u, v) du dv
e2iv2
= e2ib2 2
(substituting u = x1 + b1 )

Z
Z
= e2ib2 2 e2ib1 1
e2iv2
e2iu1 f (u, v) du dv
Z
Z
= e2ib2 2 e2ib1 1
e2i(u1 +v2 ) f (u, v) du dv
= e2ib2 2 e2ib1 1 Ff (1 , 2 )
= e2i(b2 2 +b1 1 ) Ff (1 , 2 ) .
And theres our formula, again.
The good news is, weve certainly derived the shift theorem! The bad news is, you may be saying to yourself:
This is not what I had in mind when you said the higher dimensional case is just like the one-dimensional
case. I dont have a quick comeback to that, except that Im trying to make honest statements about the
similarities and the differences in the two cases and, if you want, you can assimilate the formulas and just
skip those derivations in the higher dimensional case that bug your sense of simplicity. I will too, mostly.
8.2.3
Stretches
Theres really only one stretch theorem in higher dimensions, but Id like to give two versions of it. The
first version can be derived in a manner similar to what we did for the shift theorem, making separate
changes of variable. This case comes up often enough that its worth giving it its own moment in the
sun. The second version (which includes the first) needs the general change of variables formula for the
derivation.
Stretch Theorem, first version
1
F(f )
F(f (a1 x1 , a2 x2 )) =
|a1 ||a2 |
1 2
,
a1 a2
350

There is an analogous statement in higher dimensions.
Ill skip the derivation.
The reason that theres a second version of the stretch theorem is because theres something new that
can be done by way of transformations in higher dimensions that doesnt come up in the one-dimensional
setting. We can look at a linear change of variables in the spatial domain. In two dimensions we write
this as

u1
a b
x1
=
u2
c d
x2
or, written out,
u1 = ax1 + bx2
u2 = cx1 + dx2
The simple, independent stretch is the special case

a1 0
x1
u1
.
=
x2
0 a2
u2
For a general linear transformation the coordinates can get mixed up together instead of simply changing
independently.
A linear change of coordinates is not at all an odd a thing to do think of linearly distorting an image,
for whatever reason. Think also of rotation, which well consider below. Finally, a linear transformation as
a linear change of coordinates isnt much good if you cant change the coordinates back. Thus its natural
to work only with invertible transformations here, i.e., those for which det A 6= 0.
The general stretch theorem answers the question of what happens to the spectrum when the spatial
coordinates change linearly what is F(f (u1 , u2 )) = F(f (ax1 + bx2 , cx1 + dx2 ))? The nice answer is
most compactly expressed in matrix notation, in fact just as easily for n dimensions as for two. Let A be
an n n invertible matrix. We introduce the notation
AT = (A1 )T ,
the transpose of the inverse of A. You can check that also AT = (AT )1 , i.e., AT can be defined either
as the transpose of the inverse or as the inverse of the transpose. (AT will also come up naturally when
we apply the Fourier transform to lattices and reciprocal lattices, i.e., to crystals.)
We can now state:
Stretch Theorem, general version
F(f (Ax)) =
1
Ff (AT ) .
| det A|
Theres another way of writing this that you might prefer, depending (as always) on your tastes. Using
det AT = det A and det A1 = 1/ det A we have
1
= | det AT |
| det A|
so the formula reads
F(f (Ax)) = | det AT | Ff (AT ) .
8.2
351
Finally, Im of a mind to introduce the general scaling operator defined by

(A f )(x) = f (Ax) ,
where A is an invertible n n matrix. Then Im of a mind to write
1
F(A f )() =
Ff (AT ) .
| det A|
Your choice. Ill give a derivation of the general stretch theorem in Section ??.
Lets look at the two-dimensional case in a little more detail. To recover the first version of the stretch
theorem we apply the general version to the diagonal matrix

a1 0
A=
with det A = a1 a2 6= 0 .
0 a2
Then
A
This gives
1/a1
0
0
1/a2
1/a1
0
0
1/a2
1
1
Ff (AT ) =
Ff
F(f (a1 x1 , a2 x2 )) = F(f (Ax)) =
| det A|
|a1 ||a2 |
Works like a charm.
1 2
,
a1 a2
An important special case of the stretch theorem is when A is a rotation matrix:

cos sin
A=
sin cos
A rotation matrix is orthogonal, meaning that AAT = I:

2

cos sin
cos sin
cos + sin2
0
1 0
T
AA =
=
=
.
sin cos
sin cos
0
cos2 + sin2
0 1
Thus A1 = AT so that
AT = (A1 )T = (AT )T = A .
Also
det A = cos2 + sin2 = 1 .
The consequence of all of this for the Fourier transform is that if A is a rotation matrix then
F(f (Ax)) = Ff (A), .
In words:
A rotation in the spatial domain corresponds to an identical rotation in the frequency domain.
This result is used all the time in imaging problems.
Finally, its worth knowing that for a 2 2 matrix we can write down AT explicitly:

T

1
1
1
d b
d c
a b
a b
=
so the transpose of this is
=
c d
c d
det A c a
det A b a
This jibes with what we found for a rotation matrix.
352
The indicator function for a parallelogram As an exercise in using the stretch theorem you can
show the following. Consider a parallelogram centered at (0, 0):
One set of data that describes the parallelogram are the distances between sides, p and q, and the vectors
that give the directions of the sides. Let u be a unit vector in the direction of the sides that are p apart
and let v be a unit vector in the direction of the sides that are q apart.
The indicator function P for the parallelogram is the function that is equal to 1 on the parallelogram and
equal to 0 outside the parallelogram. The Fourier transform of P can be shown to be

p(u )
q(v )
pq
sinc
sinc
.
FP () =
| sin |
sin
sin
Shift and stretch
show:
As an example of using the general formula, lets combine a shift with a stretch and
F(f (Ax + b)) = exp(2ib AT )
1
Ff (AT )
| det A|
(I think the exponential is a little crowded to write it as e to a power here.) Combining shifts and stretches
seems to cause a lot of problems for people (even in one dimension), so let me do this in several ways.
As a first approach, and to keep the operations straight, write
g(x) = f (x + b) ,
8.2
353
and then
f (Ax + b) = g(Ax) .
Using the stretch theorem first,
F(g(Ax)) =
1
Fg(AT )
| det A|
Applying the shift theorem next gives

(Fg)(AT ) = exp(2ib AT )Ff ((AT ) .
Putting these together gives the final formula for F(f (Ax + b)).
Another way around is instead to write
g(x) = f (Ax)
and then
f (Ax + b) = f (A(x + A1 b)) = g(x + A1 b) .
Now use the shift theorem first to get
F(g(x + A1 b)) = exp(2iA1 b ) (Fg)() = exp(2ib AT ) (Fg)() .
The stretch theorem comes next and it produces
Fg() = F(f (Ax)) =
1
Ff (AT ) .
| det A|
This agrees with what we had before, as if there was any doubt.
Finally, by popular demand, I do this one more time by expressing f (Ax + b) using the delay and scaling
operators. Its a question of which comes first, and parallel to the first derivation above we can write:
f (Ax + b) = A (b f )(x) = (A b f )(x) ,
which we verify by
(A b f )(x) = (b f )(Ax) = f (Ax + b) .
And now we have
F(A (b f ))() =
1
1
F(b f )(AT ) =
exp(2iAT b)Ff (AT ) .
| det A|
| det A|
I wont give a second version of the second derivation.
8.2.4
Convolution
What about convolution? For two real-valued functions f and g on Rn the definition is
Z
f (x y)g(y) dy .
(f g)(x) =
Rn
Written out in coordinates this looks much more complicated. For n = 2, for example,
Z Z
f (x1 y1 , x2 y2 )g(y1 , y2 ) dy1 dy2 .
(f g)(x1 , x2 ) =
The intelligent person would not write out the corresponding coordinatized formula for higher dimensions
unless absolutely pressed. The intelligent person would also not try too hard to flip, drag or otherwise
visualize a convolution in higher dimensions. The intelligent person would be happy to learn, however,
that once again
F(f g)() = Ff ()Fg() and F(f g)() = (Ff Fg)() .
The typical interpretations of convolution smoothing, averaging, etc. continue to apply, when applied
by an intelligent person.
354
8.2.5
A little now, more later
Well see that things get more interesting in higher dimensions for delta functions, but the definition of the
plain vanilla is the same as before. To give the distributional definition, Ill pause, just for a moment, to
define what it means for a function of several variables to be a Schwartz function.
Schwartz functions The theory and practice of tempered distributions works the same in higher dimensions as it does in one. The basis of the treatment is via the Schwartz functions as the class of test
functions. The condition that a function of several variables be rapidly decreasing is that all partial derivatives (including mixed partial derivatives) decrease faster than any power of any of the coordinates. This
can be stated in any number of equivalent forms. One way is to require that
|x|p | q (x)| 0 as
|x| .
Ill explain the funny notation its an example of the occasional awkwardness that sets in when writing
formulas in higher dimensions. p is a positive integer, so that just gives a power of |x|, and q is a multi-index.
This means that q = (q1 , . . . , qn ), each qi a positive integer, so that q is supposed to mean
q1 ++qn
.
(xn )qn
(x1 )q1 (x2 )q2
Theres no special font used to indicate multi-indices you just have to intuit it.
From here, the definitions of tempered distributions, the Fourier transform of a tempered distribution, and
everything else, goes through just as before. Shall we leave it alone? I thought so.
in higher dimensions
The -function is the distribution defined by the pairing
h, i = (0, . . . , 0)
or
h, i = (0)
in vector notation
where (x1 , , . . . , xn ) is a Schwartz function.6 As is customary, we also write this in terms of integration
as:
Z
(x)(x) dx = (0)
Rn
You can show that is even as a distribution (once youve reminded yourself what it means for a distribution
to be even).
As before, one has
f (x)(x) = f (0)(x) ,
when f is a smooth function, and for convolution
(f )(x) = f (x) .
The shifted delta function (x b) = (x1 b1 , x2 b2 , , . . . , xn bn ) or b = b , has the corresponding
properties
f (x)(x b) = f (b)(x b) and f (x b) = f (x b) .
In some cases it is useful to know that we can factor the delta function into one-dimensional deltas, as
in
(x1 , x2 , . . . , xn ) = 1 (x1 )2 (x2 ) n (xn ) .
6
Actually, is in a larger class than the tempered distributions. It is defined by the pairing h, i = (0) when is any
smooth function of compact support.
8.2
355
Ive put subscripts on the s on the right hand side just to tag them with the individual coordinates
there are some advantages in doing this. Though it remains true, as a general rule, that multiplying
distributions is not (and cannot be) defined, this is one case where it makes sense. The formula holds
because of how each side acts on a Schwartz function.7 Lets just check this in the two-dimensional case,
and play a little fast and loose by writing the pairing as an integral. Then, on the one hand,
Z
(x)(x) dx = (0, 0)
R2
by definition of the 2-dimensional delta function. On the other hand,

Z Z
Z
(x1 , x2 )1 (x1 ) dx1 2 (x2 ) dx2
(x1 , x2 )1 (x1 )2 (x2 ) dx1 dx2 =
R2
Z
(0, x2 )2 (x2 ) dx2 = (0, 0).
=
So (x1 , x2 ) and 1 (x1 )2 (x2 ) have the same effect when integrated against a test function.
The Fourier transform of And finally the Fourier transform of the delta function is, of course,
1 (thats the constant function 1). The argument is the same as in the one-dimensional case. By duality,
the Fourier transform of 1 is . One can then shift to get
(x b) e2ib
or
F b = e2ib .
You can now see (again) where those symmetrically paired dots come from in looking at the spectral
picture for alternating black and white stripes. It comes from the Fourier transforms of cos(2 x 0 ) =
Re exp(2i x 0 ) for 0 = (1 , 0), 0 = (0, 2 ), and 0 = (3 , 3 ), since
F cos(2 x 0 ) = 21 (( 0 ) + ( + 0 )) .
7
The precise way to do this is through the use of tensor products of distributions, something we have not discussed, and will
not.
356
Scaling delta functions
Recall how a one-dimensional delta function scales:

(ax) =
1
(x) .
|a|
Writing a higher dimensional delta function as a product of one-dimensional delta functions we get a
corresponding formula. In two dimensions:
(a1 x1 , a2 x2 ) = 1 (a1 x1 )2 (a2 x2 )
1
1
1 (x1 )
2 (x2 )
=
|a1 |
|a2 |
1
1
1 (x1 )2 (x2 ) =
(x1 , x2 ),
=
|a1 | |a2 |
|a1 a2 |
and in n-dimensions
1
(x1 , . . . , xn ) .
|a1 an |
(a1 x1 , . . . , an xn ) =
Its also possible (and useful) to consider (Ax) when A is an invertible matrix. The result is
1
(x) .
| det A|
(Ax) =
See Section ?? for a derivation of this. This formula bears the same relationship to the preceding formula
as the general stretch theorem bears to the first version of the stretch theorem.
8.2.6
The Fourier transform of a radial function
For use in many applications, were going to consider one further aspects of the 2-dimensional case. A
function on R2 is radial (also called radially symmetric or circularly symmetric) if it depends only on the
distance from the origin. In polar coordinates the distance from the origin is denoted by r, so to say that
a function is radial is to say that it depends only on r (and that it does not depend on , writing the usual
polar coordinates as (r, )).
The definition of the Fourier transform is set up in Cartesian coordinates, and its clear that well be better
off writing it in polar coordinates if we work with radial functions. This is actually not so straightforward,
or, at least, it involves introducing some special functions to write the formulas in a compact way.
We have to convert
2ix
e
2
f (x) dx =
e2i(x1 1 +x2 2 ) f (x1 , x2 ) dx1 dx2
to polar coordinates. There are several steps: To say that f (x) is a radial function means that it becomes f (r). To describe all of R2 in the limits of integration, we take r going from 0 to and
going from 0 to 2. The area element dx1 dx2 becomes r dr d. Finally, the problem is the inner product
x = x1 1 +x2 2 in the exponential and how to write it in polar coordinates. If we identify (x1 , x2 ) = (r, )
(varying over the (x1 , x2 )-plane) and put (1 , 2 ) = (, ) (fixed in the integral) then
x = kxk kk cos( ) = r cos( ) .
The Fourier transform of f is thus
Z
Z Z
2ix
e
f (x) dx =
f (r)e2ir cos() r dr d .
8.2
357
Theres more to be done. First of all, because e2ir cos() is periodic (in ) of period 2, the integral
Z 2
e2ir cos() d
0
does not depend on .8 Consequently,

Z
Z 2
2ir cos()
e
d =
0
e2ir cos d .
The next step is to define ourselves out of trouble. We introduce the function
Z 2
1
J0 (a) =
eia cos d .
2 0
We give this integral a name, J0 (a), because, try as you might, there is no simple closed form expression
for it, so we take the integral as defining a new function. It is called the zero order Bessel function of the
first kind. Sorry, but Bessel functions, of whatever order and kind, always seem to come up in problems
involving circular symmetry; ask any physicist.
Incorporating J0 into what weve done,
Z 2
e2ir cos d = 2J0 (2r)
and the Fourier transform of f (r) is

2
f (r)J0 (2r) r dr
0
Lets summarize:
If f (x) is a radial function then its Fourier transform is
Z
f (r)J0 (2r) rdr
F () = 2
0
In words, the important conclusion to take away from this is that the Fourier transform of a radial
function is also radial.
The formula for F () in terms of f (r) is sometimes called the zero order Hankel transform of f (r) but,
again, we understand that it is nothing other than the Fourier transform of a radial function.
Circ and Jinc
function, is
A useful radial function to define, sort of a radially symmetric analog of the rectangle
(
1
circ(r) =
0
r<1
r1
(And one can argue about the value at the rim r = 1.) Heres the graph.
8
Weve applied this general fact implicitly or explicitly on earlier occasions when working with periodic functions, namely if
g is periodic with period 2 then
Z 2
Z 2
g( ) d =
g() d
0
Convince yourself of this; for instance let G() =

G() = G(0).
R 2
0
g( ) d and show that G () 0. Hence G() is constant, so
358
For its Fourier transform the limits of integration on r go only from 0 to 1, and so we have simply
Z 1
J0 (2r) r dr .
Fcirc() = 2
0
We make a change of variable, u = 2r. Then du = 2dr and the limits of integration go from u = 0
to u = 2. The integral becomes
Z 2
1
Fcirc() =
uJ0 (u) du .
22 0
We write the integral this way because, you will now be ecstatic to learn, there is an identity that brings
in the first-order Bessel function of the first kind. That identity goes
Z x
uJ0 (u) du = xJ1 (x) .
0
In terms of J1 we can now write

Fcirc() =
J1 (2)
It is customary to introduce the jinc function, defined by

jinc() =
J1 ()
.
2
In terms of this,
Fcirc() = 4 jinc(2) .
The graph of Fcirc is:
8.2
359
I could plot this because Bessel functions are so common (really) that they are built into many mathematical
software packages, such as Matlab or Mathematica. If you think the jinc function looks like some kind of
radially symmetric version of the sinc function youd be right. But its not obvious just how one goes from
sinc to jinc, and well have to pass on this.9
8.2.7
A Derivation of the General Stretch Theorem
The general stretch theorem says that if A is an invertible n n matrix then

1
Ff (AT ) .
| det A|
F(f (Ax)) =
To derive this lets start with the left hand side:
Z
F(f (Ax)) =
e2ix f (Ax) dx .
n
Our object is to make a change of variable, u = Ax. For this, we need to use the change of variables
formula for multiple integrals. In the form we need it, we can state:
If A is an invertible n n matrix and u = Ax then
Z
Z
g(Ax) | det A| dx =
Rn
Rn
g(u) du .
for an integrable function g.

9
Theres a symmetrization process at work involving repeated convolutions. I have notes on this. . .
360
Want to feel good (or at least OK) about this in a familiar setting? Take the case n = 1. Then
Z
Z
g(u) du ,
g(ax) |a|dx =
making the substitution u = ax. The transformation u = ax of R scales lengths, and the scaling factor is
a. (du = a dx). Thats if a is positive; the absolute value of a is in there in case a is negative thus sense
reversing. In n-dimensions the transformation u = Ax scales n-dimensional volumes, and the scaling
factor is det A. (du = det A dx.) The absolute value | det A| is in there because a matrix A with det A > 0
is sense preserving on Rn , and it is sense reversing if det A < 0. Thus, in general,
du = | det A| dx
so the substitution u = Ax leads right to the formula
Z
Z
g(Ax) | det A| dx =
Rn
Rn
g(u) du .
To apply this to the Fourier transform of f (Ax) we have

Z
Z
1
1
| det A| dx
e2iA (Ax) f (Ax)
e2ix f (Ax) dx =
| det A|
Rn
Rn
Z
1
1
=
e2iA (Ax) f (Ax) | det A| dx
| det A| Rn
Z
1
1
e2iA u f (u) du
=
| det A| Rn
(now substitute u = Ax)
If you think this looks complicated imagine writing it out in coordinates!

Next we use an identity for what happens to the dot product when theres a matrix operating on one of
the vectors, namely, for a matrix B and any vectors and u,
Bu = B T u .
We take B = A1 and then
A1 u = AT u .
With this:
1
| det A|
e2iA
1 u
f (u) du =
1
| det A|
e2iA
T u
f (u) du.
But this last integral is exactly F(f )(AT ). We have shown that
F(f (Ax)) =
1
F(f )(AT ) ,
| det A|
as desired.
Scaling the delta function
The change of variables formula also allows us to derive

(Ax) =
1
(x) .
| det A|
8.3
Higher Dimensional Fourier Series
361
Writing the pairing of (Ax) with a test function via integration not strictly legit, but it helps to
organize the calculation leads to
Z
Z
1
| det A| dx
(Ax)(A1 Ax)
(Ax)(x) dx =
| det A|
Rn
Rn
Z
1
=
(u)(A1 u) du (making the change of variables u = Ax)
| det A| Rn
1
(A1 0) (by how the delta function acts)
=
| det A|
1
=
(0) (A1 0 = 0 because A1 is linear)
| det A|
Thus (Ax) has the same effect as
8.3
1
when paired with a test function, so they must be equal.
| det A|
Its important to know that most of the ideas and constructions for Fourier series carry over directly to
periodic functions in two, three, or higher dimensions. Here we want to give just the basic setup so you
can see that the situation, and even the notation, is very similar to what weve already encountered. After
that well look at a fascinating problem where higher dimensional Fourier series are central to the solution,
but in a far from obvious way.
Periodic Functions The definition of periodicity for real-valued functions of several variables is much
the same as for functions of one variable except that we allow for different periods in different slots. To
take the two-dimensional case, we say that a function f (x1 , x2 ) is (p1 , p2 )-periodic if
f (x1 + p1 , x2 ) = f (x1 , x2 ) and f (x1 , x2 + p2 ) = f (x1 , x2 )
for all x1 and x2 . It follows that
f (x1 + p1 , x2 + p2 ) = f (x1 , x2 )
and more generally that
f (x1 + n1 p1 , x2 + n2 p2 ) = f (x1 , x2 )
for all integers n1 , n2 .
Theres a small but important point associated with the definition of periodicity having to do with properties of f (x1 , x2 ) one variable at a time or both variables together. The condition
f (x1 + n1 p1 , x2 + n2 p2 ) = f (x1 , x2 )
for all integers n1 , n2 can be taken as the definition of periodicity, but the condition f (x1 + p1 , x2 + p2 ) =
f (x1 , x2 ) alone is not the appropriate definition. The former implies that f (x1 + p1 , x2 ) = f (x1 , x2 ) and
f (x1 , x2 + p2 ) = f (x1 , x2 ) by taking (n1 , n2 ) to be (1, 0) and (0, 1), respectively, and this independent
periodicity is what we want. The latter condition does not imply independent periodicity.
For our work now its enough to assume that the period in each variable is 1, so the condition is
f (x1 + 1, x2 ) = f (x1 , x2 )
and f (x1 , x2 + 1) = f (x1 , x2 ) ,
362
or
f (x1 + n1 , x2 + n2 ) = f (x1 , x2 ) for all integers n1 , n2 .
If we use vector notation and write x for (x1 , x2 ) and (why not) n for the pair (n1 , n2 ) of integers, then
we can write the condition as
f (x + n) = f (x) ,
and, except for the typeface, it looks like the one-dimensional case.
Where is f (x1 , x2 ) defined? For a periodic function (of period 1) it is enough to know the function for
x1 [0, 1] and x2 [0, 1]. We write this as
(x1 , x2 ) [0, 1]2 .
We can thus consider f (x1 , x2 ) to be defined on [0, 1]2 and then extended to be defined on all of R2 via
the periodicity condition.
We can consider periodicity of functions in any dimension. To avoid conflicts with other notation, in this
discussion Ill write the dimension as d rather than n. Let x = (x1 , x2 , . . . , xd ) be a vector in Rd and let
n = (n1 , n2 , . . . , nd ) be an d-tuple of integers. Then f (x) = f (x1 , x2 , . . . , xd ) is periodic (of period 1 in
each variable) if
f (x + n) = f (x) for all n .
In this case we consider the natural domain of f (x) to be [0, 1]d , meaning the set of points (x1 , x2 , . . . , xd )
where 0 xj 1 for each j = 1, 2, . . . , d.
Complex exponentials, again What are the building blocks for periodic functions in higher dimensions? We simply multiply simple complex exponentials of one variable. Taking again the two-dimensional
case as a model, the function
e2ix1 e2ix2
is periodic with period 1 in each variable. Note that once we get beyond one dimension its not so helpful
to think of periodicity in time and to force yourself to write the variable as t.
In d dimensions the corresponding exponential is
e2ix1 e2ix2 e2ixd
You may be tempted to use the usual rules and write this as
e2ix1 e2ix2 e2ixd = e2i(x1 +x2 ++xd ) .
Dont do that yet.
Higher harmonics, Fourier series, et al. Can a periodic function f (x1 , x2 , . . . , xd ) be expressed as
a Fourier series using multidimensional complex exponentials? The answer is yes and the formulas and
theorems are virtually identical to the one-dimensional case. First of all, the natural setting is L2 ([0, 1]d ).
This is the space of square integrable functions:
Z
|f (x)|2 dx <
[0,1]d
8.3
363
This is meant as a multiple integral, e.g., in the case d = 2 the condition is

Z 1Z 1
|f (x1 , x2 )|2 dx1 dx2 < .
0
The inner product of two (complex-valued) functions is

Z 1Z 1
(f, g) =
f (x1 , x2 )g(x1 , x2 ) dx1 dx2 .
0
Im not going to relive the greatest hits of Fourier series in the higher dimensional setting. The only thing
I want us to know now is what the expansions look like. Its nice watch. Lets do the two-dimensional
case as an illustration. The general higher harmonic is of the form
e2in1 x1 e2in2 x2 ,
where n1 and n2 are integers. We would then imagine writing the Fourier series expansion as
X
cn1 n2 e2in1 x1 e2in2 x2 ,
n1 ,n2
where the sum is over all integers n1 , n2 . More on the coefficients in a minute, but first lets find a more
attractive way of writing such sums.
Instead of working with the product of separate exponentials, its now time to combine them and see what
happens:
e2in1 x1 e2in2 x2 = e2i(n1 x1 +n2 x2 )
= e2i nx
(dot product in the exponent!)
where we use vector notation and write n = (n1 , n2 ). The Fourier series expansion then looks like
X
c n e2inx .
n
The dot product in two dimensions has replaced ordinary multiplication in the exponent in one dimension, but the formula looks the same. The sum has to be understood to be over all points (n1 , n2 ) with
integer coefficients. We mention that this set of points in R2 is called the two-dimensional integer lattice,
written Z2 . Using this notation we would write the sum as
X
c n e2i nx .
nZ2
What are the coefficients? The argument we gave in one dimension extends easily to two dimensions (and
more) and one finds that the coefficients are given by
Z 1Z 1
Z 1Z 1
2in1 x1 2in2 x2
e2i(n1 x1 +n2 x2 ) f (x1 , x2 ) dx1 dx2
e
e
f (x1 , x2 ) dx1 dx2 =
0
0
0
0
Z
e2i nx f (x) dx
=
[0,1]2
Thus the Fourier coefficients f(n) are defined by the integral

Z
e2i nx f (x) dx
f(n) =
[0,1]2
364
It should now come as no shock that the Fourier series for a periodic function f (x) in Rd is
X
f(n)e2i nx ,
n
where the sum is over all points n = (n1 , n2 , . . . , nd ) with integer entries. (This set of points is the integer
lattice in Rd , written Zd .) The Fourier coefficients are defined to be
Z
e2i nx f (x) dx .
f(n) =
[0,1]d
Coming up next is an extremely cool example of higher dimensional Fourier series in action. Later well
come back to higher dimensional Fourier series and their application to crystallography.
8.3.1
The eternal recurrence of the same?
For this example we need to make some use of notions from probability, but nothing beyond what we used
in discussing the Central Limit Theorem in Chapter ??. For this excursion, and your safe return, you will
need:
To remember what probability means.
To know that for independent events the probabilities multiply, i.e., Prob(A, B) = Prob(A) Prob(B),
meaning that the probability of A and B occuring (together) is the product of the separate probabilities of A and B occuring.
To use expected value, which we earlier called the mean.
Though the questions well ask may be perfectly natural, you may find the answers surprising.
Ever hear of a random walk? Its closely related to Brownian motion and can also be described as a
Markov process. We wont take either of these latter points of view, but if or rather, when you
encounter these ideas in other courses, you have been warned.
Heres the setup for a random walk along a line:
Youre at home at the origin at time n = 0 and you take a step, left or right chosen with
equal probability; flip a coin; heads you move right, tails you move left. Thus at time n = 1
youre at one of the points +1 or 1. Again you take a step, left or right, chosen with equal
probability. Youre either back home at the origin or at 2. And so on.
As you take more and more steps, will you get home (to the origin)?
With what probability?
We can formulate the same question in two, three, or any number of dimensions. We can also tinker with
the probabilities and assume that steps in some directions are more probable than in others, but well stick
with the equally probable case.
9
With apologies to F. Nietzsche
8.3
365
Random walks, Markov processes, et al. are used everyday by people who study queuing problems, for
example. More recently they have been applied in mathematical finance. A really interesting treatment is
the book Random Walks and Electrical Networks by P. Doyle and J. L. Snell.
To answer the questions its necessary to give some precise definitions, and that will be helped by fixing
some notation. Think of the space case d = 3 as an example. Well write the location of a point with
reference to Cartesian coordinates. Start at the origin and start stepping. Each step is by a unit amount
in one of six possible directions, and the directions are chosen with equal probability, e.g., throw a single
die and have each number correspond to one of six directions. Wherever you go, you get there by adding
to where you are one of the six unit steps
(1, 0, 0),
(0, 1, 0),
(0, 0, 1) .
Denote any of these elementary steps, or more precisely the random process of choosing any of these
steps, by step; to take a step is to choose one of the triples, above, and each choice is made with probability
1/6. Since were interested in walks more than we are individual steps, lets add an index to step and
write step1 for the choice in taking the first step, step2 for the choice in taking the second step, and so on.
Were also assuming that each step is a new adventure the choice at the n-th step is made independently
of the previous n 1 steps. In d dimensions there are 2d directions each chosen with probability 1/2d, and
stepn is defined in the same manner.
The process stepn is a discrete random variable. To be precise:
The domain of stepn is the set of all possible walks and the value of stepn on a particular walk is
the nth step in that walk.
(Some people would call stepn a random vector since its values are d-tuples.) Were assuming that
distribution of values of stepn is uniform (each particular step is taken with probability 1/2d, in general)
and that the steps are independent. Thus, in the parlance weve used in connection with the Central Limit
Theorem, step1 , step2 , . . . , stepn are independent, identically distributed random variables.
The possible random walks of n steps are described exactly as
walkn = step1 + step2 + + stepn ,
or, for short, just
w n = s1 + s2 + + sn .
Im using the vector notation for w and s to indicate that the action is in Rd .
366
Heres a picture in R3 .
After a walk of n steps, n 1, you are at a lattice point in Rd , i.e., a point with integer coordinates. We
now ask two questions:
1. Given a particular lattice point l, what is the probability after n steps that we are at l?
2. How does walkn behave as n ?
These famous questions were formulated and answered by G. P
olya in 1921. His brilliant analysis resulted
in the following result.
Theorem In dimensions 1 and 2, with probability 1, the walker visits the origin infinitely
often; in symbols
Prob(walkn = 0 infinitely often) = 1 .
In dimensions 3, with probability 1, the walker escapes to infinity:

Prob lim |walkn | = = 1 .
n
8.3
367
One says that a random walk along a line or in the plane is recurrent and that a random walk in higher
dimensions is transient.
Heres the idea very cunning and, frankly, rather unmotivated, but who can account for genius? For
each x Rd consider
n = e2i w n x ,
where, as above, w n is a walk of n steps. For a given n the possible values of w n , as a sum of steps
corresponding to different walks, lie among the lattice points, and if w n lands on a lattice point l then
the value of n for that walk is e2i lx . What is the expected value of n over all walks of n steps? It is
the mean, i.e., the weighted average of the values of n over the possible (random) walks of n steps, each
value weighted by the probability of its occurrence. That is,
X
Expected value of n =
Prob(w n = l)e2i lx .
l
This is actually a finite sum because in n steps we can have reached only a finite number of lattice points,
or, put another way, Prob(w n = l) is zero for all but finitely many lattice points l.
From this expression you can see (finite) Fourier series coming into the picture, but put that off for the
moment.10 We can compute this expected value, based on our assumption that steps are equally probable
and independent of each other. First of all, we can write
n = e2i w n x = e2i(s 1 +s 2 ++s n )x = e2i s 1 x e2i s 2 x e2i s n x .
So we want to find the expected value of the product of exponentials. At this point we could appeal to
a standard result in probability, stating that the expected value of the product of independent random
variables is the product of their expected values. You might be able to think about this directly, however:
The expected value of e2i s 1 x e2is 2 x e2is n x is, as above, the weighted average of the values that the
function assumes, weighted by the probabilities of those values occuring. In this case wed be summing over
all steps s1 , s2 , . . . , sn of the values e2is1 x e2is2 x e2isn x weighted by the appropriate probabilities.
But now the fact that the steps are independent means
Prob(s 1 = s1 , s 2 = s2 , . . . , s n = sn ) = Prob(s 1 = s1 ) Prob(s 2 = s2 ) Prob(s n = sn )
(probabilities multiply for independent events)
=
1
,
(2d)n
and then
Expected value of n = Expected value of e2is1 x e2is2 x e2isn x
XX X
=
Prob(s 1 = s1 , s 2 = s2 , . . . , s n = sn )e2i s1 x e2is2 x e2isn x
s1
s1
10
sn
s2
XX
s2
X
sn
1
e2i s1 x e2i s2 x e2i sn x .
(2d)n
Also, though its not in the standard form, i.e., a power series, I think of P
olyas idea here as writing down a generating
function for the sequence of probabilities Prob(w n = l). For an appreciation of this kind of approach to a great variety of
problems pure and applied see the book Generatingfunctionology by H. Wilf. The first sentence of Chapter One reads:
A generating function is a clothesline on which we hang up a sequence of numbers for display. Seems pretty apt for the
problem at hand.
368
The sums go over all possible choices of s1 , s2 ,. . . ,sn . Now, these sums are uncoupled, and so the nested
sum is the product of
X 1
X 1
X 1
e2i s1 x
e2i s2 x
e2i sn x .
s1
2d
s2
2d
sn
2d
But the sums are, respectively, the expected values of e2isj x , j = 1, . . . , n, and these expected values are
all the same. (The steps are independent and identically distributed). So all the sums are equal, say, to
the first sum, and we may write
n
1 X
e2i s1 x
Expected value of n =
2d
s1
A further simplification is possible. The first step s1 , as a d-tuple has exactly one slot with a 1 and the
rest 0s. Summing over these 2d possibilities allows us to combine positive and negative terms. Check
the case d = 2, for which the choices of s1 are
(1, 0) ,
(1, 0) ,
(0, 1) ,
(0, 1) .
This leads to a sum with four terms:

X 1
X 1
e2i s1 x =
e2i s1 (x1 ,x2 )
s1
22
22
s1
1 1 2ix1
2(2e
+ 12 e2ix1 + 21 e2ix2 + 12 e2ix2 )
= 12 (cos 2x1 + cos 2x2 )

The same thing happens in dimension d, and our final formula is
n
1
X
(cos 2x1 + cos 2x2 + + cos 2xd ) .
Prob(w n = l)e2i lx =
d
Let us write
1
d
d (x) = (cos 2x1 + cos 2x2 + + cos 2xd ) .

Observe that |d (x)| 1, since d (x) is the sum of d cosines by d and | cos 2xj | 1 for j = 1, 2, . . . , d.
This has been quite impressive already. But theres more! Lets get back to Fourier series and consider the
sum of probabilities times exponentials, above, as a function of x; i.e., let
X
Prob(w n = l) e2i lx .
f (x) =
l
This is a (finite) Fourier series for f (x) and as such the coefficients must be the Fourier coefficients,
Prob(w n = l) = f(l) .
But according to our calculation, f (x) = d (x)n , and so this must also be the Fourier coefficient of d (x)n ,
that is,
Z
e2i lx d (x)n dx .
Prob(w n = l) = f(l) = \
(d )n (l) =
[0,1]d
In particular, the probability that the walker visits the origin, l = 0, in n steps is
Z
d (x)n dx .
Prob(w n = 0) =
[0,1]d
8.3
369
Then the expected number of times the walker visits the origin for a random walk of infinite length is
Prob(w n = 0) ,
n=0
and we can figure this out.11 Heres how we do this. Wed like to say that
Z
X
X
Prob(w n = 0) =
d (x)n dx
n=0
d
n=0 [0,1]
[0,1]d
d (x)
n=0
dx =
[0,1]d
1
dx
1 d (x)
using the formula for adding a geometric series. The final answer is correct, but the derivation isnt quite
legitimate: The formula for the sum of a geometric series is
rn =
n=0
1
1r
provided that |r| is strictly less than 1. In our application we know only that |d (x)| 1. To get around
this difficulty, let < 1, and write

X
Z
X
X
n
n
n
Prob(w n = 0) = lim
Prob(w n = 0) = lim
d (x) dx
n=0
= lim
n=0
Z
1 [0,1]d
1 [0,1]d
1
dx =
1 d (x)
[0,1]d
n=0
1
dx
1 d (x)
(Pulling the limit inside the integral is justified by convergence theorems in the theory of Lebesgue integration, specifically, dominated convergence. Not to worry.)
The crucial question now concerns the integral
Z
[0,1]d
1
dx .
1 d (x)
Is it finite or infinite?
This depends on the dimension and this is exactly where the dimension d enters the picture.
Using some calculus (think Taylor series) it is not difficult to show (I wont) that if |x| is small then
1 d (x) c|x|2 ,
for a constant c. Thus
1
1
,
1 d (x)
c|x|2
and the convergence of the integral were interested in depends on that of the power integral
Z
1
dx in dimension d .
2
|x|
x small
It is an important mathematical fact of nature (something you should file away for future use) that
11
For those more steeped in probability, heres a further argument why this sum is the expected number of visits to the
origin. Let Vn be the random variable which is 1 if the walker returns to the origin in n steps and is zero otherwise. The
expected valuePof Vn is then Prob(w n = 0) 1, the value of the function, 1, times the probability of that value occurring.
Now
P set V = n=0 Vn . The expected value of V is what we want and it is the sum of the expected values of the Vn , i.e.
n=0 Prob(w n = 0).
370

The power integral diverges for d = 1, 2.
The power integral converges for d 3
Let me illustrate why this is so for d = 1, 2, 3. For d = 1 we have an ordinary improper integral,
Z a
dx
, for some small a > 0 ,
2
0 x
and this diverges by direct integration. For d = 2 we have a double integral, and to check its properties
we introduce polar coordinates (r, ) and write
Z 2 Z a
Z
Z 2Z a
dx1 dx2
dr
r dr d
=
d .
2 + x2 =
2
r
x
|x| small 1
0
0 r
0
0
2
The inner integral diverges. In three dimensions we introduce spherical coordinates (, , ), and something
different happens. The integral becomes
Z Z 2Z a 2
Z
sin d d d
dx1 dx2 dx3
=
.
2
2
2
2
0 0
0
|x| small x1 + x2 + x3
This time the 2 in the denominator cancels with the 2 in the numerator and the -integral is finite. The
same phenomenon persists in higher dimensions, for the same reason (introducing higher dimensional polar
coordinates).
Lets take stock. We have shown that

Expected number of visits to the origin =
Prob(w n = 0) =
n=0
[0,1]d
1
dx
1 d (x)
and that this number is infinite in dimensions 1 and 2 and finite in dimension 3. From here we can go on
to prove P
olyas theorem as he stated it:
Prob(walkn = 0 infinitely often) = 1 in dimensions 1 and 2.
Prob(limn |walkn | = ) = 1 in dimensions 3.
For the case d 3, we know that the expected number of times that the walker visits the origin is finite.
This can only be true if the actual number of visits to the origin is finite with probability 1. Now the
origin is not special in any way, so the same must be true of any lattice point. But this means that for
any R > 0 the walker eventually stops visiting the ball |x| R of radius R with probability 1, and this is
exactly saying that Prob(limn |walkn | = ) = 1.
To settle the case d 2 we formulate a lemma that you might find helpful in this discussion.12
Lemma Let pn be the probability that a walker visits the origin at least n times and let qn be
the probability that a walker visits the origin exactly n times. Then pn = pn1 and qn = pn1 (1p1 )
12
We havent had many lemmas in this class, but I think I can get away with one or two.
8.4
III, Lattices, Crystals, and Sampling
371
To show this we argue as follows. Note first that p0 = 1 since the walker starts at the origin. Then
pn+1 = Prob(visit origin at least n + 1 times)
= Prob(visit origin at least n + 1 times given visit at least n times) Prob(visit at least n times)
= Prob(visit origin at least 1 time given visit at least 0 times) pn
(using independence and the definition of pn )
= Prob(visit at least 1 time) pn
= p1 pn
From p0 = 1 and pn+1 = p1 pn it follows (by induction) that pn = pn1 .
For the second part,
qn = Prob(exactly n visits to origin)
= Prob(visits at least n times) Prob(visits at least n + 1 times)
= pn pn+1 = pn1 (1 p1 )
Now, if p1 were less than 1 then the expected number of visits to the origin would be
n=0
nqn =
npn1 (1 p1 ) = (1 p1 )
npn1
n=0
n=0
= (1 p1 )
p1
(1 p1 )2
(Check that identity by differentiating identity
X
1
xn )
=
1x
n=0
p1
<
=
1 p1
But this contradicts the fact we established earlier, namely
Z
Expected visits to the origin =
[0,1]2
1
dx = .
1 2 (x)
Thus we must have p1 = 1, that is, the probability of returning to the origin is 1, and hence walkn must
equal 0 infinitely often with probability 1.
8.4
Our derivation of the sampling formula in Chapter ??? was a direct application and combination of the
important properties of the III function,
IIIp (t) =
(t kp) .
k=
Without redoing the whole argument here, short as it is, let me remind you what it is about III that made
things work.
372

s being what they are, IIIp is the tool to use for periodizing and for sampling:
(f IIIp )(t) =
f (t kp)
k=
f (t)IIIp (t) =
f (kp)(t kp) .
k=
For the Fourier transform,

FIIIp =
1
III .
p 1/p
It is through this property of the Fourier transform that periodizing in one domain corresponds to
sampling in the other domain. Pay particular attention here to the reciprocity in spacing between
IIIp and its Fourier transform.
The sampling formula itself says that if Ff (s) is identically 0 for |s| p/2 then

X
k
sinc(pt k) .
f
f (t) =
k=
We now want to see how things stand in two dimensions; there isnt much difference in substance between
the two-dimensional case and higher dimensions, so well stick pretty much to the plane.
8.4.1
The two-dimensional III
The formula FIIIp = (1/p)III1/p depends crucially on the fact that IIIp is a sum of impulses at evenly spaced
points this is an aspect of periodicity. Weve already defined a two-dimensional , so to introduce a
III that goes with it we need to define what evenly spaced means for points in R2 . One way of spacing
points evenly in R2 is to take all pairs (k1 , k2 ), k1 , k2 integers. The corresponding III-function is then
defined to be
X
III(x1 , x2 ) =
(x1 k1 , x2 k2 ) .
k1 ,k2 =
Bracewell, and others, sometimes refer to this as the bed of nails.
The points k = (k1 , k2 ) with integer coordinates are said to form a lattice in the plane. We denote this
particular lattice, called the integer lattice, by Z2 ; well have more general lattices in a short while. As a
model of a physical system, you can think of such an array as a two-dimensional crystal, where theres an
atom at every lattice point.
Since we prefer to write things in terms of vectors, another way to describe Z2 is to use the standard basis
of R2 , the vectors e 1 = (1, 0), e 2 = (0, 1), and write the points in the lattice as
k = k1 e 1 + k2 e 2 .
We can thus think of the elements of a lattice either as points or as vectors, and observe that the sum of
two lattice points is another lattice point and that an integer multiple of a lattice point is another lattice
point. The III-function can be written
IIIZ2 (x) =
k1 ,k2 =
It is easy to show that IIIZ2 is even.
(x k1 e 1 k2 e 2 ) =
kZ
(x k) .
8.4
373
Periodicity on Z2 and FIIIZ2 As in the one-dimensional case, IIIZ2 is the tool to use to work with
periodicity. If we form
X
(x k) ,
(x) = ( IIIZ2 )(x) =
kZ2
assuming that the sum converges, then is periodic on the lattice Z2 , or briefly, is Z2 -periodic. This
means that
(x + n) = (x)
for all x and for any lattice point n Z2 , and this is true because
X
X
(x + n) =
(x + n k) =
(x k) = (x) ;
kZ2
kZ2
the sum (or difference) of two lattice points, n k, is a lattice point, so were still summing over Z2 and
we get back .
Using periodicity, and the fact that Z2 is particularly evenly spaced as a set of points in R2 leads to the
important and remarkable formula
FIIIZ2 = IIIZ2
corresponding precisely to the one-dimensional case. Ill put the details of the derivation of this in Section
??. Its also true that
F 1 IIIZ2 = IIIZ2
because IIIZ2 is even.
At this point the most basic version of the two-dimensional sampling formula is already easily within
reach. Its much more interesting, however, as well as ultimately much more useful to allow for some
greater generality.
374
8.4.2
Lattices in general
Z2 isnt the only example of a set of evenly spaced points in the plane, though perhaps its the example
of the most evenly spaced points. Its easy to imagine oblique lattices, too. Not all crystals are square,
after all, or even rectangular, and we want to be able to use general lattices to model crystals. Well now
consider such oblique arrangements, but be warned that the subject of lattices can go on forever; the effort
here is to be brief to the point.
We adopt the vector point of view for defining a general lattice. Take any basis u 1 , u 2 of R2 and consider
all the points (or vectors) that are integer linear combinations of the two. These form:
Lattice points = p = p1 u 1 + p2 u 2 ,
p1 , p2 = 0, 1, 2, . . .
Well denote such a lattice by L. The sum and difference of two lattice points is again a lattice point, as
is any integer times a lattice point.13
The vectors u 1 and u 2 are said to be a basis for the lattice. Other vectors can also serve as a basis, and
two bases for the same lattice are related by a 2 2 matrix with integer entries having determinant 1. (I
wont go through the derivation of this.) The parallelogram determined by the basis vectors (any basis
vectors) is called a fundamental parallelogram for the lattice, or, in crystallographers terms, a unit cell.
A fundamental parallelogram for Z2 is the square 0 x1 < 1, 0 x2 < 1.14 By convention, one speaks of
the area of a lattice in terms of the area of a fundamental parallelogram for the lattice, and well write
Area(L) = Area of a fundamental parallelogram .
Two fundamental parallelograms for the same lattice have the same area since the bases are related by a
2 2 integer matrix with determinant 1 and the area scales by the determinant.
If we take the natural basis vectors e 1 = (1, 0) and e 2 = (0, 1) for R2 we get the integer lattice Z2 as
before. We can see that any lattice L can be obtained from Z2 via an invertible linear transformation A,
the one that takes e 1 and e 2 to a basis u 1 = Ae 1 and u 2 = Ae 2 that defines L. This is so precisely
because A is linear: if
p = p1 u 1 + p2 u 2 , p1 , p2 integers ,
is any point in L then
p = p1 (Ae 1 ) + p2 (Ae 2 ) = A(p1 e 1 + p2 e 2 ) ,
showing that p is the image of a point in Z2 . We write
L = A(Z2 )
A fundamental parallelogram for L is determined by u 1 and u 2 , and so
Area(L) = Area of the parallelogram determined by u 1 and u 2 = | det A| .
Here, for example, is the lattice obtained from Z2 by applying

3 1
A=
1 2
A basis is u 1 = (3, 1), u 2 = (1, 2) (Draw the basis on the lattice!) The area of the lattice is 7.
13
In mathematical terminology a lattice is a module over Z; a module is like a vector space except that you cant divide by
the scalars (the integers in this case) only add and multiply them. For a module, as opposed to a vector space, the scalars
form a ring, not a field.
14
Its a common convention to define a fundamental parallelogram to be half open, including two sides (x1 = 0 and x2 = 0
in this case) and excluding two (x1 = 1 and x2 = 1). This wont be an issue for our work.
8.4
8.4.3
375
III for a lattice
It doesnt take a great leap in imagination to think about introducing III for a general lattice: If L is a
lattice in R2 then the III function associated with L is
X
IIIL (x) =
(x p) .
pL
So theres your general sum of delta functions at evenly spaced points. We could also write the definition
as
X
IIIL (x) =
(x k1 u 1 k2 u 2 ) .
k1 ,k2 =
As L can be obtained from Z2 via some linear transformation so too can IIIL be expressed in terms of IIIZ2 .
If L = A(Z2 ) then
X
X
IIIL (x) =
(x p) =
(x Ak) .
kZ2
pL
Next, using the formula for (Ax) from earlier in this chapter,
(x Ak) = (A(A1 x k)) =
Therefore
IIIL (x) =
1
(A1 x k)
| det A|
1
III 2 (A1 x) .
| det A| Z
Compare this to our earlier formulas on how the one-dimensional III-function scales: With
IIIp (x) =
k=
(x kp)
376
and
III(px) =
(px k)
k=
we found that
III(px) =
Periodizing and sampling

with respect to the lattice. If
1
III (x)
|p| 1/p
Periodizing with IIIL via convolution results in a function that is periodic

(x) = ( IIIL )(x) =
(x p)
pL
then
(x + p) = (x)
for all x R2 and all p L. Another way of saying this is that has two independent periods, one
each in the directions of any pair of basis vectors for the lattice. Thus if u 1 , u 2 are a basis for L then
(x + k1 u 1 ) = (x) and
(x + k2 u 2 ) = (x),
k1 , k2 any integers.
IIIL is also the tool to use for sampling on a lattice, for

X
(IIIL )(x) =
(p)(x p) .
pL
Were almost ready to use this.

Dual lattices and FIIIL Of the many (additional) interesting things to say about lattices, the one
thats most important for our concerns is how the Fourier transform of IIIL depends on L. This question
leads to a fascinating phenomenon, one that is realized physically in x-ray diffraction images of crystals.
We mentioned earlier that for the integer lattice we have
FIIIZ2 = IIIZ2 .
What about the Fourier transform of IIIL ? We appeal to the general similarity theorem to obtain, for
L = AZ2 ,
1
F(IIIZ2 (A1 x))
| det A|
1
1
FIIIZ2 (AT )
=
| det A| | det A1 |
FIIIL () =
(we just get AT on the inside because were already working with A1 )
= FIIIZ2 (AT )
= IIIZ2 (AT )
Theres a much neater version of this last result, and one of genuine physical importance. But we need a
new idea.
8.4
377
In crystallography it is common to introduce the reciprocal lattice associated to a given lattice. Given a
lattice L, the reciprocal lattice is the lattice L consisting of all points (or vectors) q such that
q p = an integer for every p in the lattice L.
In some other areas of applications, and in mathematics, the reciprocal lattice is known as the dual lattice.
Ill show my heritage and generally use the term dual lattice.
Warning People in crystallography, those in Material Sciences for example, use the reciprocal
lattice all the time and define it this way. However, in some fields and for some applications
the reciprocal lattice is normalized differently to require that q p be an integer multiple of 2.
This alternate normalization is exactly tied up with the alternate ways of defining the Fourier
transform, i.e., while we use e2i x , putting the 2 in the exponential, others do not put the
2 there and have to put a factor in front of the integral, and so on. I can do no more than to
issue this warning and wish us all luck in sorting out the inconsistencies.
To develop the notion of the dual lattice a little, and to explain the terminology reciprocal, suppose we
get the lattice L from Z2 by applying an invertible matrix A to Z2 . Well show that the reciprocal lattice
L of L is given by
L = AT (Z2 ) .
Theres a maxim lurking here. Use of the Fourier transform always brings up reciprocal relations of some
sort, and in higher dimensions more often than not:
Reciprocal means inverse transpose.
Notice, by the way, that (Z2 ) = Z2 , since A in this case is the identity, i.e., Z2 is self-dual as a lattice.
This, coupled with the discussion to follow, is another reason for saying that Z2 wins the award for most
evenly spaced points in R2 .
Heres why L = AT (Z2 ). Suppose q = AT m for some m = (m1 , m2 ) in Z2 . And suppose also, because
L = A(Z2 ), that p = Am for some other m = (m1 , m2 ) in Z2 . Then
q p = AT m Am
= m A1 (Am )
=mm =
m1 m1
(because of how matrices operate with the dot product)

+ m2 m2
(an integer)
We want to draw two conclusions from the result that L = AT (Z2 ). First, we see that
Area(L ) = | det AT | =
1
1
=
.
| det A|
Area(L)
Thus the areas of L and L are reciprocals. This is probably the crystallographers main reason for using
the term reciprocal.
The second conclusion, and the second reason to use the term reciprocal, has to do with bases of L and of
L . With L = A(Z2 ) let
u 1 = Ae 1 , u 2 = Ae 2
be a basis for L. Since L = AT (Z2 ), the vectors
u 1 = AT e 1 ,
u 2 = AT e 2
378
are a basis for L . They have a special property with respect to u 1 and u 2 , namely
u i u j = ij
(Kronecker delta) .
This is simple to show, after all weve been through:

u i u j = Ae i AT e j = e i AT AT e j = e i e j = ij .
Now, in linear algebra independent of any connection with lattices bases {u 1 , u 2 } and {u 1 , u 2 } of
R2 are called dual (or sometimes, reciprocal) if they satisfy
u i u j = ij
(Kronecker delta) .
We can therefore summarize what weve found by saying

If {u 1 , u 2 } is a basis for a lattice L and if {u 1 , u 2 } is the dual basis to {u 1 , u 2 }, then {u 1 , u 2 } is a
basis for the dual lattice L .
Lots of words here, true, but its worth your while understanding what weve just done. Youre soon to
see it all in action in the sampling formula.
Heres a picture of the dual lattice to the lattice pictured earlier. Its obtained from Z2 by applying

2/7 1/7
T
A =
.
1/7 3/7
As the scales on the axes show, the dual lattice is, in this case, much more compressed than the original
lattice. Its area is 1/7.
8.4
379
Back to the Fourier transform. We showed that if L = A(Z2 ) then

FIIIL () = IIIZ2 (AT ) .
We want to call forth the reciprocal lattice. For this,
X
IIIZ2 (AT ) =
(AT n)
nZ2
(AT ( AT n))
nZ2
X
X
1
1
T
(
A
n)
=
( AT n) .
| det AT |
|
det
A|
2
2
nZ
nZ
But this last expression is exactly a sum over points in the reciprocal lattice L . We thus have
F(IIIL )() =
1
IIIL () .
| det A|
Bringing in the areas of fundamental parallelograms for L and L we can write this either in the form
F(IIIL )() = Area(L )IIIL () or
Area(L)F(IIIL )() = IIIL () .
Interchanging the roles of L and L , we likewise have

F(IIIL )() = Area(L)IIIL ()
or
Area(L )F(IIIL )() = IIIL () .
380
Formulas for the inverse Fourier transforms look just like these because the IIIs are even.
Compare these results to the formula in one dimension,
1
FIIIp = III1/p ,
p
and now youll see why I said Pay particular attention here to the reciprocity in spacing between IIIp and
its Fourier transform at the beginning of this section.
Higher dimensions Everything in the preceding discussion goes through in higher dimensions with no
significant changes, e.g., area becomes volume. The only reason for stating definitions and results in
two-dimensions was to picture the lattices a little more easily. But, certainly, lattices in three dimensions
are common in applications and provide a natural framework for understanding crystals, for example. Lets
do that next.
8.4.4
The Poisson Summation Formula, again, and F IIIZ2
Back in Chapter 5 we derived the Poisson summation formula: if is a Schwartz function then
F(k) =
(k) .
k=
k=
Its a remarkable identity and its the basis for showing that
FIII = III
for the one-dimensional III. In fact, the Poisson summation formula is equivalent to the Fourier transform
identity.
The situation in higher dimensions is completely analogous. All that we need is a little bit on higher
dimensional Fourier series, which well bring in here without fanfare; see the earlier section on Higher
dimensional Fourier series and random walks for more background.
Suppose is a Schwartz function on R2 . We periodize to be periodic on the integer lattice Z2 via
X
(x n) .
(x) = ( IIIZ2 )(x) =
nZ2
Then has a two-dimensional Fourier series

(x) =
kZ2
2ikx
b
(k)e
.
Lets see what happens with the Fourier coefficients.

Z 1Z 1
b
(k1 , k2 ) =
e2i(k1 x1 +k2 x2 ) (x1 , x2 ) dx1 dx2
0
Z 1Z
2i(k1 x1 +k2 x2 )
(x1 n1 , x2 n2 ) dx1 dx2
n1 ,n2 =
Z 1Z
n1 ,n2 = 0
e2i(k1 x1 +k2 x2 ) (x1 n1 , x2 n2 ) dx1 dx2 .
8.5
Crystals
381
Now we make the change of variables u = x1 n1 , v = x2 n2 . We can either do this separately (because
the variables are changing separately) or together using the general change of variables formula.15 Either
way, the result is
n1 ,n2 = 0
1Z 1
e2i(k1 x1 +k2 x2 ) (x1 n1 , x2 n2 ) dx1 dx2
=
=
=
=
1n1
n1 ,n2 = n1
Z 1n1
n1 ,n2 = n1
Z 1n1
1n2
1n2
1n2
e2i(k1 n1 +k2 n2 ) e2i(k1 u+k2 v) (u, v) du dv
n2
n2
n1 ,n2 = n1
Z
2i(k1 u+k2 v)
e2i(k1 (u+n1 )+k2 (v+n2 ) (u, v) du dv
n2
e2i(k1 u+k2 v) (u, v) du dv

(u, v) du dv
= F(k1 , k2 ) .
We have found, just as we did in one dimension, that the Fourier series for the Z2 -periodization of is
X
(x) =
F(k)e2i kx .
kZ2
We now evaluate (0) in two ways, plugging x = 0 into its definition as the periodization of and into
its Fourier series. The result is
X
X
F(k) =
(k) .
kZ2
kZ2
To wrap it all up, heres the derivation of

FIIIZ2 = IIIZ2
based on the Poisson summation formula. For any Schwartz function ,
X
X
hFIIIZ2 , i = hIIIZ2 , Fi =
F(k) =
(k) = hIIIZ2 , i .
kZ2
8.5
kZ2
Crystals
In a few paragraphs, heres one reason why all this stuff on dual lattices is so interesting. The physical
model for a crystal is a three-dimensional lattice with atoms at the lattice points. An X-ray diffraction
experiment scatters X-rays off the atoms in the crystal and results in spots on the X-ray film, of varying
intensity, also located at lattice points. From this and other information the crystallographer attempts to
deduce the structure of the crystal. The first thing the crystallographer has to know is that the lattice of
spots on the film arising from diffraction is the dual of the crystal lattice. (In fact, its more complicated
15
Right here is where the property of Z2 as the simplest lattice comes in. If we were working with an oblique lattice we
could not make such a simple change of variables. We would have to make a more general linear change of variables. This
would lead to a more complicated result.
382
than that, for it is the projection onto the two-dimensional plane of the film of the three-dimensional dual
lattice.)
We can explain this phenomenon atoms on a lattice, spots on the dual lattice via the Fourier
transform. What the crystallographer ultimately wants to find is the electron density distribution for the
crystal. The mathematical model for crystals puts a delta at each lattice point, one for each atom. If we
describe the electron density distribution of a single atom by a function (x) then the electron density
distribution of the crystal with atoms at points of a (three-dimensional) lattice L is
X
L (x) =
(x p) = ( IIIL )(x) .
pL
This is now a periodic function with three independent periods, one in the direction of each of the three
basis vectors that determine L. We worked with a one-dimensional version of this in Chapter 5.
The basic fact in X-ray crystallography is that the scattered amplitude of the X-rays diffracting off the
crystal is proportional to the magnitude of the Fourier transform of the electron density distribution. This
data, the results of X-ray diffraction, comes to us directly in the frequency domain. Now, we have
FL () = F()FIIIL () = F() Volume(L ) IIIL () ,
where L is the dual lattice. Taking this one more step,
X
FL () = Volume(L )
F(q)( q) .
qL
The important conclusion is that the diffraction pattern has peaks at the lattice points of the reciprocal
lattice. The picture is not complete, however. The intensities of the spots are related to the magnitude
of the Fourier transform of the electron density distribution, but for a description of the crystal it is also
necessary to determine the phases, and this is a hard problem.
Heres a picture of a macroscopic diffraction experiment. On the left is an array of pinholes and on the
right is the diffraction pattern. The spots on the right are at the lattice points of the reciprocal lattice.
The goal of X-ray diffraction experiments is to determine the configuration of atoms from images of this
type. Making the analysis even harder is that for 3D crystal lattices the images on an X-ray film are the
projection onto the image plane of the 3D configuration. Just how difficult it may be to infer 3D structure
from 2D projections is illustrated by a famous experiment: Fun in Reciprocal Space published in the
distinguished American journal The New Yorker.
8.6
8.6
Bandlimited Functions on R2 and Sampling on a Lattice
383
Lets develop the sampling formula in two dimensions. A function f on R2 is bandlimited if Ff is identically
zero outside of some bounded region. We always assume that f is real valued, and hence Ff () = Ff ().
Thus, as weve pointed out before, if Ff () 6= 0 then Ff () 6= 0 and so, as a point set in R2 , the spectrum
is symmetric about the origin.
We want to derive a sampling formula associated with a lattice L by following the recipe of first periodizing
Ff via IIIL , then cutting off, and then taking the inverse Fourier transform. The result will be a sinc
reconstruction of f from its sampled values but just where those sampled values are is whats especially
interesting and relevant to what weve just done.
To get the argument started we assume that the support of Ff lies in a parallelogram. This parallelogram
determines a fundamental parallelogram for a lattice L, and the spectrum gets shifted parallel to itself
and off itself through convolution with IIIL . This periodization is the first step and its analogous to the
one-dimensional case when the spectrum lies in an interval, say from p/2 to p/2, and the spectrum gets
shifted around and off itself through convolution with IIIp . Recall that the crucial limitation is that the
spectrum only goes up to p/2 and down to p/2, while IIIp has s spaced p apart. The spacing of the
s is big enough to shift the spectrum off itself and no smaller spacing will do. Correspondingly in two
dimensions, the parallelogram containing the spectrum determines a lattice with big enough spacing for
a III based on the lattice to shift the spectrum off itself.
Using the general stretch theorem, well be able to get the general result by first deriving a special case,
when the spectrum lies in a square. Suppose, then, that Ff () is identically zero outside the (open) square
|1 | < 1/2, |2 | < 1/2. We work with the integer lattice Z2 with basis e 1 and e 2 . The (open) fundamental
parallelogram for Z2 is 0 < 1 < 1, 0 < 2 < 1 and the spectrum is inside the center fourth of four copies
of it, as pictured.
Periodizing Ff by IIIZ2 shifts the spectrum off itself, and no smaller rectangular lattice will do for this.
384
We then cut off by the two-dimensional rect function (x1 , x2 ) = (x1 )(x2 ) and this gives back Ff :
Ff () = (1 ) (2 )(Ff IIIZ2 )() .
This is just as in the one-dimensional case, and now its time to take the inverse Fourier transform. Using
FIIIZ2 = IIIZ2 , or rather F 1 IIIZ2 = IIIZ2 , and invoking the convolution theorem we obtain
f (x) = f (x1 , x2 ) = (sinc x1 sinc x2 ) (f (x) IIIZ2 (x))

X
= (sinc x1 sinc x2 ) f (x)
(x k1 e 1 k2 e 2 )
k1 ,k2 =
= (sinc x1 sinc x2 )
f (k1 , k2 )(x1 k1 , x2 k2 )
k1 ,k2 =
f (k1 , k2 ) sinc(x1 k1 ) sinc(x2 k2 ) .
k1 ,k2 =
In solidarity with the general case soon to follow, lets write this square sampling formula as
f (x) =
f (k1 e 1 + k2 e 2 ) sinc(x e 1 k1 ) sinc(x e 2 k2 ) .
k1 ,k2 =
Now suppose that the spectrum of Ff lies in the (open) parallelogram, as pictured, with u 1 and u 2 parallel
to the sides and as long as the sides.
Let A be the 2 2 matrix that takes e 1 to u 1 and e 2 to u 2 , so that A maps the lattice Z2 to the lattice L
with basis u 1 and u 2 . Let B = AT (hence B T = A) and remember that B takes Z2 to the dual lattice
L of L. A basis for L (the dual basis to u 1 and u 2 ) is
u 1 = Be 1 ,
u 2 = Be 2 .
8.6
385
Next let
g(x) = f (Bx) .
According to the general stretch theorem,
Fg() =
1
Ff (B T ) = | det A| Ff (A) .
| det B|
The determinant factor out front doesnt matter; whats important is that the spectrum of g is in the
square 1/2 < 1 < 1/2, 1/2 < 2 < 1/2, since the corresponding points A lie in the parallelogram
containing the spectrum of f , i.e., Fg is identically zero outside the square.
We now apply the square sampling formula to g to write
g(x) =
g(k1 e 1 + k2 e 2 ) sinc(x e 1 k1 ) sinc(x e 2 k2 )
k1 ,k2 =
With y = Bx we can then say

f (y) =
=
=
=
f (B(k1 e 1 + k2 e 2 )) sinc(B 1 y e 1 k1 ) sinc(B 1 y e 2 k2 )
k1 ,k2 =
f (k1 Be 1 + k2 Be 2 ) sinc(AT y e 1 k1 ) sinc(AT y e 2 k2 )
k1 ,k2 =
f (k1 u 1 + k2 u 2 ) sinc(y Ae 1 k1 ) sinc(y Ae 2 k2 )
k1 ,k2 =
f (k1 u 1 + k2 u 2 ) sinc(y u 1 k1 ) sinc(y u 2 k2 ) .
k1 ,k2 =
Were done. Change y to x for psychological comfort, and the lattice sampling formula says that
f (x) =
f (k1 u 1 + k2 u 2 ) sinc(x u 1 k1 ) sinc(x u 2 k2 ).
(8.1)
k1 ,k2 =
This is a sinc reconstruction formula giving the function in terms of sample values on a lattice. But its
the dual lattice! Heres how to remember the highlights:
386

The spectrum of f lies in a parallelogram, which determines a lattice with basis u 1 and u 2 .
That lattice determines a dual lattice (in the spatial domain) with dual basis u 1 and u 2 .
The sincs use data from the lattice, while the sample points are exactly the points in the dual lattice.
Look back at the one-dimensional sampling formula and tell yourself what you see of this picture.
Exercise
mula?
What should we mean by sampling rate vis `a vis the two-dimensional lattice sampling for-
The next topics on this path would be to investigate aliasing and to consider the case of a finite spectrum
and finite sampling. Another time, another class.
8.7
Naked to the Bone
Our final topic in the course will be a quick development of the use of the Fourier transform in medical
imaging. Well find that the two-dimensional Fourier transform is perfectly suited to the problem of
recovering a density function a function representing bones, internal organs, the whole lot from the
projections of that density obtained by passing parallel beams of X-rays through a two-dimensional cross
section of the body. (For the discussion of the use of the Fourier transform Im not making a distinction
between the original methods of tomography using X-rays and those of magnetic resonance imaging.)
For an account of the history of medical imaging, I recommend the book Naked to the Bone: Medical
Imaging in the Twentieth Century by Bettyann Kevles, from which I stole the title of this section.
Dimmer and dimmer What happens when light passes through murky water? It gets dimmer and
dimmer the farther it goes, of course this is not a trick question. If the water is the same murkiness
throughout, meaning, for example, uniform density of stuff floating around in it, then its natural to assume
that the intensity of light decreases by the same percent amount per length of path traveled. Through
absorption, scattering, etc., whatever intensity comes in, a certain percentage of that intensity goes out;
over a given distance the murky water removes a percentage of light, and this percentage depends only on
the distance traveled and not on where the starting and stopping points are.16 Were assuming here that
light is traveling in a straight line through the water.
Constant percent change characterizes exponential growth, or decay, so the attenuation of the intensity of
light passing through a homogeneous medium is modeled by
I = I0 ex ,
where I0 is the initial intensity, x is the distance traveled, and is a (positive) murkiness constant. x has
dimension of length and has dimension 1/length and units murkiness/length. is constant because
we assume that the medium is homogeneous. We know the value of I0 , and one measurement of x and I
will determine . In fact, what we do is to put a detector at a known distance x and measure the intensity
when it arrives at the detector.
16
Optical fibers provide an interesting and important study in the progress of making something glass in this case less
murky. In the Appendix 8.12 Ive attached a graph showing just how dramatic the progress has been.
8.7
Naked to the Bone
387
Now suppose the water is not uniformly murky, but rather the light passes through a number of layers,
each layer of uniform murkiness. If the ith layer has murkiness constant i and is of length xi , and if
there are n layers, then the intensity of light that reaches the detector can be modeled by

X
n
i xi .
I = I0 exp
i=1
Clearly, if the murkiness is described by a function (x), then the intensity arriving at the detector is
modeled by

Z
I = I0 exp (x) dx ,
L
where L is the line the light travels along. Its common to call the number

Z
I
(x) dx = ln
p=
I
0
L
the attenuation coefficient.
Can we recover the density function (x) from knowledge of the intensity? Not so easily. Certainly not
from a single reading many arrangements of murkiness along the path could result in the same final
intensity at the detector.
If we could vary the detector along the path and record the results then we would be able to determine
(x). That is, if we could form
Z
p() =
(x) dx ,
0
as a function of a variable position along the line (0 is some fixed starting point the source) then we
could find from p by finding the derivative p . The trouble is moving the detector through the murky
water along the path.
Tomography X-rays are light, too, and when they pass through murky stuff (your body) along a straight
line they are attenuated and reach a detector on the other end at a reduced intensity. We can continue to
assume that the attenuation, the decrease in intensity, is exponentially decreasing with the path length.
The exponential of what? What do the X-rays pass through?
From the start we set this up as a two-dimensional problem. Take a planar slice through your body. The
gunk in this two-dimensional slice bones, organs, other tissue is of variable density; lets say its
described by an unknown function (x1 , x2 ). We consider (x1 , x2 ) to be zero outside the section of the
body. Take a line L through this slice in the plane of the slice, the path that an X-ray would follow
and parameterize the line by x1 (s), x2 (s), where s is the arclength parameter going from s0 to s1 . (The
arclength parameter means that we move along the line at unit speed.) Then the density along the line
is (x1 (s), x2 (s)) and the attenuation of the X-ray intensity along the line is

Z s1
(x1 (s), x2 (s)) ds
I = I0 exp
s0
Instead of writing out the parameters and limits, we often write the integral simply as
Z
(x1 , x2 ) ds .
L
Well refer to this as a line integral of along L.
388

The fundamental problem of tomography17 is to determine the function (x, y) from these line integrals, taken over many lines through the region.
For example whats inside?
In trying to solve this problem, whats not allowed is to move the detector through the body thats not
covered by HMO plans. What is allowed is to rotate the source (and the detector) to get X-rays circling
around the two-dimensional cross-section of the body, and what well have are families of parallel X-rays.
Before laying all this out, it pays to organize our study of the problem.
8.8
The Radon Transform
For each line L, cutting through the slice, the integral

Z
(x1 , x2 ) ds
L
is a number. The operation line determines number thus defines a real-valued function of L. The whole
subject of tomography is about this function. To work with it effectively we need to be able describe the
set of all lines not the (Cartesian) equation of a given line, but some kind of parametric description for
the collection of lines. This will allow us to write the integral as a function of these parameters.
There are many ways to describe the collection of all lines in the plane. One that may seem most natural
to you is to use the slope-intercept form for the equation of a line; a line can be written as y = mx + b
where m is the slope and b is the y-intercept. A line can thus be associated with a unique pair (m, b) and
17
tomos means section in Greek
8.8
The Radon Transform
389
vice versa. Theres a catch here, however vertical lines (lines x = constant, infinite slope) are left out
of this description.
Another approach, one that allows us to describe all lines and that is well suited for the function of L,
above, goes as follows. First, a line through the origin is determined by its unit normal vector n. Now,
n and n determine the same line, so we represent all the (distinct) normal vectors as (cos , sin ) for
an angle satisfying 0 < , measured counterclockwise from the x1 -axis. In other words, there is
a one-to-one correspondence between the s with 0 < and the collection of all lines through the
origin.
A line not through the origin can then be described by its unit normal vector together with the directed
distance of the line from the origin, a positive number if measured in the direction of n and a negative
number if measured in the direction n. Call this directed distance . Thus < < .
The set of pairs (, ) provides a parameterization for the set of all lines in the plane. Once again:
A pair (, ) means, in this context, the unique line with normal vector n = (cos , sin ) which is at
a directed distance from the origin, measured in the direction n if > 0 and in the direction n
if < 0.
Anytime youre confronted with a new coordinate system you should ask yourself what the situation is
when one of the coordinates is fixed and the other is free to vary. In this case, if is fixed and varies we
390
get a family of parallel lines.
For the other case, when is fixed, we have to distinguish some cases. The pairs (0, ) correspond to lines
through the origin. When is positive and varies from 0 to (including 0, excluding ) we get the
family of lines tangent to the upper semicircle of radius (including the tangent at (, 0) excluding the
tangent at (, 0)). When < 0 we get lines tangent to the lower semicircle (including the tangent at
(||, 0), excluding the tangent at (||, 0)).
Using the coordinates (, ) we therefore have a transform of the function (x1 , x2 ) to a function R(, )
defined by
Z
(x1 , x2 ) ds .
R(, ) =
L(,)
This is called the Radon transform of , introduced by Johann Radon way back in 1917! The fundamental question of tomography can then be stated as:
Is there an inversion formula for the Radon transform? That is, from knowledge of the values R(, )
can we recover ?
8.9
Getting to Know Your Radon Transform
391
Weve indicated the dependence of the integral on and by writing L(, ), but we want to use the
coordinate description of lines to write the integral in a still more convenient form. Using the dot product,
the line determined by (, ) is the set of points (x1 , x2 ) with
= x n = (x1 , x2 ) (cos , sin ) = x1 cos + x2 sin .
or described via the equation
x1 cos x2 sin = 0 ,
< x1 < , < x2 < .
Now consider the delta function along the line, that is,
( x1 cos x2 sin )
as a function of x1 , x2 . This is also called a line impulse and its an example of the greater variety one has
in defining different sorts of s in two-dimensions. With some interpretation and argument (done in those
notes) one can show that integrating a function f (x1 , x2 ) against the line impulse associated with a line
L results precisely in the line integral of f along L. This is all well need here, and with that the Radon
transform of (x1 , x2 ) can be expressed as
Z Z
R()(, ) =
(x1 , x2 )( x1 cos x2 sin ) dx1 dx2 .
This is the form well most often work with. One also sees the Radon transform written as
Z
(x)( x n) dx .
R()(, n) =
R2
This expression suggests generalizations to higher dimensions interesting, but we wont pursue them.
Projections Its often convenient to work with R()(, ) by first fixing and letting vary. Then were
looking at parallel lines passing through the domain of , all perpendicular to a particular line making
an angle with the x1 -axis (that line is the common normal to the parallel lines), and we compute the
integral of along these lines.
This collection of values, R()(, ) with fixed, is often referred to as a projection of , the idea being
that the line integrals over parallel lines at a fixed angle are giving some kind of profile, or projection, of
in that direction.18 Then varying gives us a family of projections, and one speaks of the inversion
problem as determining (x1 , x2 ) from its projections.
This is especially apt terminology for the medical applications, since thats how a scan is made:
1. Fix an angle and send in a bunch of parallel X-rays at that angle.
2. Change the angle and repeat.
8.9
We want to develop a few properties of the Radon transform, just enough to get some sense of how to work
with it. First, a few comments on what kinds of functions (x1 , x2 ) one wants to use; its interesting but
we wont make an issue of it.
18
Important: Dont be fooled by the term projection. You are not geometrically projecting the shape of the twodimensional cross section (that the lines are cutting through). You are looking at the attenuated, parallel X-rays that emerge
as we move a source along a line. The line is at some angle relative to a reference axis.
392
Inspired by honest medical applications, we would not want to require that the cross-sectional density
(x1 , x2 ) be smooth, or even continuous. Jump discontinuities in (x1 , x2 ) correspond naturally to a change
from bone to muscle, etc. Although, mathematically speaking, the lines extend infinitely, in practice the
paths are finite. In fact, the easiest thing is just to assume that (x1 , x2 ) is zero outside of some region
its describing the density of a slice of a finite extent body, after all.
Examples There arent too many cases where one can compute the Radon transform explicitly. One
example is the circ function, expressed in polar coordinates as
(
1
r1
circ(r) =
0
r>1
We have to integrate the circ function along any line. Think in terms of projections, as defined above.
From the circular symmetry, its clear that the projections are independent of .
Because of this we can take any convenient value of , say = 0, and find the integrals over the parallel
lines in this family. The circ function is 0 outside the unit circle, so we need only to find the integral (of
the function 1) over any chord of the unit circle parallel to the x2 -axis. This is easy. If the chord is at a
distance from the origin, || 1, then
R(1)(, 0) =
1 2
1 2
p
1 dx2 = 2 1 2 .
8.9
Thus for any (, ),
393
( p
2 1 2
R circ(, ) =
0
|| 1
|| > 1
Gaussians again Another example where we can compute the Radon transform exactly is for a Gaussian:
2
g(x1 , x2 ) = e(x1 +x2 ) .

Any guesses as to what Rg is? Lets do it.
Using the representation in terms of the line impulse we can write
Z Z
2
2
e(x1 +x2 ) ( x1 cos x2 sin ) dx1 dx2 .
Rg(, ) =
We now make a change of variables in this integral, putting

u1 = x1 cos + x2 sin ,
u2 = x1 sin + x2 cos .
This is a rotation of coordinates through an angle , making the u1 -axis correspond to the x1 -axis. The
Jacobian of the transformation is 1, and we also find that
u21 + u22 = x21 + x22 .
394
In the new coordinates the integral becomes:

Z Z
2
2
Rg(, ) =
e(u1 +u2 ) ( u1 ) du1 du2

Z Z
2
u21
e
( u1 ) du1 eu2 du2
=
2
2
e eu2 du2 (by the sifting property of )
=
Z
2
2
eu2 du2
=e
=e
(because the Gaussian is normalized to have area 1)
Writing this in polar coordinates, r = x21 + x22 , we have shown that

2
R(er ) = e .
How about that.
Linearity, Shifts, and Evenness We need a few general properties of the Radon transform.
Linearity:
integrand.
R(f + g) = R(f ) + R(g). This holds because integration is a linear function of the
8.10
Appendix: Clarity of Glass
Shifts:
395
This is a little easier to write (and to derive) in vector form. Let n = (cos , sin ). The result is
R((x b)) = (R)( b n, )
In words: shifting x by b has the effect of shifting each projection a distance b n in the -variable.
To derive this we write the definition as
R((x b)) =
(x b)( x n) dx
R2
If b = (b1 , b2 ) then the change of variable u1 = x1 b1 and u2 = x2 b2 , or simply u = x b with

u = (u1 , u2 ), converts this integral into
Z
(u)( (u + b) n) du
R((x b)) =
2
ZR
(u)( u n b n)) du
=
R2
= (R)( b n, )
Evenness: Finally, the Radon transform always has a certain symmetry it is always an even function
of and . This means that
R(, + ) = R(, ) .
Convince yourself that this makes sense in terms of the projections. The derivation goes:
Z Z
R(, + ) =
(x1 , x2 )( x1 cos( + ) x2 sin( + )) dx1 dx2

Z Z
=
(x1 , x2 )( x1 ( cos ) x2 ( sin )) dx1 dx2
Z
Z
=
(x1 , x2 )( + x1 cos + x2 sin ) dx1 dx2

Z Z
=
(x1 , x2 )( x1 cos x2 sin ) dx1 dx2 (because is even)
= R(, )
8.10
Appendix: Clarity of Glass
Heres a chart showing how the clarity of glass has improved over the ages, with some poetic license in
estimating the clarity of the windows of ancient Egypt. Note that on the vertical axis on the left the tick
marks are powers of 10 but the units are in decibels which already involve taking a logarithm! The big
jump in clarity going to optical fibers was achieved largely by eliminating water in the glass.
396
8.11
Medical Imaging: Inverting the Radon Transform
Lets recall the setup for tomography. We have a two-dimensional region (a slice of a body) and a density
function (x1 , x2 ) defined on the region. The Radon transform of is obtained by integrating along
lines that cut across the region. We write this as
Z Z
R(, ) =
(x1 , x2 )( x1 cos x2 sin ) dx1 dx2 .
Here (, ) are coordinates that specify a line; (0 < ) is the angle the normal to the line makes with
the x1 -axis and ( < < ) is the directed distance of the line from the origin. (x1 cos x2 sin )
is a line impulse, a -function along the line whose (Cartesian) equation is x1 cos x2 sin = 0.
If we fix and vary , then R(, ) is a collection of integrals along parallel lines through the region, all
making the same angle, + /2, with a reference axis, the x1 -axis. This set of values is referred to as a
projection of . Thus one often speaks of the Radon transform as a collection of projections parameterized
by an angle .
In practice (x1 , x2 ) is unknown, and what is available are the values R(, ). These values (or rather
a constant times the exponential of these values) are what your detector registers when an X-ray reaches
8.11
397
it having gone through the region and having been attenuated according to its encounter with (x1 , x2 ).
The problem is to reconstruct (x1 , x2 ) from these meter readings, in other words to invert the Radon
transform.
Among those who use these techniques, (x1 , x2 ) is often referred to simply as an image. In that terminology the problem is then to reconstruct the image from its projections.
The Projection-Slice Theorem The inversion problem is solved by a result that relates the twodimensional Fourier transform of to a one-dimensional Fourier transform of R(), taken with respect
to . Once F is known, can be found by Fourier inversion.
The formulation of this relation between the Fourier transforms of an image and its projections is called
the Projection-Slice Theorem 19 and is the cornerstone of tomography. Well go through the derivation,
but it must be said at once that, for practical applications, all of this has to be implemented numerically,
i.e., with the DFT (and the FFT). Much of the early work in Computer Assisted Tomography (CAT)
was in finding efficient algorithms for doing just this. An important issue are the errors introduced by
approximating the transforms, termed artifacts when the reconstructed image (x1 , x2 ) is drawn on a
screen. We wont have time to discuss this aspect of the problem.
Starting with
R(, ) =
(x1 , x2 )( x1 cos x2 sin ) dx1 dx2 ,
what is its Fourier transform with respect to , regarding as fixed? For lack of a better notation, we
write this as F (R()). Calling the frequency variable r dual to we then have
Z
e2ir R(, ) d
F R()(r, ) =
Z
Z Z
2ir
e
=
(x1 , x2 )( x1 cos x2 sin ) dx1 dx2 d

Z
Z Z
=
( x1 cos x2 sin )e2ir d dx1 dx2
(x1 , x2 )
Z
Z
(x1 , x2 )e2ir(x1 cos +x2 sin ) dx1 dx2
=

Z Z
=
(x1 , x2 )e2i(x1 r cos +x2 r sin ) dx1 dx2
Check out what happened here: By interchanging the order of integration we wind up integrating the line
impulse against the complex exponential e2ir . For that integration we can regard (x1 cos x2 sin )
as a shifted -function, and the integration with respect to produces e2i(x1 r cos +x2 r sin ) . Now if we let
1 = r cos
2 = r sin
the remaining double integral is
Z Z
Z
2i(x1 1 +x2 2 )
e
(x1 , x2 ) dx1 dx2 =
19
Also called the Central Slice Theorem, or the Center Slice theorem.
R2
e2ix (x) dx .
398
This is the two-dimensional Fourier transform of .
We have shown
The Projection-Slice Theorem:
F R()(r, ) = F(1 , 2 ),
1 = r cos , 2 = r sin .
Observe that
r 2 = 12 + 22
and
tan =
2
.
1
This means that (r, ) are polar coordinates for the (1 , 2 )-frequency plane. As varies between 0 and
(including 0, excluding ) and r between and we get all the points in the plane.
Reconstructing the image That last derivation happened pretty fast. Lets unpack the steps in using
the projection-slice theorem to reconstruct an image from its projections.
1. We have a source and a sensor that rotate about some center. The angle of rotation is , where
0 < .
2. A family of parallel X-rays pass from the source through a (planar) region of unknown, variable
density, (x1 , x2 ), and are registered by the sensor.
For each the readings at the meter thus give a function g () (or g(, )), where is the (directed)
distance that a particular X-ray is from the center of the beam of parallel X-rays.
Each such function g , for different s, is called a projection.
3. For each we compute Fg (r), i.e., the Fourier transform of g () with respect to .
4. Since g () also depends on so does its Fourier transform. Thus we have a function of two variables,
G(r, ), the Fourier transform of g (). The projection-slice theorem tells us that this is the Fourier
transform of :
F(1 , 2 ) = G(r, ), where 1 = r cos , 2 = r sin .
Thus (F)(1 , 2 ) is known.
5. Now take the inverse two-dimensional Fourier transform to recover :
Z
e2ix F() d .
(x) =
R2
Running the numbers Very briefly, lets go through how one might set up a numerical implementation
of the procedure weve just been through. The function that we know is g(, ) thats what the sensor
gives us, at least in discrete form. To normalize things we suppose that g(, ) is zero for || 1. This
means, effectively, that the region were passing rays through is contained within the circle of radius one
the region is bounded so we can assume that it lies within some disk, so we scale to assume the the
region lies within the unit disk.
8.11
399
Suppose we have M equal angles, j = j/M , for j = 0, . . . , M 1. Suppose next that for each angle we
send through N X-rays. Were assuming that 1 1, so the rays are spaced = 2/N apart and we
index them to be
N
N
2n
, n = ,..., 1.
n =
N
Then our projection data are the M N values

gnj = g(n , j ) ,
j = 0, . . . , M 1 , n =
N
N
,..., 1.
2
2
The first step in applying the projection slice theorem is to find the one-dimensional Fourier transform of
g(, j ) with respect to , which, since the function is zero for || 1, is the integral
Fg(r, j ) =
e2ir g(, j ) d .
We have to approximate and discretize the integral. One approach to this is very much like the one we
took in obtaining the DFT (Chapter 6). First, were integrating with respect to , and we already have
sample points at the n = 2n/N ; evaluating g at those points gives exactly gnj = g(n , j ). Well use these
for a trapezoidal rule approximation.
We also have to discretize in r, the frequency variable dual to . According to the sampling theorem, if
we want to reconstruct Fg(r, j ) from its samples in r the sampling rate is determined by the extent of
g(, j ) in the spatial domain, where the variable is limited to 1 1. So the sampling rate in r
is 2 and the sample points are spaced 1/2 apart:
rm =
m
,
2
m=
N
N
,..., 1.
2
2
The result of the trapezoidal approximation using n = 2n/N and of discretizing in r using rm = m/2 is
2
Fg(rm , j )
N
2
N
N/2
X
e2in rm gnj
n=N/2+1
N/2
X
e2inm/N gnj .
n=N/2+1
(The 2 in 2/N comes in from the form of the trapezoidal rule.) Up to the constant out front, this is a DFT
of the sequence (gnj ), n = N/2 + 1, . . . , N/2. (Here n is varying, while j indexes the projection.) That
is,
2
Fg(rm , j ) F (gnj )[m] .
N
Computing this DFT for each of the M projections j (j = 0, . . . , M 1) gives the data Fg(rm , j ). Call
this
Gmj = F (gnj )[m] .
The next step is to take the two-dimensional inverse Fourier transform of the data Gmj . Now theres an
interesting problem that comes up in implementing this efficiently. The Gmj are presented as data points
based on a polar coordinate grid in the frequency domain:
400
The vertices in this picture are the points (rm , j ) and thats where the data points Gmj live. However,
efficient FFT algorithms depend on the data being presented on a Cartesian grid. One way this is often
done is to manufacture data at Cartesian grid points by taking a weighted average of the Gmj at the polar
grid points which are nearest neighbors:
GCartesian = wa Ga + wb Gb + wc Gc + wd Gd .
Choosing the weighting factors wa , wb , wc and wc is part of the art, but the most significant introductions
of error in the whole process come from this step.
The final picture is then created by
(grid points in spatial domain) = F 1 (GCartesian ) .
8.11
401
This is your brain. This is your brain on Fourier transforms Here are some pictures of a Fourier
reconstruction of a model brain.20 . The brain is modeled by a high density elliptical shell (the skull)
with lower density elliptical regions inside.
Its possible to compute explicity the Radon transform for lines going through an elliptical region, so the
sampling can be carried out based on these formulas. There are 64 projections (64 j s) each sampled at 64
points (64 n s) in the interval [1, 1]. Heres the plot of the values of the projections (the Radon transforms
along the lines). As in pictures of the (Fourier) spectrum of images, the values here are represented via
shading; white represents large values and black represents small values. The horizontal axis is and the
vertical is .
20
See the paper: L A. Shepp and B. F. Logan, The Fourier reconstruction of a head section, IEEE Trans. Nucl. Sci., NS-21
(1974) 2143.
402
And here is the reconstructed brain.

Chap8 PDF

Uploaded by

Copyright:

Available Formats

Chap8 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chap8 PDF

Uploaded by

Copyright:

Available Formats

Chapter 8

n-dimensional Fourier Transform

Space, the Final Frontier

Chapter 8 n-dimensional Fourier Transform

The Fourier transform

The inverse Fourier transform of a function g() is

so for two dimensions,

e2i(x1 1 +x2 2 ) f (x1 , x2 ) dx1 dx2 .

Space, the Final Frontier

To be neater, we again use the notation

Chapter 8 n-dimensional Fourier Transform

at unit speed. The dot product of x(t) and is

Space, the Final Frontier

and the complex exponential is a function of t along the line:

Chapter 8 n-dimensional Fourier Transform

Space, the Final Frontier

Chapter 8 n-dimensional Fourier Transform

Space, the Final Frontier

Finding a few Fourier transforms: separable functions

Chapter 8 n-dimensional Fourier Transform

We can factor (x1 , x2 ) as the product of two one-dimensional rect functions:

Space, the Final Frontier

For the Fourier transform of the 2-dimensional we then have

Chapter 8 n-dimensional Fourier Transform

The direct generalization of the (basic) rect function to n dimensions is

Fg() = e1 e2 en = e(1 +2 ++n ) = e|| .

Getting to Know Your Higher Dimensional Fourier Transform

As for one dimension, we see that g is its own Fourier transform.

Getting to Know Your Higher Dimensional Fourier Transform

Chapter 8 n-dimensional Fourier Transform

(Remember, b involves b.) Each to their own taste.

If f (x) F () then f (x b) e2ibF ().

Getting to Know Your Higher Dimensional Fourier Transform

and theres our formula.

Chapter 8 n-dimensional Fourier Transform

Ill skip the derivation.

Getting to Know Your Higher Dimensional Fourier Transform

Finally, Im of a mind to introduce the general scaling operator defined by

An important special case of the stretch theorem is when A is a rotation matrix:

Chapter 8 n-dimensional Fourier Transform

Getting to Know Your Higher Dimensional Fourier Transform

Applying the shift theorem next gives

I wont give a second version of the second derivation.

Chapter 8 n-dimensional Fourier Transform

A little now, more later

(x1 )q1 (x2 )q2

The -function is the distribution defined by the pairing

Getting to Know Your Higher Dimensional Fourier Transform

by definition of the 2-dimensional delta function. On the other hand,

Chapter 8 n-dimensional Fourier Transform

Scaling delta functions

Recall how a one-dimensional delta function scales:

The Fourier transform of a radial function

e2i(x1 1 +x2 2 ) f (x1 , x2 ) dx1 dx2

Getting to Know Your Higher Dimensional Fourier Transform

does not depend on .8 Consequently,

e2ir cos d = 2J0 (2r)

and the Fourier transform of f (r) is

Convince yourself of this; for instance let G() =