Академический Документы
Профессиональный Документы
Культура Документы
A bstract-The srmpliag theorem forbandbited signals of W e energy error in reading the sample values produces only a
can be interpreted in two ways, associated with
the names of Nyquist and correspondingly small error in the recovered signal.
sbnsaoo. 2) Every square-summable sequence of numbers may be
1) Every signal of W e energy and bandwidth W Hz may be com- transmitted at the rate of 2 W per second over an ideal
pletely recovered, in a simple way, from a knowledge of its samples channel of bandwidth W Hz, by being represented as
taken at the rate of 2W per secwd (Nyquist rate). Moreover, the
reeoverybseMe,inthe~thatasmfllerrorinrefdiagsample
the samples (at the points k/2W) of a bandlimited
values produces only a correspondingly small error in the recovered signal of finite energy.
Signnl.
2) Every sqalre-srrmmable sequesce of nmnbers may be tnutsmitted at
Thus the sampling theorem serves as basis for the inter-
the rate of 2Wper second over an ideal channel of bandwidth W H z , changeability of analog signals and digital sequences, so
by being represented as the samples of an easily conshocted band- valuable in modem communication systems. The rate of
limited signrl of W e energy. 2W samples per second for W Hz of bandwidth is often
The practical importaoce of these results, together with the restrictiom called the Nyquist rate.
implicit in the sampling theorem, make itnatnral to ask whether the above In view of the. practical importance of these results, it is
rates cannotbe improved, by ppssing to differently chosensampling instam@ natural to ask whether the Nyquist rate prescribed above
or to badpass or mdliband (rather than badimited) signak, or to more
elaborate compotatioos.In tbis paper wedraw a distiactioo between re- for stable sampling or for data transmission cannot some-
combwthgasignalfromitssamples,anddoiisoinastibleway,andwe how be improved upon. Since 1)and 2) refer to a very special
argue that only stable sampliq is meaaingfal in practice. We then prove situation-to signals which in frequency occupy only a
that : single band and to their values at regularly spaced instants-
1) stablesamplingcannotbeperformedataratelowerthanthe it is conceivable that signals might be recoverable from
Nyquist, their values taken at a lower rate, if the sampling instants
2) data cannot be transmitted as samples at a rate higher than the were chosen differently; or if the signals had their fre-
Nyquist,
quencies in a union of several bands; or at thecost of more
regardless of the locatioa of sampling instants, the nature of the set of fre-
qwacies whicb the signals occupy,or the method of comhcth.
These
computing than is required by the simple formula (1).Were
this possible, we would have a more efficient reduction of
cwclmions apply not merely to fiuiteswrgy, but Pt90 to bomded, signals.
continuous signals to digital form, with a consequent saving
INTRODUCTION of bandwidth. Similarly, it is conceivable thatarbitrary
HE sampling theorem states that f(t) is a signal of square-summable numbers could be prescribed at a higher
finite energy bandlimited to W Hz if and only if rate as the sample values of a signal, if the sampling instants
were properly chosen, or if the carrying signals were allowed
sin 271W(t- k/2W) frequencies in an appropriate union of bands. Were this
f ( t )= & " 271W(t - k / 2W )
with 1 < cx) ; (1) possible, we would have a more efficientwayof sending
data over given bandwidth.
thereupon The purposeof this paper is to prove that stable sampling
rm of signals cannot be performed at less than the Nyquist
rate, nor can data be transmitted as samples at a greater
rate, regardless of the location of sampling instants, the
nature of the set of frequencies which the signals occupy,
or the method of computation. Although discussed in the
This may be read in two ways, each of which has found context of finite-energy signals, these results also apply to
important applications. bounded signals.
1) Every signal of finite energy and bandwidth W Hz
FORMULATION OF THE PROBLEM AND
may be completely recovered from a knowledge of its
STATEMENT OF RESULTS
samples taken at the rate of 2W per second. More-
over-indispensable for any implementation in prac- We consider a fixed set S of N-frequency bands whose
tice-the recovery is stable, in the sense that a small total finite length (counting negative as well as positive
frequencies) we denote by m(S), and the space g ( S ) of all
Manuscript received March 3,1967; revised June 12, 1967. signals of finite energy with frequencies contained only in
The author is with Bell Telephone Laboratories,Inc., Murray Hill, N. J. S . Specifically, we make the following definition.
Authorized licensed use limited to: Universidad Nacional de Colombia. Downloaded on January 17, 2010 at 19:48 from IEEE Xplore. Restrictions apply.
1702 PROCEEDINGS OF THE IEEE, OCTOBER 1967
Dejinition: B(S) consists of all square-integrable func- We show in the Appendix that points of a set of inter-
tions f ( t ) whose Fourier transform F(w)is supported on s, polation for B(S)are necessarily separated by at least some
i.e., F(w)= 0 for o 4 S. positive distance d, and that the interpolation can be per-
We have in the past called such signals multiband. formed in a stable way, i.e., the function f ( t ) E B ( S )can be
We start by considering those sets of points t= {tk} at chosen so that
which the values of a signal f ( t ) E B ( S )can be used to re-
construct the signal everywhere. We assume from the outset
that these points separated
are by at least some positive
jrnlf(t)12
-m
dt 1 lak12
set of uniqueness, whereas the reverse is not true. More- We begin with an intuitive explanation of the proof. In
over, our definition does not commit us to any particular what follows wewilluse the term “sampling” to mean
recovery scheme,and so describes the most general circum- exclusively “stable sampling.”
stances in which sampling can be performed in practice. Suppose { t k } is a set of sampling for B(S) and I is an
Takingup next the problem of transmitting data as interval containing n ofthe points of {tk}.Supposef(t)EW(S)
samples, we consider instants t = {tk} at which square- has little of its energy outside I ; we will referto such a func-
summable values can be prescribed arbitrarily for a signal tion as being well concentrated on I . If we can presume that
f ( t ) E B ( S ) Here
. too we formulate the matterindependently the samples f ( t k ) at points tk outside Z are also small, then,
of any method of computation. by virtue of our discussion of stability, replacing these
Dejinition : The set of points { tk} is a set of interpolation sample values by zero should not too drastically affect our
for W ( S )if, given any set of numbers {ak} with C lak12< co,
there exists f ( t ) E g ( S )with f(tk)=ak. With log’ r = max (log r, 0).
Authorized licensed use limited to: Universidad Nacional de Colombia. Downloaded on January 17, 2010 at 19:48 from IEEE Xplore. Restrictions apply.
SAMPLING,
LANDAU: TRANSMISSION,
DATA AND NYQUIST RATE 1703
fIf (t)I2 dt
N
-
1
log+ zir - 1
Irn
i= 1 n2
(3)
A = I f(t)I2 dt m(S)r
2--
271
A log+ r - B,
-m
achieves its largest value. This is a standard problem of with constants A and B depending only on S . Thus, from
maximizing a quadratic form and leads immediately to the ( 5 ) and (8),
1 I.z(r, S ) = double integral of (kernel12. Let S be bounded and {tk} be a set of interpolation for
(6)
k=O g ( S ) , whose points are separated by at least d >O. Let I be
any compact set, I - be the set of points whose distance to
To evaluate the latter most easily, we observe that (4) may the complement of I exceeds d/2, and n ( I - ) be the number
be converted to other equationswithout changmg the eigen- of points of { tk} contained in I - . Then An(l-) - l(r, S )2 6 >0,
values. One such transformation, obtained by introducing where 6 depends on S and {tk} but not on I .
into (4) the definition of w, multiplying both sides by e-iz", We have already remarked that the eigenvalues A k ( z , S )
and integrating over I , gives remain unchanged when we reverse the roles of the time and
Authorized licensed use limited to: Universidad Nacional de Colombia. Downloaded on January 17, 2010 at 19:48 from IEEE Xplore. Restrictions apply.
1704 PROCEEDINGS OF THE IEEE,OCTOBER 1967
frequency domains. It is also easy to see from (4) that they Proof: We follow the proof of Theorem 1, interchanging
are unaffected by a rescaling of time and frequency which the roles of Lemmas 1 and 2. Thus with Z an interval of
replaces Z by aZ and S by a- ' S , with a any constant.2 By length r, n(Z-)2 n(Z)-2, and, by Lemma 2, there exists a
applying both of these permissible transformations, we con- constant 6 >0 independent of r such that
clude that i k - l(r, S ) is also the kth eigenvalue p f the prob-
lem of concentrating functions whose frequencies lie in a
single interval of length 2n (i.e., bandlimited functions) on
the set (r/2n)S. Butnow theordinary sampling theorem
shows the integer points to be a set both of sampling and
interpolation. When S is the union of N intervals, the num-
ber of these points contained in S + is at most (2n)-'rrn(S) so that
+ 2N, and their number in S - is at least (2n)-'rm(S)- 2 N . (n(Z) - 3) - ([rm(S)(2n)-'] + 2 N ) < C'log' r + D',
Thus Lemmas 1 and 2 establish the existence of absolute
constants yo and 6, such that
whence
4(2x)-'rm(s)1+2N(r,S ) Yo < 1, (10) n(Z) 5 rm(S)(2n)-' + Clog' r + D.
i[(2x)-'rm(S)]-2N-1(r, 2 > O; (11)
Theorem 2 is established.
here we use the notation [x] to denote the integer part of
It follows immediately that the density of a set of inter-
x. It follows from their construction that y o and do may be
polation cannot be higher than rn(S)/2n.
taken to be 3, but we will not prove this here.
In the preceding argument we used Lemmas 1 and 2, to- CONCLUDING
REMARKS
gether with known sets of sampling and interpolation, to
determine the behavior of the { i k ( r ,S ) } . Now the lemmas 1 ) The aforementioned density results apply equally well
allow us in turn to apply this information to arbitrary sets to functions of several variables and to frequency sets S
of sampling or interpolation. more complicated than finite unions of intervals. A complete
account of this extension is to be found elsewhere[31.
Theorem 1 2) It may well be argued that functions of finite energy
If { t k } is a set of sampling for @S), every interval of are too special a class, and that bounded functions present
'
length r must contain at least (rm(S)(2n)- -A log' r - B) greater practical interest. The problem of describing fre-
of the points of { t k } ,with A and B appropriate constants. quency content is rather severe in this case, the Fourier
Proof: Let Z be an interval of length r, and n(Z) the num- transform not being available, but meaningful definitions
ber of points of { t k } contained in I ; since I is a single in- can be given (using distributions, for example) for the class
+
terval, n(Z + ) < n(Z) 2. By Lemma 1, there exists a constant g b ( S )of bounded signals with frequencies in S ; we will not
y < 1 independent of r such that discuss the details here. We can then formulate the notions
of stable sampling and interpolation as follows.
& ( I ) + 2(r, S ) Y < 1. (12) Definition : The points { t k } form a set of stable sampling
But from ( 1 1) for g b ( S ) if there exists a constant K such that
-m
If(')12
dt I 1 lakl2 = 1lf(tk)12.
some positive distance d, and the interpolation can be per-
formed in a stable way. This shows that when interpolation is performed in d?(S),it
Proof: By writing f € W ( S ) as the inverse transform of its is necessarily stable. Proposition 1 is established.
Fourier transform, and applying Schwarz's inequality and Lemma 1
Parseval's theorem to this representation, we find
1
Let S be bounded and { t k } be a set of sampling for B(S),
whose points are separated by at least d >O. Let I be any
IK, lf(t)I2 dt, (15) compact set, I + be the set of points whose distance to I
lm
-m is less than 4 2 , and n(I+ ) be the number of points of { t k }
+
contained in I . Then )(I, S)I y < 1 , where y depends
If'(t)12 IK, lf(t)I2 dl, (16) on S and ( t k ) but not on I .
-m
Proof: By the Weyl-Courant lemma,
with constants K , and K , depending only on S . Thus for
signals in W ( S )a bound on energy implies a corresponding
bound on both the signal and its derivative.
If the points ( t k } are not separated by at least some fixed
positive distance, we can find two disjoint subsequences, J-m
which we denote by { p k } and { q k ) , with the property that
Ipk-qkI I l / k 2 , k = 1 , 2 , . . . . Now let us prescribe the values with ck any subspace of dimension k. Thus to prove Lemma
l / k at t = p k , - l / k at t = 4 k , and 0 at every remaining point 1 , it is sufficient to show that we can keep the concentration
of { t k } ; since these values are square-summable, and { t k } is of a function of B(S)over I bounded away from 1 by im-
a set of interpolation, they will be assumed by some posing n(I +) orthogonality conditions. The following
simple device accomplishes this. Let h b ) be a square-
f(t)Eg(S).But then 2 / k = f b k ) - f ( q k ) 5 l p k - qkl lf'(sk)l with
sk apoint lyingbetween p k and q k . Hence I f ' ( s k ) l > 2 k , integrable function which vanishes for ly[ > d / 2 and whose
k = 1 , 2, . . . , contradicting ( 1 6 ) . This establishes the first Fourier transform H ( o ) satisfies IH(o)l> 1 for OES.Such a
part of the proposition. function certainly exists, since the transforms of functions
To prove the second part, we appeal toan abstract vanishing on lyl> d / 2 can be made to approximate uniformly
principle. In W(S), let usview functions as vectors and an arbitrary continuousfunction on any bounded set. Now
energy as the square of length ; this makes B(S) into a given f€&(S), we form
Hilbert space. Let d?'(S) be the subspace of g ( S ) consisting
of those functions which vanish at all the points { t k } . This
subspace is closed, since, by ( 1 5), convergence in energy of
functions of to a function h(t) implies uniform con- 1 r
Now let us consider the transformation T which assigns and by Parseval's theorem
to a square-summable sequence { a t } that function of 8 ( S )
r m r r
which interpolates to ak at t = t k . By hypothesis, T is de-
fined on the whole space 1, of square-summable sequences,
maps 1, into the Hilbert space d?(S), and, in view of ( 1 5 ) ,
has a closed graph. Then by the closed graph theorem (see
Loomi~[~], p. 1 8 ) T is bounded, i.e., there exists a constant
= Im -m
Ig(t)12d t .
1706 OF THE IEEE, OCTOBER
PROCEEDINGS 1967
Now subjecting f to the n(Z +) orthogonality conditions with some fixed constant K . For each t k , we let & ( t ) E & ( S )
be the function whose value is 1 at t k and 0 at every Z j ,
a j # k ; these functions are linearly independent.
f ( y ) h ( t k - y ) d y = 0, tk E I+ 9
We let h be the same as in the proof of Lemma 1, and
--a:
construct functions + k E a ( S ) such that
ensures that g ( t k ) = O , t k E I + , whereupon the sum in (20)
is extended only over t k $ I + .For those t k , the corresponding
integrals in (19) are taken over non-overlapping intervals,
4k(t) =
1 “
271 - m
7
Il/kb)h(t - y )d y ; (24)
each contained outside I ; thus combining (21), (20), and
(19) we find to see that these exist, we need only take the Fourier trans-
form of (24) and note that I H ( o ) l 2 1 for OES.The $’s are
likewise linearly independent, so that, letting C be the sub-
space of W ( S ) spanned by the functions +k with t k in I - ,
we obtain a subspace of dimension n(I -).
or Now given ~ E Cwe , form the function g of (18). In view
I
j
If (t)12 d t
< l - - =1y < l .
of (24), g is a linear combination of the $ k , hence lies in
&(S) and satisfies (21).However, since the only q i k figuring
in g are those with t k in I - , the values of g necessarily
jm If @)I2 d t
--a:
Kc vanish at every other t k , so that
I-m
J-m
with ck any subspace of dimension k. Any proof of (17) can which, applied to (22), proves Lemma 2. As in Lemma 1,
be modified to yield(22), but the following independent the constant 6 does not depend on I.
argument serves as well. Let L k - l be the subspace of B(S)
spanned by &,,. . . ,dk-2,the first k - 1 eigenfunctions. REFERENCES
Since the dimension of Lk-lis lower than that of ck, any [‘I A. Beurling and P. Malliavin, “On the closure of characters and
linear transformation of C, into & - I , in particular the zeros of entire functions,” Acta Mathematica, vol. 118, pp. 79-93, May
orthogonal projection onto L,-,, must annihilate some 1967.
H . J. Landau, “A sparse regular sequence of exponentials closed
nonzero element of ck. Thus Ckcontains a function ff0 on large sets,” Bull. Am. Math. Soc.,vol. 70, pp. 5 6 5 6 9 , July 1964.
orthogonal to Lk- and (22) follows from the definition of -, “Necessary density conditions for sampling and interpolation
of certainentire functions,” ActaMathematica, vol. 117, pp. 37-52,
‘ k - 1. February 1967.
We have shown in Proposition 1 that it is possible to [41 H. J. Landau andH. 0.Pollak, “Prolate spheroidal wave functions,
perform the interpolation in a subspace &(S) of a(S),in Fourier analysis and uncertainty.111,” Bell Sys. Tech.J., vol. 41, pp. 1295-
which every function g ( t ) satisfies 1336, July 1962.
c5] L. H. Loomis, An Introduction toAbstract Harmonic Analysb.
m New York: Van Nostrand, 1953.
F. Riesz and B. S. Nagy, Functional Analysis. New York: Ungar,
(23) 1955.
-m