Circulant-Matrices

September 7, 2017

In [1]: using PyPlot, Interact

1 Circulant Matrices
In this lecture, I want to introduce you to a new type of matrix: circulant matrices. Like Hermitian
matrices, they have orthonormal eigenvectors, but unlike Hermitian matrices we know exactly what their
eigenvectors are! Moreover, their eigenvectors are closely related to the famous Fourier transform and Fourier
series. Even more importantly, it turns out that circulant matrices and the eigenvectors lend themselves to
incredibly efficient algorithms called FFTs, that play a central role in much of computational science and
engineering.
Consider a system of n identical masses m connected by springs k, sliding around a circle without friction.
2
Similar to lecture 26, the vector ~s of displacements satifies m ddt2~s = −kA~s, where A is the n × n matrix:
 
2 −1 −1
−1 2 −1 
 

 −1 2 −1 

A=
 .. .. .. 
 . . . 


 −1 2 −1 

 −1 2 −1
−1 −1 2
(This matrix is real-symmetric and, less obviously, positive semidefinite. So, it should have orthogonal
eigenvectors and real eigenvalues λ ≥ 0.)
For example, if n = 7:

In [2]: A = [ 2 -1 0 0 0 0 -1
-1 2 -1 0 0 0 0
0 -1 2 -1 0 0 0
0 0 -1 2 -1 0 0
0 0 0 -1 2 -1 0
0 0 0 0 -1 2 -1
-1 0 0 0 0 -1 2]

https://github.com/stevengj/1806-spring17/raw/master/lectures/cyc

Figure 1: a ring of springs

1
Out[2]: 7×7 Array{Int64,2}:
2 -1 0 0 0 0 -1
-1 2 -1 0 0 0 0
0 -1 2 -1 0 0 0
0 0 -1 2 -1 0 0
0 0 0 -1 2 -1 0
0 0 0 0 -1 2 -1
-1 0 0 0 0 -1 2

This matrix has a very special pattern: every row is the same as the previous row, just shifted to the
right by 1 (wrapping around “cyclically” at the edges). That is, each row is a circular shift of the first row.
This is called a circulant matrix. A 4 × 4 circulant matrix looks like:
 
c0 c1 c2 c3
c3 c0 c1 c2 
C= c2 c3 c0 c1 


c1 c2 c3 c0
The general form of an n × n circulant matrix C is:
 
c0 c1 c2 ··· cn−1
cn−1
 c 0 c1 c2 ··· 

C=
cn−2 cn−1 c1 ··· 

 .. .. .. .. .. 
 . . . . . 
c1 c2 ··· cn−1 c0
When working with circulant matrix, it is convenient to number entries from 0 to n − 1 rather than from
1 to n!

1.1 Multiplying by circulant matrices: Convolutions

Suppose we have an n × n circulant matrix C that we want to multiply by a vector x = (x0 , x1 , . . . , xn ). It
turns out that the result is a very special kind of operation:
 
c0 c1 c2 ··· cn−1  
x0
cn−1
 c 0 c 1 c 2 · · · 
  x1 
y = Cx = cn−2 cn−1 c0
 ··· 
  .. 

 .. . .. . .. . .. . ..   . 
 .
xn−1
c1 c2 · · · cn−1 c0
Let’s write down a formula for the entries of y:

y0 = c0 x0 + c1 x1 + c2 x2 + · · · y1 = cn−1 x0 + c0 x1 + c1 x2 + · · · y3 = cn−2 x0 + cn−1 x1 + c0 x2 + · · ·

Can you see the pattern? This is one of those cases that is actually clearer if we write out the summation:
n
X
yk = cj−k xj
j=0

There is a slight problem with this formula: the subscript j −k can be < 0! No problem: we just interpret
the subscript periodically, i.e. we let c−1 = cn−1 , c−2 = cn−2 , and so on. Equivalently, we define cj±n = cj .
(We could say that the subscripts are modulo n.)
Multiplying by a circulant matrix is equivalent to a very famous operation called a circular convolution.
Convolution operations, and hence circulant matrices, show up in lots of applications: digital signal pro-
cessing, image compression, physics/engineering simulations, number theory and cryptography,
and so on.

2
2 Eigenvectors of circulant matrices
One amazing property of circulant matrices is that the eigenvectors are always the same. The eigen-
values are different for each C, but since we know the eigenvectors they are easy to diagonalize.
We can actually see one eigenvector right away. Let’s call it x(0) :
 
1
1
 
x(0) = 1
 
 .. 
.
1
This is an eigenvector because multiplying Cx(0) simply sums each row of C. But since each row of
C contains the same entries (just in a different order), the sum is the same:

Cx(0) = (c0 + c1 + · · · + cn−1 ) x(0)

| {z }
λ0

Thus, one of the eigenvalues λ0 of C is simply the sum of the row entries.
How can we find the other eigenvectors?

2.1 Roots of unity

The eigenvectors are simple to write down in terms of a very special value: a primitive root of unity:
2πi
ωn = e n

The quantity ωn has the very special property that ωnn = e2πi = 1 = ωn0 , but no smaller power equals 1.
Therefore, ωnj+n = ωnj ωnn = ωnj : the exponents of ω[U+2099] are periodic. (Just like the cj !)
For example, let’s plot the powers of ω7 :

In [3]: ω 7 = exp(2π*im / 7)

Out[3]: 0.6234898018587336 + 0.7818314824680298im

In [4]: fig = figure()

@manipulate for n in slider(1:20, value=7)
ω = exp(2π*im/n)
withfig(fig) do
for j = 1:n
z = ω ^ j
plot([0,real(z)], [0,imag(z)], ls="solid", color=(1,.7,.7))
plot(real(z), imag(z), "ro")
text(real(z), imag(z), "\$\\omega_{$n}^{$j}\$")
end
axis("square")
grid()
xlabel("real part")
ylabel("imaginary part")
title("$n-th roots of unity")
xlim(-1.2,1.2)
ylim(-1.2,1.2)
end
end

Interact.Slider{Int64}(Signal{Int64}(7, nactions=1),"",7,1:20,"horizontal",true,"d",true)

3
 Out[4]:

They are called “roots of unity” because ωnj for j = 0, . . . , n − 1 (or 1, . . . , n) are all the solutions z to

z n = 1.

2.2 Eigenvectors: The discrete Fourier transform (DFT)

In terms of ωn , the eigenvectors of a circulant matrix are easy: the k-th eigenvector x(k) (k = 0, . . . , n − 1)
for any n×n circulant matrix is simply

ωn0k
 
 ωn  1k
2k 
 
x =  ωn 
(k) 
 .. 
 . 
(n−1)k
ωn
Therefore, the matrix F whose columns are the eigenvectors is:


F = x(0) x(1) ··· x(n−1)
with entries

(k) 2πi
Fjk = xj = ωnjk = e n jk .

4
 Multiplying a vector by F is called a discrete Fourier transform (DFT). This is one of the most important
matrices in the world! (It is sort of a finite, computer-friendly analogue to a Fourier series if you’ve seen
those before.)
Before we show this, let’s try it:

In [5]: # define a function to create the n×n matrix F for any n:

F(n) = [exp((2π*im/n)*j*k) for j=0:n-1, k=0:n-1]

Out[5]: F (generic function with 1 method)

The 2 × 2 and 4 × 4 DFT matrices F are quite simple, for example

 
1 1
F2×2 =
1 −1
 
1 1 1 1
1 i −1 −i 
F4×4 = 
1 −1 1 −1
1 −i −1 i

In [6]: round.(F(4))

Out[6]: 4×4 Array{Complex{Float64},2}:

1.0+0.0im 1.0+0.0im 1.0+0.0im 1.0+0.0im
1.0+0.0im 0.0+1.0im -1.0+0.0im -0.0-1.0im
1.0+0.0im -1.0+0.0im 1.0-0.0im -1.0+0.0im
1.0+0.0im -0.0-1.0im -1.0+0.0im 0.0+1.0im

Let’s check that it diagonalizes our 7 × 7 mass-and-spring matrix A from earlier. We should have
F −1 AF = Λ:

In [7]: round.(inv(F(7)) * A * F(7), 3) # F−1 AF = Λ, rounded to 3 digits

Out[7]: 7×7 Array{Complex{Float64},2}:

0.0-0.0im -0.0-0.0im -0.0-0.0im ... 0.0-0.0im -0.0-0.0im
-0.0-0.0im 0.753-0.0im 0.0+0.0im 0.0-0.0im 0.0+0.0im
-0.0+0.0im 0.0-0.0im 2.445+0.0im -0.0+0.0im 0.0+0.0im
0.0+0.0im 0.0+0.0im 0.0-0.0im 0.0-0.0im 0.0+0.0im
0.0-0.0im 0.0+0.0im 0.0-0.0im -0.0+0.0im -0.0+0.0im
-0.0+0.0im 0.0-0.0im 0.0-0.0im ... 2.445-0.0im -0.0-0.0im
0.0+0.0im 0.0+0.0im -0.0+0.0im -0.0-0.0im 0.753+0.0im

Compare the diagonal entries to the eigenvalues:

In [8]: eigvals(A)

Out[8]: 7-element Array{Float64,1}:

-9.41142e-16
0.75302
0.75302
2.44504
2.44504
3.80194
3.80194

In [9]: diag(inv(F(7)) * A * F(7)) # diagonal entries

5
Out[9]: 7-element Array{Complex{Float64},1}:
0.0-1.2326e-32im
0.75302-5.32581e-17im
2.44504+2.87857e-16im
3.80194+1.90437e-16im
3.80194+2.12572e-16im
2.44504-1.93109e-16im
0.75302+4.73147e-17im

Yup!
Since A is real-symmetric, you may wonder why the eigenvectors are not real. But they could be chosen
real. For a real-symmetric circulant matrix, the real and imaginary parts of the eigenvectors are themselves
eigenvectors. This is why most of the eigenvalues come in pairs! (The only eigenvalues that don’t come in
pairs correspond to eigenvectors x(k) that are purely real, e.g. x(0) = (1, 1, . . . , 1).) These real and imaginary
eigenvectors turn out to correspond to a discrete cosine transform (DCT) and a discrete sine transform
(DST).

2.3 Derivation and eigenvalues

Why does this work? It’s easy to see if we take our formula from above for Cx and multiply it by an
eigenvector. Let y = Cx(k) . Then the `-th component is:
n−1
X n−1
X
y` = cj−` ωnjk = ωn`k cj−` ωn(j−`)k
j=0 j=0

But the remaining sum is now independent of `: because both cj and ωnj are periodic in j, all j → j − `
does is to re-arrange the numbers being summed (a circular shift), so you get the same sum. And ωn`k = x(k) ,
so we have:

Cx(k) = λk x(k)
where
n−1
X
λk = cj ωnjk
j=0

But if we define a vector ĉ = (λ0 , λ1 , . . . , λn−1 ), then

ĉ = F c
That is, the eigenvalues are the DFT of c (where c = first row of C).
Let’s check it:

In [10]: F(7) * A[:,1] # DFT of first row/column of A

Out[10]: 7-element Array{Complex{Float64},1}:

0.0+0.0im
0.75302+1.11022e-16im
2.44504-1.11022e-16im
3.80194-6.66134e-16im
3.80194-8.88178e-16im
2.44504-2.22045e-16im
0.75302+8.88178e-16im

In [11]: eigvals(A)

6
Out[11]: 7-element Array{Float64,1}:
-9.41142e-16
0.75302
0.75302
2.44504
2.44504
3.80194
3.80194

Yup, they match!

3 Unitarity
The DFT matrix F is special in many ways. It is symmetric, but not Hermitian, so its eigenvalues are not
real. However, it has orthogonal columns:

In [12]: round.(F(7)’ * F(7), 3)

Out[12]: 7×7 Array{Complex{Float64},2}:

7.0+0.0im -0.0-0.0im -0.0+0.0im ... 0.0+0.0im 0.0+0.0im 0.0+0.0im
-0.0+0.0im 7.0+0.0im -0.0-0.0im -0.0+0.0im 0.0+0.0im 0.0+0.0im
-0.0-0.0im -0.0+0.0im 7.0+0.0im -0.0+0.0im 0.0+0.0im 0.0+0.0im
0.0-0.0im -0.0-0.0im -0.0-0.0im 0.0+0.0im -0.0+0.0im -0.0+0.0im
0.0-0.0im -0.0-0.0im -0.0-0.0im 7.0+0.0im 0.0+0.0im -0.0+0.0im
0.0-0.0im 0.0-0.0im 0.0-0.0im ... 0.0-0.0im 7.0+0.0im -0.0-0.0im
0.0-0.0im 0.0-0.0im 0.0-0.0im -0.0-0.0im -0.0+0.0im 7.0+0.0im

(It is a straightforward exercise to show this, e.g. using the geometric-series summation
√ formula.)
The columns are orthogonal but not orthonormal because they have length n. But this means that if
we divide by their length, then:
1
√ F
n
is a unitary matrix. Equivalently:
1 H 1
F −1 = F = F̄ ,
n n
where we have used the fact that F T = F . This is the inverse discrete Fourier transform (IDFT).
Note that this means that every circulant matrix C has orthogonal eigenvectors (the columns of
F). (Even if the matrix C is not Hermitian or one of the similar cases we have seen so far!)

4 Fast Fourier transforms (FFTs)

The product F x for a vector x is the DFT of x. At first glance, it seems like it would require ∼ n2 operations
(n dot products) like any other matrix–vector multiplication.
One of the most amazing facts of computational science is that it is possible to compute the DFT F x in
only ∼ n log n operations (and ∼ n storage), by a fast Fourier transform (FFT) algorithm. FFT algorithms
mean that DFTs and circulant matrices become practical to deal with even for huge n, and are are the core
of a huge number of practical computational algorithms.
(From a linear-algebra viewpoint, it turns out that the “dense” matrix F factors into a product of
∼ log n sparse matrices.)
FFTs aren’t too complicated to understand (the simplest ones can be derived in a few lines of algebra),
but they are a bit outside the scope of 18.06. But fft(x) functions are built-in to Julia and every other
computational-science environment:

7
 Figure 2: FFTW logo

In [13]: fft(A[:,1]) # computes the same thing as F(7) * A[:,1], but much faster

Out[13]: 7-element Array{Complex{Float64},1}:

0.0+0.0im
0.75302+0.0im
2.44504+0.0im
3.80194+0.0im
3.80194+0.0im
2.44504+0.0im
0.75302+0.0im

But now that we know that F diagonalizes any circulant matrix, it leads to an amazing fact: you can
multiply Cx for any circulant matrix in ˜ n log n operations and ˜ n work.
Specifically, C = F ΛF −1 = F Λ
n F̄ , so to multiply Cx we just need to:

1. Multiply x̂ = F̄ x. This can be done by an (inverse) FFT in ∼ nlogn operations.

2. Multiply each component of x̂ by the eigenvalues λk /n. The eigenvalues can be computed by multiply
F by the first row of C, which can be done by an FFT in ∼ n log n operations.
3. Multiply by F via another FFT in ∼ n log n operations.

This means that circulant matrices are perfectly suited to iterative solver algorithms (e.g. the
power method or steepest-descent), just like sparse matrices!

4.1 A personal note

FFTs are near and dear to your instructor’s heart — 20 years ago, partially for my final project in Prof. Alan
Edelman’s class 18.337, I developed a library for FFTs along with a friend of mine. Being arrogant MIT
graduate students, we called it FFTW: The Fastest Fourier Transform in the West. Fortunately, FFTW
somewhat lived up to its name, and it is now the FFT library in Matlab, in Julia, and in many other software
packages.