Least Squares Quantization in PCM

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

IEEE TRANSACTIONSON INFORMATIONTHEORY,VOL. IT-28, NO.

2, MARCH 1982

129

Least Squares Quantization in PCM


STUART P. LLOYD

received amplitudes s!i), as in (1), to produce an exact


reproduction of the ongmal band-limited signal s.
PCM is a modification of this. Instead of sending the
exact sample values (3), one partitions the voltage range of
the signal into a finite number of subsetsand transmits to
the receiver only the information as to which subset a
sample happens to fall in. Built into the receiver there is a
source of fixed representative voltages-quanta-one
for each of the subsets.When the receiver is informed that
a certain sample fell in a certain subset,it usesits quantum
for that subset as an approximation to the true sample
value and constructs a band-limited signal based on these
approximate sample values.
We define the noise signal as the difference between the
receiver-output
signal and the original signal and the noise
I. INTRODUCTION
power as the averagesquare of the noise signal. The probHE BASIC IDEAS in the pulse-code modulation lem we consider is the following: given the number of
(PCM) system [ 11,[2, ch. 191are the Shannon-Nyquist quanta and certain statistical properties of the signal, desampling theorem and the notion of quantizing the sample termine the subsetsand quanta that are best in minimizing
values.
the noise power.
The sampling theorem asserts that a signal voltage s(t),
- 00 < t < cc, containing only frequencies less than W
II.
QUANTIZATION
cycles/s can be recovered from a sequenceof its sample
Let us formulate the quantization process more exvalues according to
plicitly. A quantization scheme consists of a class of sets
s(t) = f$ s(tj)K(t - t,),
-ccoOttm,
(1) {Q,, Qz>.. ->Q,} and a set of quanta {q,, q2;. -,q,}. The
{Q,} are any v disjoint subsets of the voltage axis which,
jz-00
taken together, cover the entire voltage axis. The {qa} are
where s(tj) is the value of s at thejth sampling instant
any v finite voltage values. The number v of quanta is to be
regarded
throughout as a fixed finite preassignednumber.
-coCj-Co3,
t,=&k
We associatewith a partition (Q,} a label function y(x),
- 00 < x < co, defined for all (real) voltages x by
and where
y(x) = 1 if x liesin Q,,
sin2rWt
-KloOttco,
K(t) =
2mJ,J,7t T
(2)
y(x) =2 if x liesin Q2,
(4)
is a sin t/t pulse of the appropriate width.
The pulse-amplitude modulation (PAM) system [2, ch.
161 is based on the sampling theorem alone. One sends
y(x) = v if x liesin Q,.
over the system channel, instead of the signal values s(t)
If s(t,) is the jth sample of the signal s, as in Section I,
for all times t, only a sequence
then we denote by aj the label of the set that this sample
. . . >s(t-I>, &), s(t*>, * * (3) falls in:
of samplesof the signal. The (idealized) receiver constructs
-m<j<co.
aj = Y(s(tj)),
the pulses K(t - tj) and adds them together with the
In PCM the signal sent over the channel is (in some code
or another) the sequenceof labels
Manuscript received May 1, 1981. The material in this paper was
. . . ,a-,,a,,q;~~,
(5)
Abstract-It
has long been realized that in pulse-code modulation
(PCM), with a given ensemble of signals to handle, the quantum values
should be spaced more closely in the voltage regions where the signal
amplitude is more likely to fall. It has been shown by Panter and Dite that,
in the limit as the number of quanta becomes infinite, the asymptotic
fractional density of quanta per unit voltage should vary as the one-third
power of the probability density per unit voltage of signal amplitudes. In
this paper the corresponding result for any finite number of quanta is
derived; that is, necessary conditions are found that the quanta and
associatedquantization intervals of an optimum finite quantization scheme
must satisfy. The optimization criterion used is that the average quantization noise power be a minimum. It is shown that the result obtained here
goes over into the Panter and Dite result as the number of quanta become
large. The optimum quantization schemesfor 26 quanta, b = 1,2, t ,7, are
given numerically for Gaussian and for Laplacian distribution of signal
amplitudes.

presented in part at the Institute of Mathematical Statistics Meeting,


Atlantic City, NJ, September lo- 13, 1957.
The author is with Bell Laboratories, Whippany Road, Whippky, NJ
0798 1.

each aj being one of the integers { 1,2,. . +,v}. The technology of this transmission does not concern us, except that

001%9448/82/0300-0129$00.75 01982 IEEE

IEEE TRANSACTIONSONINFORMATIONTHEORY,VOL. IT-28,~o.

130

we assume that such a sequencecan be delivered to the


receiver without error.
The receiver uses the fixed voltage 4, as an approximation to all sample voltages in Q,, (Y= 1,2,. * *, v. That is,
the receiver, being given the value of uj in the sequence(5),
proceeds as if the jth sample of s had value q, and
produces the receiver-output signal
r(t) =

q,K(r

-co<ttoQ.

tj)P

j=-*

To put it another way, the system mutilates an actual


sample voltage value x to the quantized value y(x) given
by
(6)
Y(X) = 4U(X) -cQcx)xxcc,
and we may expressthe receiver output in terms of this as
r(t)

y(s(tj))K(t

ple,
F(x)

= P{s( t) 5 X})

-@3(x<m,

is independent of t, - 00 < t -C 00, as indicated by the


notation. Then the averagepower of the s process,assumed
to be finite, is constant in time:
S = E{s(t)}

=/:mxzdF(x),

-co ==ct < co.

w-9
Moreover, the r and n processeshave this same property;
the averagereceiver-output power R is given by
R = E{r2(t)}

=I-Ly(x)

dF(x),

-00 < t < 00,


(11)

where y(x) is defined in (6), and the noise power N is


N = E{n2(t)}

-03~<~<.

tj)9

~,MARCH 1982

=J-tz2(x)dF(x),

-00 <t < co,

jz-03

(1-a

(7)

Hence the noise signal, defined as


=

n(t)

r(t)

s(t),

-m<ttcc,

is given by

n(t)= 5 z(s(t,))K(t
- t,), -co~tt<,

with z(x) as in (9). (Detailed proofs of these statements,


together with further assumptions used, are given in Appendix A.) The stochastic processproblem is thus reduced
to a problem in a single real variable: choose the {Q,} and
{q,} so that the rightmost integral in (12) is as small as
possible.

jz-00

IV.

(8)

where
-Mexico,
z(x) =y(x> - x,
(9)
may be regarded as the quantization error added to a
sample which has voltage value x.
Note that we assumethat the receiver uses the nonrealizable pulses (2). If other pulses are used (e.g:, step functions
or other realizable pulses) there will be sampling noise, in
general, even without quantization [3]. Our noise (8) is due
strictly to quantization.
Finally we must emphasizethat we assumethat the {Q,}
and {q,) are constant in time. In deltamodulation and its
refinements the {Q,} and {q,} change from sampling
instant to sampling instant, depending on the past behavior of the signal being handled. Such systems are very
difficult to treat theoretically.
III.

THE BEST QUANTA

We consider first the problem of minimizing N with


respect to the quanta {qa} when the {Q,} are fixed preassigned sets.
The dF integral in (12) may be written more explicitly as
N=

( (q,-x)2dF(x).

(13)

(The sets {Q,} must be measurable[dF] if (1 l)-( 13) are to


have meaning, and we assumealways that this is the case.)
If we regard the given F as describing the distribution of
unit probability mass on the voltage axis [5, p. 571,then
(13) expressesN as the total moment of inertia of the
sets {Q,} around the respectivepoints (4,). It is a classical
result that such a moment assumes its minimum value
when each {q,} is the center of mass of the corresponding
{Q,} (see,e.g., [5, p. 1751).That is,

NOISE POWER

Instead of working with a particular band-limited signal,


we assumethat there is given a probabilistic family of such
signals. That is, the s of the preceding sections and hence
the various signals derived from it are to be regarded as
stochastic processes[4]. We denote the underlying probability measure by P{ .} and averageswith respect to this
measure(expectations) by E{ . }.
We use the following results of the probabilistic treatment. We assume that the s process is stationary, so that
the cumulative probability distribution function of a sam-

4a =

JQ,
/ QndF(x)

a = 1,2;**,v >

(14

are the uniquely determined best quanta to use with a


given partition {Q,}.
To avoid the continual mention of trivial cases we assume always that F is increasing at least by v + 1 points,
so that the quantization noise does not vanish. Then none
of the denominators in (14) will vanish, at least in an

131

LLOYD: LEASTSQUARESQUANTIZATION IN PCM

optimum scheme. For if Q, has vanishing mass it can be


combined with some set Q, of nonvanishing mass (discarding q,) to give a scheme with v - 1 quanta and the
same noise. Then one of the sets of this scheme can be
divided into two sets and new quanta assigned to give a
scheme with v quanta and noise less than in the original
scheme.(We omit the details.)
If the expressionon the right in (14) is substituted for q,
in (13), there results
N = S -

partition to use with the given quanta is


Q, = {x: - 00 <XIX,)
Q,=

(16)

Qv-,= {x:x,-2(xIx,-,}
Q,= {x:x,~~~x~co},
where the endpoints {x~} are given

ji q:lQ dF(x),
a
LY=l

where the {q,} here are the optimum ones of (14). The sum
on the right is the receiver-output power from (11). Hence
when the {q,} are centers of mass of the {Q,}, optimum or
not, then S = R + N, which implies that the noise is
orthogonal to the receiver output. One expects this in a
least squaresapproximation, of course.
V. THE BEST PARTITION

{X:X<XIX~}

Xl = &I,

q2)

x2 = 5(42

q3)

07)

X,-I = tkv-1 + 4,).


We have assumed, as we shall hereafter, that the indexing
is such that q, -C q2 < . . * -C q,.
VI.

QUANTIZATION

PROCEDURES

Now we find the best sets {Q,} to use with a fixed


From Sections IV and V we know that we may confine
preassignedset of quanta (4,). The considerations of this our attention to quantization schemesdefined by 2v - 1
section are independent of those of the preceding section. numbers
In particular, the best {Q,} for given {q,} may not have
the {q,} as their centers of mass.
ql<x,
<q,<x,<
-..<qv-,<xp-,(qy,
(18)
We assume that the given {q,} are distinct since it will
never happen in an optimum schemethat q, = qP for some where the {x~} are the endpoints of the intervals {Q,}, as
(Y# p. For if q, = qp, then Q, and QB are effectively one in (16), and the { qa} are the corresponding quanta. We will
set Q, U Qp as far as the noise is concerned (13), and this regard such a set of numbers as the Cartesian coordinates
set can be redivided into two sets and these two sets can be of a point
given distinct quantum values in such a way as to reduce
the noise. (We omit the details.)
Consider the probability mass in a small interval around in (2v - 1)-dimensional Euclidean space E2y-I. The noise
voltage value x. According to (13) any of this mass which is as a function of p has the form
assignedto q, (i.e., which lies in Q,) will contribute to the
+jx2(q2 - x)dF(x)
+ .a.
noise at rate (q, - x)~ per unit mass. To minimize the N(p) =Jx-cc (ql - x)dF(x)
noise, then, any mass in the neighborhood of x should be
assignedto a qa for which (qa - x)~ is the smallest of the
+,-;,
- x)dF(x).
(19)
h-1
numbers (q, - x)~, (q2 - ~)~;..,(q~
- x)~. In other
words,
In an optimum schemethe {qa} will be centers of mass
of
the corresponding {Q,}, (14), and the ( xa} will lie
Q, > [X : (q, - x) < ( qp - x) for all /3 # a>,
midway between adjacent {q,}, (17). From the derivations
a = l,,..,V,
these conditions are sufficient that N(p) be a minimum
modulo sets of measurezero [ dF]. This simplifies to
with respect to variations in each coordinate separatelyand
hence are necessaryconditions at a minimum of N(p). As
Q, 1 {x:(qp - q&x - f(q, + qp)) <OforallP+a),
it turns out, however, they are not sufficient conditions for
a = 1,2;**,v.
05) a minimum of N(p). Points at which (14) and (17) are
satisfied, which we term stationary points, while never local
It is straightforward that the best {Q,} are determined by maxima, may be saddle points of N(p). Moreover, among
(15) as the intervals whose endpoints bisect the segments the stationary points there may be several local minima,
between successive{q,}, except that the assignment of the only one of which is the sought absolute minimum of N(p).
endpoints is not determined. To make matters definite we These complications are discussedfurther in Appendix B.
let the {Q,} be left-open and right-closed, so that the best The author has not been able to determine sufficient
conditions for an absolute minimum.
The derivations suggest one trial-and-error method for
If C(x) is a condition on x, then {x: C(x)} denotes the set of all x
finding
stationary points. A trial point p(l) in E2-, is
which satisfy C(x).

IEEE TRANSACTIONSON INFORMATIONTHEORY,VOL. IT-28, NO. 2, MARCH 1982

132

chosen as follows. The endpoints

will lie equidistant from x, (17), and from (19) the noise
will not change if we reassign the mass at x, to Q,, 1,
-00 <xp<xp<
. . . <xp, < CQ
retaining the given (4,) as quanta. But q, and qa+ , are
are chosen arbitrarily except that each of the resulting definitely not centers of mass of the corresponding mod{Qt)} should have nonvanishing mass. Then the centers of ified sets,and the noise will strictly decreaseas qa and qa+ 1
mass of these sets are taken as the first trial quanta {q:)}.
are moved to the new centers of mass. Thus the given
These values will not satisfy the midpoint conditions configuration is not a local minimum, contrary to assump(17), in general, so that the second trial point pc2) is taken tion. From this result and (19) we seethat N(p) is continuto be
ous in a neighborhood of a local minimum. We have
proved
also that there is no essential loss of generality in
qL2= q;),
a = 1,2;. ,V
assuming the form (16) for the {Q,}.
,@)
a = 1,2;-*,v - 1,
We refer to the above trial-and-error method as Method
a = i(q2
a + qa?,),
I. Another trial-and-error method is the following one,
with appropriate modifications if any of the resulting { QL2)} Method II. To simplify the discussion we assume for the
have vanishing mass. This step does not increasethe noise, moment that F is continuous and nowhere constant. We
in view of the discussion in Section V; that is, N(pc2))5
choose a trial value ql satisfying
N( p)).
The new {qf)}, centers of mass (c.m.) of the old {Q$},
41 < ( xdF(x).
J-cc
will not be centers of mass of the new {QL2)}, in general;
The condition that q, be the center of mass of Q, detrial point pc3) is determined by
termines x1 as the unique solution of
x(3) = .(2)
ff = 1,2,*.*,v - 1,
,a,

= (ac.m.

of Qi3)),

a = 1,2;*.,v.

For the resulting noise we have N(pC3))5 N(pc2)).


We continue in this way, imposing conditions (14) and
(17) alternately. There results a sequenceof trial points
p9 p27 . . .
(20)
such that

41=

p))

L ....

The noise is nonnegative, so that lim,N( pcm)) will exist,


and we might hope that the sequence(20) had as a limit a
local minimum of N(p).
If the sequence(20) has no limit points then some of the
{xLm)} must become infinite with m; this corresponds to
quantizing into fewer than v quanta. Since we have assumed that F increasesat least by v -t 1 points there will
be quantizing schemeswith v quanta for which the resulting noise is less than the optimum noise for v - 1 quanta,
obviously. If p( is such a schemethen (20) will have limit
points, using the property that N(p()) is a decreasing
sequence.2
Supposep() is such a limit point. If each of the coordinate values {x&m)} of p( is a continuity point of F then it
is easy to see that the coordinates of p() will satisfy both
(14) and (17). In particular, if N(p) has a unique stationary
point p,, (which is the minimum sought), then the sequence
(20), unless it diverges,will converge to p,,.
Note, by the way, that at a local minimum of N(p) the
numbers {xa} are necessarily continuity points of F. Suppose to the contrary that there is a nonvanishing amount of
mass concentrated at one of the,endpoints {x0>, and that
the adjacent sets Q, and Q,,, are as in (16), so that the
mass at x, belongs to Q,. The centers of mass q, and qa+ 1
21t seems likely that this condition A( p()) 5 (optimum noise for Y - 1
quanta) is stronger than necessary for the nondivergence of (20).

/::

dF(x)

-00

The quantities q, and x1 now being known, the first of


conditions (17) determines q2 as
q2

N( p() r N(

x xdF(x)

= 2x1

41.

If this q2 lies to the *right of the center of mass of the


interval (x,, co) then the trial chain terminates, and we
start over again with a different trial value q,. Otherwise,
x1 and q2 being known, the second of conditions (14):

servesto determine x2 uniquely. Now the second of conditions (17) gives


cl3 = 2x2 -

q2.

We continue in this way, obtaining successively


41, Xl, * . .,q,- ,, x,,- ,, q,; the last step is the determination
of q, according to
(21)
4, = 2x,-, - 4,-l.
However in this procedure we have not used the last of
conditions (14):

J
4, =

co xdF(x)

yz

dF(x)

(22)

X,-I
and the q, obtained from (21) will not satisfy (22) in
general. The discrepancy between the right members of
(21) and (22) will vary continuously with the starting,value
q,, and the method consists of running through such chains

133

LLOYD: LEASTSQUARESQUANTIZATION IN PCM

using various starting values until the discrepancy is reduced to zero.


This method is applicable to more general F, with some
obvious modifications. When F has intervals of constancy
the {xJ may not be uniquely determined by conditions
(14), and a trial chain may involve severalarbitrary parameters besides q,. Discontinuities of F will cause no real
trouble, since we know that the {xu} of an optimum
scheme are continuity points of F, a trial chain that does
not have this property is discarded. We note that Method
II may be used to locate all stationary points of N(p).
VII.

TABLE I
GAUSSIAN,Y = 2

a
1

co

4a
0.4528
1.5104

1
2

X,
0.9816
w

TABLE III
GAUSSIAN,Y = 8

EXAMPLES

. ..<q.-,<x,-,<q,;

x,

TABLE II
GAUSSIAN,~ = 4

In all of the examples we now consider, the distribution


of sample values is absolutely continuous with a sample
probability density f = F, which is an even function. If
N(p) has a unique stationary point, which we assumeto be
the case in the examples treated, then the optimum {qa}
and {xa} will clearly be symmetrically distributed around
the origin. In applications we are usually interested in
having an even number of quanta, v = 2~~ so we renumber
the positive endpoints and quanta according to
O=xo<q,<x,<

40
0.7919

a
1
2
3
4
5
6
7
8

-cQ~x-=c<.

The optimum schemesfor v = 2, b = 1,2,. . . ,7, are given


in Tables I-VIL3 respectively. The corresponding noise
values appear in Table VIII together with the quantities
v*N and vx,. The behavior of these latter with increasing v
hint at the existence of asymptotic properties; we examine
this question in the next section.
Since some of the tables were never completed, those tables although
mentioned in text are not included in this paper.

0.7584
0.3880
0.6568
0.9423
1.2562
1.6181
2.0690
2.1326

0.%82
0.5224
0.7996
1.0993
1.4371
1.8435
2.4008
co

TABLE VIII
GAUSSIAN;OPTIMUMNOISEFORVARIOUSVALUES OF Y
Y

Method II of the preceding section shows that N(p) in this


case has a unique stationary point, which is necessarily an
absolute minimum. The optimum schemeis the usual one
with v equal intervals of width l/(2&)
each; the quanta
being the midpoints of these intervals. The minimum value
of the noise is the familiar N = l/v*.
Another case of possible interest is the Gaussian:

X,
0.5006
1.0500
1.7480
co

TABLE IV
GAUSSXAN,~= 16

(23)

the endpoints and quanta for the negative half-axis .are the
negativesof these.
We normalize to unit signal power S = 1. The {q,} and
{x~} for other values of S are to be obtained by multiplying the numbers in the tables by \rs.
The simplest case is the uniform distribution:

4a
0.245 1
0.7560
1.3439
2.1520

1
2
3
4

2
4
8
16
32
64
128
(03)

Y2N

YXl

0.3634
0.1175
3.455 x 10-2
9.500 x 10-3

1.452
1.880
2.205
2.430

3.93
4.00
4.13

(0)

(2.72)

(4.34)

For speech signals a distribution which has been found


useful empirically is the Laplacian: 4

-lxl!h -m<xX<.
f(x)=c-.6

The optimum quantizing schemesfor this distribution for


v=2b,b=
1,2,-e. ,7, are given in Tables IX-XV, respectively. The corresponding N, v*N, and vx, values are given
in Table XVI; again, we notice certain regularities.
VIII.

ASYMPTOTIC PROPERTIES

Let us assume that the distribution F is absolutely continuous with density function f = F, which is itself dif4The author is indebted to V. Vyssotsky of the Acoustics Research
Group for this information (private communication).

134

IEEE TRANSACTIONSON INFORMATIONTHEORY,VOL. IT-28, NO. 2, MARCH 1982

ferentiable, and that for each v there is a unique optimum


quantization scheme.We revert to our original numbering
(10
Let the quantities {h,} be defined by
a = 1,2 7,V - 1,
ha = x, - 4, = qa+l - x,,

TABLE XVII
APPROXIMATELASTENDPOINT
FROMASYMPTOTICFORMULA

2
4
8
16
32
64
128

so that, for a = 2,3,-a. ,v - 1, Q, consists of an interval


of length h, to the right of q, together with an interval of
length ha-, to the left of qa. We have already imposed the
optimizing conditions (17) in the very definition of the
{h,}. The center of mass conditions (14) (except for the
first and last) may be written as
qa+ha x - q&(x)

dx = 0,

a = 2,3;..,v

- 1.

s, OL
-h a-,(

If we expand f here in Taylors series around q,, the


integration gives

t(hi - hit-,)fk,)

+ +@a+ hi-,)f(d

= o( h3,) + o( hi-,),

a = 2,3;*.,v

Gaussian
b
0
1.168
1.992
2.657

Laplacian
bv

and we find that the left-hand member of (24) is indeed


o(h3) = o(l/v3) provided that

g(x)
--g(x)

- fb>
3fb)

-09~x~co.

(26)

The normalized solution of (26) is


- 1. (24)

f 13(x)

g(x) = ( fl/3(x,)dx
The numbers in Tables VIII and XVI suggest the ex--ooxx~oo,
(27)
istence of an asymptotic fractional density of quanta.
J-cc
Accordingly, we define the function g,(x), - cc < x < cc,
provided that the integral in the denominator exists.
by
The noise power becomes
-cQooxxq,
d-4 = 09
N=I
II- 1
4, <x 5 q&+1,
12v2
2vh,
(25)
a = 1,2,-a-,v - 1,
(28)
=[I^
f1i3(x)dx]3+o(f),
= 0,
q,<x<m.
12v2 --oo
The definition is arranged so that (for given v) the sets
Q2, Q3,.. -,Q,-, subtend equal areas of l/v each under
the graph of g,(x) versusx. We will proceed as if a limiting
density,
g(x) = p~&)~

-caooxx<,

neglecting the contributions from the end quanta.6 For the


Gaussian example then, the numbers v*N of Table VIII
should have a limit easily evaluated from (28) as
v*N + no/2(=
2.72), and in the Laplacian case, Table
XVI, we find v*N --) 9/2.
The quantities denoted by vx, in Tables VIII and XVI
should have the limiting value

existed.
We wish to express g in terms of the given sample
density function f. To do this we will use conditions (24),
lim v( ha- 1 + h,) = -!together with the following further assumptions. We asv-m
g(O)
400
sume that g has a derivative, and we assumethat for given
x and k the difference c(x) = g,(x) - g(x), -cc < x (
comparing with (25). In the Gaussian example we find
co, has the property
l/g(O) = G(=
4.34), and for the Laplacian: l/g(O) =
3a( = 4.24).

For large values of v the sets {Q,} should subtend


approximately equal areas of l/v each under the graph of
g(x) versus x, so that the number b, defined by

In (24), then, we may approximate h, - ha-, by


ha - ha-,

$ = Jbmd:) dx
Y

1
1
i&z, + ha) - dqa - ha-,) 1

g(qJ
= - 2v2g3(qa)
5Thenotationu(v)

a = 2,3;..,v

= O(Z)(Y)) meansinourcaselim,,,u(v)/o(v)

- 1,

= 0.

might be expected to be near the rightmost division point.


Comparing Table XVII with Tables I-VII and IX-XVI we
see that the approximation is surprisingly good, at least in
the examplesconsidered.
60ther derivations of (27)-(28) are given in [6] and [7].

135

LLOYD: LEASTSQUARESQUANTIZATION IN PCM

convergeboundedlyto

ACKNOWLEDGMENT
The numerical results presented in the tables are due to
Miss M. C. Gray and her assistants in the Numerical
Analysis and Digital ProcessesGroup; the programming of
Method I for the IBM-650 electronic computer was done
by Miss C. A. Conn.
After substantial progress had been made on the work
described here there appeared in [ 1l] a review of a paper
by J. Lukaszewicz and H. Steinhauson optimum go/no-go
gauge sets. The present author has not been able to obtain
a copy of this paper, but it seemslikely that these authors
have treated a problem similar or identical to the one
discussedin Sections IV-VI. M. P. Schtitzenbergerin [12]
examines the quantization problem in the casewhere v = 2
and where F increasesat 3 or 4 points.
APPENDIX

Suppose s(t), - w < t < co, is a continuous parameter stochastic process, real, separable, measurable, stationary, and of
finite power:

-w<t<co,

S=E{s2(t)} =jm 2dF(x)<w,


-cc

(where F is the first-order distribution of the process, Section III).


Then s has a spectral representation

-co<t<w,
s(t) = jm e2*iAtdS(A),
(29)
--m
where the spectral process [(A), - cc < X < co, has orthogonal
increments [4, p. 5271. To say that s is band-limited to the
frequency band - W 5 X 5 W is to say that the [ process has
vanishing increments outside of this band with probability one,
and (29) becomes
s(t) =j-w;Joe2itdt(A).

-w<t<w.

(30)

Since we are particularly concerned with the behavior of & at the


band edges,we rewrite (30) as
s(t) = jpw~~oe 2rfxtd6(X)

+ 26,cos2aWt

S(X)=/lily,(h)

where the real random variables 6, and 6, describe the jumps of 6


at the band edges:

A=

i(t) =

X I W, has Fourier

J(t) = 1;i.z. / w+ S,(X) G(X)


-w-o
w-o
Z
PiA,,
+ 26, cos2?rWt
/ - w+o
= s(t) + 26,sin27rWt,
-w<t<w,

from [4, p. 4291. (The corresponding result for deterministic


functions is given in [9].) Since the orthogonal increments property of 5 requires E{ 6 ,S,} = 0 together with
E=

(8:)

=E{8;}

=+E{)[(W+O)-+W-O)12},

we see from (32) that the sampling series (31) represents s with
probability 1, if and only if, the [ process has no power concentrated at the band edges,
E(J[()W+O)-[(W-0)\2}

. . . ?d4-,)),
cpMGd)~ cp(4~1))> ...
(33)
constitute a stationary discrete-parameter stochastic process. If
the number
@ = E{q2(+,))}

=j_m_lp(x)

pwwq

Q+(t,))

-co

=j-wwe2ilh/(2Wd$(h),

<j<

cc),

-co cjc

co,

where the n process has orthogonal increments ([4, p. 4811,with a


change of scale).
A certain continuous parameter stochastic process e(t), -cc
< t < co, may be defined in terms of the n process by
=/-wwe2xdq(A),

-co

<t < co.

This process is stationary in the wide sense and has the given
process (33) as its samples, clearly
-al

<j -=cco.

= j~ww~{ldd~~ I} = j~mv2(x)df(x),
-w<t~co.

-wCj<w,

f - t,),

dF(x),

is finite then the process (33) admits a spectral representation

The 0 process is represented by the sampling series


e(i)=j=~ciP(s(~,))K(i-4),

-[

=O.

(Other proofs of this result appear in [3] and in [lo].)


Let s be as above and suppose q(x), - 00 < x < co, is a Bake
function. Then the random variables

E{e(i)}

in the notation of Section I. This function is of bounded variation, so that the partial sums
i

(32)

Moreover,

(2W))
29rW(t - j/ (2W))

jzz

(31)

converges(in stochasticmeansquare)to

e(t,) = cp(S(l,)),

= sin2mW(t -j/

s(t,)K(t - t,)

j=-m

1 w e-2Tijh/(2W)e277ihrd~
I = 2~ / _ w

ax)

w,
-cw,

from [8]. Hence,usingthe representation


(30) for the samples,the
samplingseries

e(t)

~(*W+O)-[(*W-O)=6,-Ci8,.

= K(t - tj),

-w<x<

,,+

- 26,sin2mWt,
-w~t~w,

For fixed t, the function e2jx, - WI


coefficients

= e2niXt

-woOttwoO,

(34)

- w 5 x 5 w,

if and only if, the spectral process q has no power concentrated at

136

IEEETRANSACTIONS
ON INFORMATIONTHEORY,VOL. IT-28, NO. 2, MARCH 1982

the band edges.The argumentsareidenticalto thosegivenabove


for the s processitself.
The I and n processes
of SectionsII and III are of the form just
described,sincethe functionsy(x), (6), and z(x), (9), will differ
from certain Baire functionsonly on setsof measurezero [dF]

when we assume, as we do, that the sets {Q,} are measureable

CP

Cl

[W.
The well-known mean-erogodic property,
b,

1.i.m.

m
2

m II-m'+1

mf~--m~-m j=mj

b2

b,

b3

x-

Fig. 1. The density f(x) vanishesoutside of the intervals (b,, b2),


(ha bd

(-l)'cP(s(tj))

=Tj(W+O)-q(w-0)+q(-w+0)-l)(-W-O)
=2Re[q(W+O)--(tW-0)],

([4, p. 4911) shows that the requirement that 9 have no power at


the band edges is equivalent to the condition

Finally we note that if the .$process has a discrete component


at frequencies *A,, then depending on the form of cp, the
derived n process is likely to have discrete components at all of
the harmonic frequencies krnh, (modulo 2W) of A,, m =
1,2;.. . In particular, if A, is rational then the n process may
have a discrete component at the band edges, a possibility which
must be excluded if (34) is to hold.
APPENDIX

-I

O I
XI

X-

Fig. 2. The densityf( x) vanishesoutsideof the interval(- 1,l).

into two equalparts and (- 1,O)is ignored.The parametersfor


this configuration work out to be
c2 - 9c,
41=4c,
2

A simple example shows that the conditions (14) and (17) are

c2

not sufficient for an absolute minimum of N. Suppose F is


absolutely continuous, with a density f = F as shown in Fig. 1,
where c,(b, - b,) + c2(b4 - b,) = 1. If Y > 1 quanta are desired, let v,, v2 > 0 be any integers such that v, + v2 = v, and
divide the interval (b,, b2) into v, equal intervals and (b,, b4)
into v2 equal intervals; let the quanta be the midpoints of these v

Xl =T,

3(c2
q2

then the division point which separates the right-hand {Q,} in


(h,, b3), so that the conditions (14) and (17) will be satisfied.
Thus we have v - 1 distinct local minima of N. (If c,, respectively c2, is small enough there may even be another minimum,
corresponding to v, = 0, respectively v2 = 0.) Which of these is
the true minimum dependson the values of the parameters.
(Explicitly, the noise has the value

NE&-

intervals.If we supposethat b, < +(b, + b,) and b, > &(b2+ b4)


(h, , b2) and the left-hand{Q,) in (b,, b4) will lie in the interval

- 3c,

4c2

cl)
9

(c2-3c~)3
16~;

Hence if this configurationexiststhen it is better than the one


first mentioned. We note, by the way, that method I of Section
VI will convergeto the (- l/2, 0, l/2) configuration,if the
starting value xl, is negative; if c2 > 3c,, however, this config-

uration is only a saddlepoint of N.


Authors

Note 1981:

This is nearly a verbatim reproduction of a draft


manuscript,
which was circulated for comments at Bell
12v:
12v;

Laboratories; the Mathematical Research Department log


and the v2/v, for which this is a minimumis givenby
date is July 3 1, 1957. I wish to thank the editors for their
invitation
to publish this antique samizdat in the present
c2
3
v2/
(h
- b3)
issue.
v,/(b2-b,)=

( ;
1
The main reason the paper was not submitted for publiagreeing with (27).)
cation previously was that the numerical calculations were
The following interesting example is due to J. L. Kelly, Jr., of
never completed. The Gaussian v = 32 case was done on
the Visual Research Group. Let the density f be as in Fig. 2, with
c2 > c,, c, + c2 = 1. The signalpoweris S = l/3, independently the IBM 650 card programmable calculator; the Laplacian
of c, and c2. Supposev = 2. One configurationfor which condi- cases were done only for v = 2. Some time later the 650
was replaced by an IBM 701 electronic computer, but no
tions (14) and (17) are satisfied is q, = - 4, x, = 0, q2 = 4,
clearly, and the resultingnoise in N = l/ 12. When c2 > 3c, ; quantizing program was written for it.
I was not satisfied with not having conditions for a
however, there is another solution-it is the one which in the
limit c, = 0, c2 = 1 goes into the scheme where (0,l) is divided
unique minimum but would have published the paper
N

c,(b2

bJ3

c2(hl

b313

137

IEEE TRANSACTIONSON INFORMATIONTHEORY,VOL. IT-28, NO. 2, MARCH 1982

without this. Later, P. E. Fleischer of Bell Laboratories


gave a neat sufficient condition in his paper [ 131.
In the examplesof Appendix B, the direct current can be
removed by changing the origin; the noise is not affected.
The results of the paper are valid for the uncentered
processesused.
I was aware when I wrote the paper that the methods for
quantizing a real random variable extend to other loss
functions. In the least squares case the process quantizing
noise is just the noise per sample; the generalization is
more complicated and was omitted.

[41
[51

[61
[71

PI
[91

[lOI
[Ill

REFERENCES
B. M. Oliver, J. R. Pierce, and C. E. Shannon, The philosophy of
PMC, Proc. I. R. E., vol. 36, pp. 1324-1331, 1948.
PI H. S. Black, Modulution Theory. Princeton, NJ: Van Nostrand,
1953.
[31 S. P. Lloyd and B. McMillan, Linear least squares filtering and
prediction of sampled signals, in Proc. Symp. on Modern Network

[II

[I21
1131

Synthesis, vol. 5. Brooklyn, NY: Polytechnic Institute of Brooklyn,


1956, pp. 221-247.
J. L. Doob, Stochustic Processes. New York: Wiley, 1953.
H. Cramer, Mathematical Method of Statistics. Princeton, NJ:
Princeton University, 195 1.
P. F. Panter and W. Dite, Quantization distortion in pulse-count
modulation with nonuniform snacing
- of levels. Proc. I. R. E.. vol.
39, pp. 44-48, 1951.
B. Smith, Instantaneous companding of quantized signals, Bell
Syst. Tech. J., vol. 36, pp. 653-709, 1957.
A. Zygmund, Trigonometrical Series. New York: Dover, 1955, p.
47.
H. P. Kramer, A generalized sampling theorem, Bull. Am. Math.
Sot. vol. 63, p. 117, 1957.
E. Parzen, A simple proof and some extensions of the sampling
theorem, Department of Statistics, Stanford Univ., CA, Tech.
Report 7, Dec. 1956.
.I. Kukaszewicz and H. Steinhaus, On measuring by comparison,
Zustos. MU., vol. 2, pp. 225-231, 1955; Math. Reviews, vol. 17, p.
757, 1956.
M. P. Schtitzenberger, Contribution aux applications statistiques
de la theorie de linformation, Pub. de Ilnst. de Stutis. de IUniversite de Paris, vol. 3, Fast. 1-2 pp. 56-69, 1954.
P. E. Fleischer, Sufficient conditions for achieving minimum distortion in quantizer, IEEE Int. Convention Record, part I, vol. 12,
pp. 104-111, 1964.

The Hexagon Theorem


DONALD J. NEWMAN

A k&act- We wish to minimize ls 12 - f( 2) I, S the unit square, over


all choices off(Z) which take just n values. For large n it is shown that
f(Z) approachesthat function which on each face of a hexagonal mesh
has the value of the centroid of that face. This problem in the design of
quantizers for a pulse-codemodulation (PCM) system arose in which pairs
of adjacent samples are quantized in an effort to capitalize on the correlation between samples and on the geometry of the plane.

E WISH TO MINIMIZE /s 1Z - f(Z) 12,S the unit


square, over all choices of f(Z) which take just n
values. For any fixed n this seemsa hopelesslycomplicated
problem. For n large, A. J. Goldstein has conjectured that
the minimizingf( Z) approachesa hexagonal mesh function. Indeed, if S, , S,, . . . , S,, are the usual tessellation of S
into equal regular hexagons and some boundary pieces,

Manuscript received Oct. 2 I, 198 1,


The author was with Bell Laboratories as a consulting engineer. He is
now with the Mathematics Department, Temple University, Philadelphia,
PA, 19122.
Editors Note: This paper first appeared as a Bell Laboratories Technical Memorandum in May 1964. The principal theorem of this paper was
first proved by L. Fejes Toth in his book Lagerungen in der Ebene, auf der
Kugel und im Raum, Springer-Verlag, Berlin, 1953. Newmans paper
provided the first exposition in English of the result and a new and simple
proof.

and f(Z) = centroid of S, for Z E Si then

where
(JXZ- 5
18fi .
This fact and the following theorem verify Goldsteins
conjecture in a very sharp way.
Theorem: If S is the unit square and f( Z) is any (measurable complex valued) function on S, which takes only n
distinct values, then

p - f(Z)12
>f
where
5
u=iip.
Proof:

Let f*(Z)

be the actual minimizer. (Existence is

trivial.)

001%9448/82/0300-0137$00.75 01982 IEEE

You might also like