SIAMREVIEW c 1999Societyfor Industrial andAppliedMathematics

Vol. 41, No. 1, pp. 135147

TheDiscreteCosineTransform

Gilbert Strang

Abstract. Each discrete cosine transform (DCT) uses N real basis vectors whose components are
cosines. In the DCT-4, for example, the jth component of v
k
is cos(j +
1
2
)(k+
1
2
)

N
. These
basis vectors are orthogonal and the transform is extremely useful in image processing. If
the vector x gives the intensities along a row of pixels, its cosine series

c
k
v
k
has the
coecients c
k
= (x, v
k
)/N. They are quickly computed from a Fast Fourier Transform.
But a direct proof of orthogonality, by calculating inner products, does not reveal how
natural these cosine vectors are.
We prove orthogonality in a dierent way. Each DCT basis contains the eigenvectors
of a symmetric second dierence matrix. By varying the boundary conditions we get the
established transforms DCT-1 through DCT-4. Other combinations lead to four additional
cosine transforms. The type of boundary condition (Dirichlet or Neumann, centered at
a meshpoint or a midpoint) determines the applications that are appropriate for each
transform. The centering also determines the period: N 1 or N in the established
transforms, N
1
2
or N +
1
2
in the other four. The key point is that all these eigenvectors
of cosines come from simple and familiar matrices.
Keywords. cosine transform, orthogonality, signal processing
AMSsubject classications. 42, 15
PII. S0036144598336745
Introduction. Just as the Fourier series is the starting point in transforming
and analyzing periodic functions, the basic step for vectors is the Discrete Fourier
Transform (DFT). It maps the time domain to the frequency domain. A vector
with N components is written as a combination of N special basis vectors v
k
. Those
are constructed from powers of the complex number w = e
2i/N
:
v
k
=
_
1, w
k
, w
2k
, . . . , w
(N1)k
_
, k = 0, 1, . . . , N 1 .
The vectors v
k
are the columns of the Fourier matrix F = F
N
. Those columns are
orthogonal. So the inverse of F is its conjugate transpose, divided by v
k

2
= N.
The discrete Fourier series x =

c
k
v
k
is x = Fc. The inverse c = F
1
x uses
c
k
= (x, v
k
)/N for the (complex) Fourier coecients.
Two points to mention, about orthogonality and speed, before we come to the
purpose of this note. First, for these DFT basis vectors, a direct proof of orthogonality
is very ecient:
(v
k
, v

) =
N1

j=0
(w
k
)
j
( w

)
j
=
(w
k
w

)
N
1
w
k
w

1
.

Received by the editors December 12, 1997; accepted for publication (in revised form) August
6, 1998; published electronically January 22, 1999.
http://www.siam.org/journals/sirev/41-1/33674.html

Massachusetts Institute of Technology, Department of Mathematics, Cambridge, MA 02139

(gs@math.mit.edu, http://www-math.mit.edu/gs).
135
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
136 GILBERT STRANG
The numerator is zero because w
N
= 1. The denominator is nonzero because k = .
This proof of (v
k
, v

) = 0 is short but not very revealing. I want to recommend a

dierent proof, which recognizes the v
k
as eigenvectors. We could work with any
circulant matrix, and we will choose below a symmetric A
0
. Then linear algebra
guarantees that its eigenvectors v
k
are orthogonal.
Actually this second proof, verifying that A
0
v
k
=
k
v
k
, brings out a central point
of Fourier analysis. The Fourier basis diagonalizes every periodic constant coecient
operator. Each frequency k (or 2k/N) has its own frequency response
k
. The
complex exponential vectors v
k
are important in applied mathematics because they
are eigenvectors!
The second key point is speed of calculation. The matrices F and F
1
are full,
which normally means N
2
multiplications for the transform and the inverse transform:
y = Fx and x = F
1
y. But the special form F
jk
= w
jk
of the Fourier matrix allows a
factorization into very sparse and simple matrices. This is the Fast Fourier Transform
(FFT). It is easiest when N is a power 2
L
. The operation count drops from N
2
to
1
2
NL, which is an enormous saving. But the matrix entries (powers of w) are complex.
The purpose of this note is to consider real transforms that involve cosines. Each
matrix of cosines yields a Discrete Cosine Transform (DCT). There are four established
types, DCT-1 through DCT-4, which dier in the boundary conditions at the ends
of the interval. (This dierence is crucial. The DCT-2 and DCT-4 are constantly
applied in image processing; they have an FFT implementation and they are truly
useful.) All four types of DCT are orthogonal transforms. The usual proof is a direct
calculation of inner products of the N basis vectors, using trigonometric identities.
We want to prove this orthogonality in the second (indirect) way. The basis
vectors of cosines are actually eigenvectors of symmetric second-dierence matrices.
This proof seems more attractive, and ultimately more useful. It also leads us, by
selecting dierent boundary conditions, to four less familiar cosine transforms. The
complete set of eight DCTs was found in 1985 by Wang and Hunt [10], and we want
to derive them in a simple way. We begin now with the DFT.
1. ThePeriodic CaseandtheDFT. The Fourier transform works perfectly for
periodic boundary conditions (and constant coecients). For a second dierence
matrix, the constant diagonals contain 1 and 2 and 1. The diagonals with 1
loop around to the upper right and lower left corners, by periodicity, to produce a
circulant matrix:
A
0
=
_

_
2 1 1
1 2 1
.
.
.
1 2 1
1 1 2
_

_
.
For this matrix A
0
, and every matrix throughout the paper, we look at three things:
1. the interior rows,
2. the boundary rows (rows 0 and N 1),
3. the eigenvectors.
The interior rows will be the same in every matrix! The jth entry of A
0
u is
u
j1
+ 2u
j
u
j+1
, which corresponds to u

. This choice of sign makes each

matrix positive denite (or at least semidenite). No eigenvalues are negative.
At the rst and last rows (j = 0 and j = N 1), this second dierence involves
u
1
and u
N
. It reaches beyond the boundary. Then the periodicity u
N
= u
0
and
u
N1
= u
1
produces the 1 entries that appear in the corners of A
0
.
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
THEDISCRETECOSINETRANSFORM 137
Note: The numbering throughout this paper goes from 0 to N 1, since SIAM
is glad to be on very friendly terms with the IEEE. But we still use i for

1! No
problem anyway, since the DCT is real.
We now verify that v
k
= (1, w
k
, w
2k
, . . . , w
(N1)k
) is an eigenvector of A
0
. It
is periodic because w
N
= 1. The jth component of A
0
v
k
=
k
v
k
is the second
dierence:
w
(j1)k
+ 2w
jk
w
(j+1)k
=
_
w
k
+ 2 w
k
_
w
jk
=
_
e
2ik/N
+ 2 e
2ik/N
_
w
jk
=
_
2 2 cos
2k
N
_
w
jk
.
A
0
is symmetric and those eigenvalues
k
= 2 2 cos
2k
N
are real. The smallest is

0
= 0, corresponding to the eigenvector v
0
= (1, 1, . . . , 1). In applications it is very
useful to have this at DC vector (direct current in circuit theory, constant gray level
in image processing) as one of the basis vectors.
Since A
0
is a real symmetric matrix, its orthogonal eigenvectors can also be chosen
real. In fact, the real and imaginary parts of the v
k
must be eigenvectors:
c
k
= Re v
k
=
_
1, cos
2k
N
, cos
4k
N
, . . . , cos
2(N 1)k
N
_
,
s
k
= Imv
k
=
_
0, sin
2k
N
, sin
4k
N
, . . . , sin
2(N 1)k
N
_
.
The equal pair of eigenvalues
k
=
Nk
gives the two eigenvectors c
k
and s
k
. The
exceptions are
0
= 0 with one eigenvector c
0
= (1, 1, . . . , 1), and for even N also

N/2
= 4 with c
N/2
= (1, 1, . . . , 1, 1). Those two eigenvectors have length

N,
while the other c
k
and s
k
have length
_
N/2. It is these exceptions that make the
real DFT (sines together with cosines) less attractive than the complex form. That
factor

2 is familiar from ordinary Fourier series. It will appear in the k = 0 term
for the DCT-1 and DCT-2, always with the at basis vector (1, 1, . . . , 1).
We expect the cosines alone, without sines, to be complete over a half-period.
In Fourier series this changes the interval from [, ] to [0, ]. Periodicity is gone
because cos 0 = cos . The dierential equation is still u

= u. The boundary con-

dition that produces cosines is u

(0) = 0. Then there are two possibilities, Neumann

and Dirichlet, at the other boundary:
Zero slope: u

() = 0 gives eigenfunctions u
k
(x) = cos kx;
Zero value: u() = 0 gives eigenfunctions u
k
(x) = cos
_
k +
1
2
_
x.
The two sets of cosines are orthogonal bases for L
2
[0, ]. The eigenvalues from u

k
=
u
k
are = k
2
and =
_
k +
1
2
_
2
.
All our attention now goes to the discrete case. The key point is that every
boundary condition has two fundamental approximations. At each boundary, the
condition on u can be imposed at a meshpoint or at a midpoint. So each problem
has four basic discrete approximations. (More than four, if we open up to further
renements in the boundary conditionsbut four are basic.) Often the best choices
use the same centering at the two endsboth meshpoint centered or both midpoint
centered.
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
138 GILBERT STRANG
In our problem, u

(0) = 0 at one end and u

() = 0 or u() = 0 at the other end

yield eight possibilities. Those eight combinations produce eight cosine transforms.
Starting from u(0) = 0 instead of u

(0) = 0, there are also eight sine transforms. Our

purpose is to organize this approach to the DCT (and DST) by describing the second
dierence matrices and identifying their eigenvectors.
Each of the eight (or sixteen) matrices has the tridiagonal form
A =
_

_

1 2 1
1 2 1

1 2 1

_

_
. (1)
The boundary conditions decide the eigenvectors, with four possibilities at each end:
Dirichlet or Neumann, centered at a meshpoint or a midpoint. The reader may object
that symmetry requires o-diagonal 1s in the rst and last rows. The meshpoint
Neumann condition produces 2. So we admit that the eigenvectors in that case need
a rescaling at the end (only involving

2) to be orthogonal. The result is a beautifully

simple set of basis vectors. We will describe their applications in signal processing.
2. TheDCT. The discrete problem is so natural, and almost inevitable, that it
is really astonishing that the DCT was not discovered until 1974 [1]. Perhaps this
time delay illustrates an underlying principle. Each continuous problem (dierential
equation) has many discrete approximations (dierence equations). The discrete case
has a new level of variety and complexity, often appearing in the boundary conditions.
In fact, the original paper by Ahmed, Natarajan, and Rao [1] derived the DCT-
2 basis as approximations to the eigenvectors of an important matrix, with entries

|jk|
. This is the covariance matrix for a useful class of signals. The number (near
1) measures the correlation between nearest neighbors. The true eigenvectors would
give an optimal KarhunenLo`eve basis for compressing those signals. The simpler
DCT vectors are close to optimal (and independent of ).
The four standard types of DCT are now studied directly from their basis vectors
(recall that j and k go from 0 to N 1). The jth component of the kth basis vector
is
DCT-1: cos jk

N1
(divide by

2 when j or k is 0 or N 1) ,
DCT-2: cos
_
j +
1
2
_
k

N
(divide by

2 when k = 0) ,
DCT-3: cos j
_
k +
1
2
_

N
(divide by

2 when j = 0) ,
DCT-4: cos
_
j +
1
2
_ _
k +
1
2
_

N
.
Those are the orthogonal columns of the four DCT matrices C
1
, C
2
, C
3
, C
4
. The
matrix C
3
with top row
1

2
(1, 1, . . . , 1) is the transpose of C
2
. All columns of C
2
, C
3
,
C
4
have length
_
N/2. The immediate goal is to prove orthogonality.
Proof. These four bases (including the rescaling by

2) are eigenvectors of sym-
metric second dierence matrices. Thus each basis is orthogonal. We start with ma-
trices A
1
, A
2
, A
3
, A
4
in the form (1), whose eigenvectors are pure (unscaled) cosines.
Then symmetrizing these matrices introduces the

2 scaling; the eigenvectors become

orthogonal. Three of the matrices were studied in an unpublished manuscript [12] by
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
THEDISCRETECOSINETRANSFORM 139
David Zachmann, who wrote down the explicit eigenvectors. His paper is very useful.
He noted earlier references for the eigenvalues; a complete history would be virtually
impossible.
We have seen that A
0
, the periodic matrix with 1, 2, 1 in every row, shares
the same cosine and sine eigenvectors as the second derivative. The cosines are picked
out by a zero-slope boundary condition in the rst row.
3. BoundaryConditionsat MeshpointsandMidpoints. There are two natural
choices for the discrete analogue of u

(0) = 0:
Symmetry around the meshpoint j = 0: u
1
= u
1
;
Symmetry around the midpoint j =
1
2
: u
1
= u
0
.
The rst is called whole-sample symmetry in signal processing; the second is half -
sample. Symmetry around 0 extends (u
0
, u
1
, . . .) evenly across the left boundary to
(. . . , u
1
, u
0
, u
1
, . . .) . Midpoint symmetry extends the signal to (. . . , u
1
, u
0
, u
0
, u
1
, . . .)
with u
0
repeated. Those are the simplest reections of a discrete vector. We substitute
the two options for u
1
in the second dierence u
1
+ 2u
0
u
1
that straddles the
boundary:
Symmetry at meshpoint: u
1
= u
1
yields 2u
0
2u
1
;
Symmetry at midpoint: u
1
= u
0
yields u
0
u
1
.
Those are the two possible top rows for the matrix A:
meshpoint: = 2 2 and midpoint: = 1 1 .
At the other boundary, there are the same choices in replacing u

() = 0. Substituting
u
N
= u
N2
or u
N
= u
N1
in the second dierence u
N2
+ 2u
N1
u
N
gives the
two forms for the Neumann condition in the last row of A:
meshpoint: = 2 2 and midpoint: = 1 1 .
The alternative at the right boundary is the Dirichlet condition u() = 0. The
meshpoint condition u
N
= 0 removes the last term of u
N2
+ 2u
N1
u
N
. The
midpoint condition u
N
+ u
N1
= 0 is simple too, but the resulting matrix will be a
little surprising. The 2 turns into 3:
meshpoint: = 1 2 and midpoint: = 1 3 .
Now we have 2 4 = 8 combinations. Four of them give the standard basis functions
of cosines, listed above. Those are the DCT-1 to DCT-4, and they come when the cen-
tering is the same at the two boundaries: both meshpoint centered or both midpoint
centered. Zachmann [12] makes the important observation that all those boundary
conditions give second-order accuracy around their center points. Finite dierences
are one-sided and less accurate only with respect to the wrong center! We can quickly
write down the matrices A
1
to A
4
that have these cosines as eigenvectors.
4. TheStandardCosineTransforms. Notice especially that the denominator in
the cosines (which is N 1 or N) agrees with the distance between centers. This
distance is an integer, measuring from meshpoint to meshpoint or from midpoint to
midpoint. We also give the diagonal matrix D that makes D
1
AD symmetric and
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
140 GILBERT STRANG
makes the eigenvectors orthogonal:
DCT-1
Centers j = 0 and N 1
Components cos jk

N1
D
1
= diag
_
2, 1, . . . , 1,

2
_
A
1
=
_

_
2 2
1 2 1

1 2 1
2 2
_

_
DCT-2
Centers j =
1
2
and N
1
2
Components cos
_
j +
1
2
_
k

N
D
2
= I
A
2
=
_

_
1 1
1 2 1

1 2 1
1 1
_

_
DCT-3
Centers j = 0 and N
Components cos j
_
k +
1
2
_

N
D
3
= diag(

2, 1, . . . , 1)
A
3
=
_

_
2 2
1 2 1

1 2 1
1 2
_

_
DCT-4
Centers j =
1
2
and N
1
2
Components cos
_
j +
1
2
_ _
k +
1
2
_

N
D
4
= I
A
4
=
_

_
1 1
1 2 1

1 2 1
1 3
_

_
Recently Sanchez et al. [7] provided parametric forms for all matrices that have
the DCT bases as their eigenvectors. These are generally full matrices of the form
Toeplitz plus near-Hankel. Particular tridiagonal matrices (not centered dier-
ences) were noticed by Kitajima, Rao, Hou, and Jain. We hope that the pattern of
second dierences with dierent centerings will bring all eight matrices into a common
structure. Perhaps each matrix deserves a quick comment.
DCT-1: The similarity transformation D
1
1
A
1
D
1
yields a symmetric matrix.
This multiplies the eigenvector matrix for A
1
by D
1
1
. (Notice that Ax = x leads to
(D
1
AD)D
1
x = D
1
x.) The eigenvectors become orthogonal for both odd N and
even N, when D
1
1
divides the rst and last components by

2:
N = 3
_
1

2
, 1,
1

2
_ _
1

2
, 0,
1

2
_ _
1

2
, 1,
1

2
_
for k = 0, 1, 2 ;
N = 4
_
1

2
, 1, 1,
1

2
_
. . .
_
1

2
, 1, 1,
1

2
_
for k = 0, 1, 2, 3 .
The rst and last eigenvectors have length

N 1; the others have length

_
(N 1)/2.
DCT-2: These basis vectors cos
_
j +
1
2
_
k

N
are the most popular of all, because
k = 0 gives the at vector (1, 1, . . . , 1). Their rst and last components are not
exceptional. The boundary condition u
1
= u
0
is a zero derivative centered on a
midpoint. Similarly, the right end has u
N
= u
N1
. When those outside values are
eliminated, the boundary rows of A
2
have the neat 1 and 1.
I believe that this DCT-2 (often just called the DCT) should be in applied math-
ematics courses along with the DFT. Figure 1 shows the eight basis vectors (when
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
THEDISCRETECOSINETRANSFORM 141
c
0
c
1
c
2
c
3
c
4
c
5
c
6
c
7
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-40
-35
-30
-25
-20
-15
-10
-5
0
Normalized Frequency
M
a
g
n
it
u
d
e

R
e
s
p
o
n
s
e

(
d
B
)
Fig. 1. The eight DCT-2 vectors and their Fourier transforms (absolute values).
N = 8). On the right are the Fourier transforms of those vectors. Maybe you can see
the rst curve |e
2ij/8
| and especially its second lobe, rising to 13 decibels (which is
20 log
10
13) below the top. This is not a big dropo! Like the closely connected Gibbs
phenomenon, it does not improve as N increases. A good lowpass lter can drop by
40 or 50 db. The other seven transforms vanish at zero frequency (no leakage of the
direct current DC term). Those seven vectors are orthogonal to (1, 1, . . . , 1).
This basis was chosen for the JPEG algorithm in image compression. Each 8 8
block in the image is transformed by a two-dimensional DCT. We comment below
on the undesirable blocking artifacts that appear when the transform coecients are
compressed.
DCT-3: The vectors cos j
_
k +
1
2
_

N
are the discrete analogues of cos(k +
1
2
)x.
The Neumann condition at the left and Dirichlet condition at the right are centered
at meshpoints. For orthogonality we need the factor D
1
3
that divides the rst com-
ponents by

2. This basis loses to the DCT-4.
DCT-4: We had never seen the nal entry 3 in the matrix A
4
but MATLAB
insisted it was right. Now we realize that a zero boundary condition at a midpoint
gives u
N
u
N1
(the extension is antisymmetric). Then 1, 2, 1 becomes 1, 3.
The eigenvectors are even at the left end and odd at the right end. This attractive
property leads to j +
1
2
and k +
1
2
and a symmetric eigenvector matrix C
4
. Its
applications to lapped transforms are described below.
Remember our proof of orthogonality! It is a verication that the cosine vectors
are eigenvectors of A
1
, A
2
, A
3
, A
4
. For all the 1, 2, 1 rows, this needs to be done
only once (and it reveals the eigenvalues = 2 2 cos ). There is an irreducible
minimum of trigonometry when the jth component of the kth vector c
k
is cos j in
types 1 and 3, and cos(j +
1
2
) in types 2 and 4:
cos(j 1) + 2 cos j cos(j + 1) = (2 2 cos ) cos j ,
cos
_
j
1
2
_
+ 2 cos
_
j +
1
2
_
cos
_
j +
3
2
_
= (2 2 cos ) cos
_
j +
1
2
_
.
This is Ac
k
=
k
c
k
on all interior rows. The angle is = k

N1
for type 1 and = k

N
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
142 GILBERT STRANG
for type 2. It is =
_
k +
1
2
_

N
for A
3
and A
4
. This leaves only the rst and last
components of Ac
k
=
k
c
k
to be veried in each case.
Let us do only the fourth case, for the last row 1, 3 of the symmetric matrix
A
4
. A last row of 1, 1 would subtract the j = N 2 component from the j = N 1
component. Trigonometry gives those components as
j = N 1 :cos
_
N
1
2
__
k +
1
2
_

N
= sin
1
2
_
k +
1
2
_

N
,
j = N 2 :cos
_
N
3
2
__
k +
1
2
_

N
= sin
3
2
_
k +
1
2
_

N
.
We subtract using sina sinb = 2 cos
_
b+a
2
_
sin
_
ba
2
_
. The dierence is
2 cos
_
k +
1
2
_

N
sin
1
2
_
k +
1
2
_

N
. (2)
The last row of A
4
actually ends with 3, so we still have 2 times the last component
(j = N 1) to include with (2):
_
2 2 cos
_
k +
1
2
_

N
_
sin
1
2
_
k +
1
2
_

N
. (3)
This is just
k
times the last component of c
k
. The nal row of A
4
c
k
=
k
c
k
is
veried.
There are also discrete sine transforms DST-1 through DST-4. The entries of the
basis vectors s
k
are sines instead of cosines. These s
k
are orthogonal because they
are eigenvectors of symmetric second dierence matrices, with a Dirichlet (instead of
Neumann) condition at the left boundary. In writing about the applications to signal
processing [9], we presented a third proof of orthogonalitywhich simultaneously
covers the DCT and the DST, and shows their fast connection to the DFT matrix of
order 2N. This is achieved by a neat matrix factorization given by Wickerhauser [11]:
e
i/4N
R
T
F
2N
R =
_
C
4
0
0 iS
4
_
.
The entries of S
4
are sin(j +
1
2
)(k +
1
2
)

N
. The connection matrix R is very sparse,
with w = e
i/2N
:
R =
1

2
_
D D
E E
_
with
D = diag(1, w, . . . , w
N1
) ,
E = antidiag(w, w
2
, . . . , w
N
) .
Since R
T
and F
2N
and R have orthogonal columns, so do C
4
and S
4
.
5. CosineTransformswithN
1
2
andN+
1
2
. There are four more combinations
of the discrete boundary conditions. Every combination that produces a symmetric
matrix will also produce (from the eigenvectors of that matrix) an orthogonal trans-
form. But you will see N
1
2
and N +
1
2
in the denominators of the cosines, because
the distance between centers is no longer an integer. One center is a midpoint and
the other is a meshpoint.
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
THEDISCRETECOSINETRANSFORM 143
The transforms DCT-5 to DCT-8, when they are spoken of at all, are called odd.
They are denoted by DCT-IO to DCT-IV O in [5] and [7]. Three of the tridiagonal
matrices (A
5
, A
6
, A
8
) are quite familiar:
DCT-5
Centers j = 0 and N
1
2
Components cos jk

N
1
2
D
5
= diag(

2, 1, . . . , 1)
A
5
=
_

_
2 2
1 2 1

1 2 1
1 1
_

_
DCT-6
Centers j =
1
2
and N 1
Components cos
_
j +
1
2
_
k

N
1
2
D
6
= diag(1, . . . , 1,

2)
A
6
=
_

_
1 1
1 2 1

1 2 1
2 2
_

_
DCT-7
Centers j = 0 and N
1
2
Components cos j
_
k +
1
2
_

N
1
2
D
7
= diag(

2, 1, . . . , 1)
A
7
=
_

_
2 2
1 2 1

1 2 1
1 3
_

_
DCT-8
Centers j =
1
2
and N
Components cos
_
j +
1
2
_ _
k +
1
2
_

N+
1
2
D
8
= I
A
8
=
_

_
1 1
1 2 1

1 2 1
1 2
_

_
.
We could study A
8
by reection across the left boundary, to produce the pure Toeplitz
1, 2, 1 matrix (which is my favorite example in teaching). The eigenvectors become
discrete sines on a double intervalalmost. The length of the double interval is not
2N, because the matrix from reection has odd order. This leads to the new period
length N +
1
2
in the cosines.
Notice that A
5
has the boundary conditions (and eigenvector components) in
reverse order from A
6
. The rst eigenvectors of A
5
and A
6
are (1, 1, . . . , 1), corre-
sponding to k = 0 and = 0. This at vector can represent a solid color or a xed
intensity by itself (this is terric compression). The DCT-5 and DCT-6 have a coding
gain that is completely comparable to the DCT-2.
So we think through the factors that come from D
6
= diag(1, . . . , 1,

2). The
symmetrized D
1
6
A
6
D
6
has

2 in the two lower right entries, where A

6
has 1 and
2. The last components of the eigenvectors are divided by

2; they are orthogonal
but less beautiful. We implement the DCT-6 by keeping the matrix C
6
with pure
cosine entries, and accounting for the correction factors by diagonal matrices:
4
2N1
C
6
diag
_
1
2
, 1, . . . , 1
_
C
T
6
diag
_
1, . . . , 1,
1
2
_
= I. (4)
The cosine vectors have squared length
2N1
4
, except the all-ones vector that is ad-
justed by the rst diagonal matrix. The last diagonal matrix corrects the Nth com-
ponents as D
6
requires. The inverse of C
6
is not quite C
T
6
(analysis is not quite
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
144 GILBERT STRANG
the transpose of synthesis, as in an orthogonal transform) but the corrections have
trivial cost. For N = 2 and k = 1, the matrix identity (4) involves cos
1
2

3/2
=
1
2
and
cos
3
2

3/2
= 1:
4
3
_
1
1
2
1 1
_ _
1
2
1
_ _
1 1
1
2
1
_ _
1
1
2
_
=
_
1
1
_
.
Malvar has added a further good suggestion: Orthogonalize the last N 1 basis
vectors against the all-ones vector. Otherwise the DC component (which is usually
largest) leaks into the other components. Thus we subtract from each c
6
k
(with k > 0)
its projection onto the at c
6
0
:
c
6
k
= c
6
k

(1)
k
2N
(1, 1, . . . , 1) . (5)
The adjusted basis vectors are now the columns of

C
6
, and (5) becomes
C
6
=

C
6
_

_
1
1
2N
+1
2N
. . .
1
.
.
.
1
_

_
.
This replacement in equation (4) also has trivial cost, and that identity becomes

C
6

C
1
6
= I. The coecients in the cosine series for x are y =

C
1
6
x. Then x is
reconstructed from

C
6
y (possibly after compressing y). You see how we search for a
good basis. . . .
Transforms 5 to 8 are not used in signal processing. The half-integer periods are
a disadvantage, but reection oers a possible way out. The reected vectors have an
integer double period and they overlap.
6. Convolution. The most important algebraic identity in signal processing is the
convolution rule. A slightly awkward operation in the time domain (convolution, from
a Toeplitz matrix or a circulant matrix) becomes beautifully simple in the frequency
domain (just multiplication). This accounts for the absence of matrices in the leading
textbooks on signal processing. The property of time invariance (delay of input simply
delays the output) is always the starting point.
We can quickly describe the rules for doubly innite convolution and cyclic con-
volution. A vector h of lter coecients is convolved with a vector x of inputs. The
output is y = h x with no boundary and y = h
c
x in the cyclic (periodic) case:
y
n
=

h
k
x
nk
or y
n
=

k+n(mod N)
h
k
x

. (6)
Those are matrix-vector multiplications y = Hx. On the whole line (n Z) the
doubly innite matrix H is Toeplitz; the number h
k
goes down its kth diagonal. In
the periodic case (n Z
N
) the matrix is a circulant; the kth diagonal continues with
the same h
k
onto the (k N)th diagonal. The eigenvectors of these matrices are pure
complex exponentials. So when we switch to the frequency domain, the matrices are
diagonalized. The eigenvectors are the columns of a Fourier matrix, and F
1
HF is
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
THEDISCRETECOSINETRANSFORM 145
diagonal. Convolution with h becomes multiplication by the eigenvalues H() in the
diagonal matrix:
(7)
_

h
k
e
ik
_ _

e
i
_
=

y
n
e
in
is H()X() = Y () ,
(7)
N
_

N1
0
h
k
w
k
__

N1
0
x

_
=

N1
0
y
n
w
n
is H(w)X(w) = Y (w) .
The innite case (discrete time Fourier transform) allows all frequencies || . The
cyclic case (DFT) allows the N roots of w
N
= 1. The multiplications in (7) agree
with the convolutions in (6) because e
ikx
e
ix
= e
i(k+)x
and w
k
w

= w
k+
. The
question is: What convolution rule goes with the DCT?
A complete answer was found by Martucci [5]. The nite vectors h and x are
symmetrically extended to length 2N or 2N1, by reection. Those are convolved in
the ordinary cyclic way (so the double length DFT appears). Then the output is re-
stricted to the original N components. This symmetric convolution h
s
x corresponds
in the transform domain to multiplication of the cosine series.
The awkward point, as the reader already knows, is that a symmetric reection
can match u
1
with u
0
or u
1
. The centering can be whole sample or half sample at
each boundary. The extension of h can be dierent from the extension of x! This
conrms again that discrete problems have an extra degree of complexity beyond
continuous problems. (And we resist the temptation to compare combinatorics and
linear algebra with calculus.)
In the continuous case, we are multiplying two cosine expansions. This corre-
sponds to symmetric convolution of the coecients in the expansions.
7. The DCT in Image Processing. Images are not innite, and they are not
periodic. The image has boundaries, and the left boundary seldom has anything
to do with the right boundary. A periodic extension can be expected to have a
discontinuity. That means a slow decay of Fourier coecients and a Gibbs oscillation
at the jumpthe one place where Fourier has serious trouble! In the image domain
this oscillation is seen as ringing. The natural way to avoid this discontinuity is
to reect the image across the boundary. With cosine transforms, a double-length
periodic extension becomes continuous.
A two-dimensional (2D) image may have (512)
2
pixels. The gray level of the
pixel at position (i, j) is given by an integer x(i, j) (between 0 and 255, thus 8 bits
per pixel). That long vector x can be ltered by x h, rst a row at a time (j xed)
and then by columns (using the one-dimensional (1D) transforms of the rows). This
is computationally and algebraically simplest: the 2D Toeplitz and circulant matrices
are formed from 1D blocks.
Similarly the DCT-2 is applied to rows and then to columns; 2D is the tensor
product of 1D with 1D. The JPEG compression algorithm (established by the Joint
Photographic Experts Group) divides the image into 8 8 blocks of pixels. Each
block produces 64 DCT-2 coecients. Those 64-component vectors from the separate
blocks are compressed by the quantization step that puts coecients into a discrete
set of bins. Only the bin numbers are transmitted. The receiver approximates the
true cosine coecient by the value at the middle of the bin (most numbers go into the
zero bin). Figures 2ad show the images that the receiver reconstructs at increasing
compression ratios and decreasing bit rates:
1. the original image (1:1 compression, all 8 bits per pixel);
2. medium compression (8:1, average 1 bit per pixel);
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
146 GILBERT STRANG
(a) (b)
(c) (d)
Fig. 2 (a) Original Barbara gure. (b) Compressed at 8:1. (c) Compressed at 32:1. (d) Compressed
at 128:1.
3. high compression (32:1, average
1
4
bit per pixel);
4. very high compression (128:1, average
1
16
bit per pixel).
You see severe blocking of the image as the compression rate increases. In telecon-
ferencing at a very low bit rate, you can scarcely recognize your friends. This JPEG
standard for image processing is quick but certainly not great. The newer standards
allow for other transforms, with overlapping between blocks. The improvement is
greatest for high compression. The choice of basis (see [8]) is crucial in applied mathe-
matics. Sometimes form is substance!
One personal comment on quantization: This more subtle and statistical form of
roundo should have applications elsewhere in numerical analysis. Numbers are not
simply rounded to fewer bits, regardless of size. Nor do we sort by size and keep only
the largest (this is thresholding, when we want to lose part of the signalit is the basic
idea in denoising). The bit rate is controlled by the choice of bin sizes, and quantiza-
tion is surprisingly cheap. Vector quantization, which puts vectors into multidimen-
sional bins, is more expensive but in principle more ecient. This technology of coding
is highly developed [3] and it must have more applications waiting to be discovered.
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p
THEDISCRETECOSINETRANSFORM 147
A major improvement for compression and image coding was Malvars [4] ex-
tension of the ordinary DCT to a lapped transform. Instead of dividing the image
into completely separate blocks for compression, his basis vectors overlap two or more
blocks. The overlapping has been easiest to develop for the DCT-4, using its evenodd
boundary conditionswhich the DCT-7 and DCT-8 share. Those conditions help to
maintain orthogonality between the tail of one vector and the head of another. The
basic construction starts with a symmetric lowpass lter of length 2N. Its coecients
p(0), . . . p(2N 1) are modulated (shifted in frequency) by the DCT-4:
The kth basis vector has jth component p(j) cos
_
(k +
1
2
)(j +
N+1
2
)

N

.
There are N basis vectors of length 2N, overlapping each block with the next block.
The 1D transform matrix becomes block bidiagonal instead of block diagonal. It is still
an orthogonal matrix [4, 9] provided p
2
(j)+p
2
(j +N) = 1 for each j. This is Malvars
modulated lapped transform (MLT), which is heavily used by the Sony mini disc and
Dolby AC-3. (It is included in the MPEG-4 standard for video.) We naturally wonder
if this MLT basis is also the set of eigenvectors for an interesting symmetric matrix.
Coifman and Meyer found the analogous construction [2] for continuous wavelets.
The success of any transform in image coding depends on a combination of
propertiesmathematical, computational, and visual. The relation to the human
visual system is decided above all by experience. This article was devoted to the
mathematical property of orthogonality (which helps the computations). There is no
absolute restriction to second dierence matrices, or to these very simple boundary
conditions. We hope that the eigenvector approach will suggest more new transforms,
and that one of them will be fast and visually attractive.
WebLinks.
JPEG http://www.jpeg.org/public/jpeglinks.htm
DCT http://www.cis.ohio-state.edu/hypertext/faq/usenet/
compression-faq/top.html (includes source code)
Author http://www-math.mit.edu/gs/
REFERENCES
[1] N. Ahmed, T. Natarajan, and K. R. Rao, Discrete cosine transform, IEEE Trans. Comput.,
C-23 (1974), pp. 9093.
[2] R. Coifman and Y. Meyer, Remarques sur lanalyse de Fourier ` a fenetre, C. R. Acad. Sci.
Paris, 312 (1991), pp. 259261.
[3] N. J. Jayant and P. Noll, Digital Coding of Waveforms, Prentice-Hall, Englewood Clis,
NJ, 1984.
[4] H. S. Malvar, Signal Processing with Lapped Transforms, Artech House, Norwood, MA, 1992.
[5] S. Martucci, Symmetric convolution and the discrete sine and cosine transforms, IEEE Trans.
Signal Processing, 42 (1994), pp. 10381051.
[6] K. R. Rao and P. Yip, Discrete Cosine Transforms, Academic Press, New York, 1990.
[7] V. Sanchez, P. Garcia, A. Peinado, J. Segura, and A. Rubio, Diagonalizing properties of
the discrete cosine transforms, IEEE Trans. Signal Processing, 43 (1995), pp. 26312641.
[8] G. Strang, The search for a good basis, in Numerical Analysis 1997, D. Griths, D. Higham,
and A. Watson, eds., Pitman Res. Notes Math. Ser., Addison Wesley Longman, Harlow,
UK, 1997.
[9] G. Strang and T. Nguyen, Wavelets and Filter Banks, Wellesley-Cambridge Press, Wellesley,
MA, 1996.
[10] Z. Wang and B. Hunt, The discrete W-transform, Appl. Math. Comput., 16 (1985), pp. 1948.
[11] M. V. Wickerhauser, Adapted Wavelet Analysis from Theory to Software, AK Peters, Natick,
MA, 1994.
[12] D. Zachmann, Eigenvalues and Eigenvectors of Finite Dierence Matrices, unpublished
manuscript, 1987, http://epubs.siam.org/sirev/zachmann/. .
D
o
w
n
l
o
a
d
e
d

0
2
/
2
1
/
1
4

t
o

1
8
.
1
1
1
.
9
2
.
1
5
0
.

R
e
d
i
s
t
r
i
b
u
t
i
o
n

s
u
b
j
e
c
t

t
o

S
I
A
M

l
i
c
e
n
s
e

o
r

c
o
p
y
r
i
g
h
t
;

s
e
e

h
t
t
p
:
/
/
w
w
w
.
s
i
a
m
.
o
r
g
/
j
o
u
r
n
a
l
s
/
o
j
s
a
.
p
h
p