619
CORRESPONDENCE
of one communication link lock onto the crosscorrelation peaks
obtained by correlating with the encoding sequence of a different
communication link. Thus the successful use of spread spectrum
communication systems in multiplexing
applications depends
upon the construction of large families of encoding sequences with
uniformly low crosscorrelation values. In this paper we present
an analytical technique for the construction of such families of linear
binary encoding sequences.
Equation (17) is equivalent to the recursion
(SNR), a,,
2
2
2
a,  an+ = ___N
whose solution is
a; = a; 1 + qq.
[
II.
Just as in Schalkwijklzl we generate the estimates recursively, i.e.,
a2,
@NW,
e"@
_
1)
+
@
p
8'(n)
= 1+ 1
1 I (StiR)i % (20)
N
J. P. M. SCHALKWIJK
L. I. BLUESTEIN
Communication Systems Laboratories
Sylvania Electronic Systems
Div. of Sylvania Elec. Prods., Inc.
Waltham, Mass. 02154
REFERENCES
11 T. J. Gobliok,
Theoretical
limitations
on the transmission
of data from
an&g
murces,
IEEE
Trans.
Information
Theory,
vol.
ITl&
pp. 558567,
October
1965.
A coding
scheme for additive
noise chitnnels
with
El J. P. M. Schalkwijk,
IEEE
Trans.
Information
Theory.
feedback,
Part
II:
Bandlimited
signals,
vol. ITU,
pp. 183189,
April 1966.
Ial T. Kailath,
An application
of Shannons
r&edistortion
theory
to analog
Proc. Princeton
Symp. on Systm
Science,
oommunicrttion
over feedback
channels,
Mrtrch
1967.
141 T. J. Cruise,
Achievement
of ratedistortion
bound
over additive
white
noise channel utilialng
& noiseless feedback
channel,
Proc. IEEE
(Letters), vol. 55,
pp. 583584,
April 1967.
NOTATION
Following the notation of Zierlex+l we denote by V(f) the vector
space of linear sequences generated by the recursion relation corresponding to the polynomial f of degree n and further identify
the members of V(f) with binary 2*  1 tuples. If f is a primitive
polynomial over the field K = (0, 1) then h e V(f) implies h is a
maximal linear sequence. W e denote by ]lhll the number of ones in
the sequence h and by & the sequence such that R(i) = h(z) + 1.
The correlation function 8 of two binary sequences a, b has been
defined as
where x is the unique isomorphiim of the additive group (0, 1)
onto the multiplicative group { 1, 1). W e note that (?(a, b) (T) is
simply described ss the number of agreements  number of disagreements of the sequences a and b for each 7 and that
O(a, b) = 2 
1  2 I/a + bjj.
In what follows a will always denote a primitive 2  1 root of
unity in a splitting field of x21 + 1 and the minimal polynomial
of & will be denoted by fi. Finally, we note the following result of
Bose and Chaudhuri.131
Theorem 1
Optimal
plexing
Binary
Sequences for Spread Spectrum
I.
Multi
INTRODUCTION
Linear shit register sequences (see Zierlerlll and Gold121) have
found extensive applications in spread spectrum communication
systems. The binary sequences generated by shit register devices
serve as the encoding mechanism of such systems which, when
added to the baseband information, results in a wideband lowpowerdensity signal which has statistical properties similar to
noise. The casual listener is thus denied access to the baseband
information which can be recovered from the wideband signal
only through correlation with a stored reference sequence in the
receiver which is an exact replica of the original encoding sequence.
The usefulness of the maximal linear sequences in spread spectrum
communications depends in large part on their ideal autocorrelation
properties. The autocorrelation function of a binary sequence h is
defined as O,(T) = (number of agreements  number of disagreements) when the sequence h is compared with a cyclic shift of itself.
It is well known that for maximal linear sequences s,(O) = period
of the sequence and &,(7) =  1 for 7 Z 0. The detection by the
receiver of the high inphase correlation value oh(O) determines
the synchronization between transmitter and receiver necessary
for the removal of the encoding sequence and the recovery of the
baseband information. In multiplexing applications many systems
will be operating in the same neighborhood and each communication
link will employ a different maximal encoding sequence. In general,
the crosscorrelation function between different maximal sequences
may be relatively large. Thus dierent systems operating in the
same environment can interfere with the successful attainment
and maintenance of proper synchronization by having the receiver
Manuscript
received
July
Let LYbe any primitive element of the splitting field of zm1 + 1.
Let fi be the minimal polynomial of at. Let
22, 1966;
revised
February
4, 1967.
2n + 1
715, f2, .   LJ
lcm
Then a, b E V(g) implies Ila + bj j > 2k.
III.
STATEMENT
AND
PROOF
OF RESULT
Our techniques for the construction of large families of encoding
sequences with uniformly low crosscorrelation values is based on the
following result.
Theorem d
Let (Y be any primitive element of GF(2). Let fi be the minimal
polynomial of DI. Let ft be the minimal polynomial of CY~where
(+1)2) + 1
t = p
b (n+2)2) + 1
(n odd)
(n even).
Then a e V(f,) and b E V(ft) implies [@(a, b) 1 5 t.
The significance of this theorem is that it tells how to select
shit register tap connections which will generate maximal linear
sequences with a known bound on the crosscorrelation function.
Since r~ is primitive, the sequence generated by the shift register
corresponding to fi is maximal. Since
2(n+1,/2
and
s/n+z)/z
are both relatively prime to 2n  1 for n $ 0 mod the 4 sequence
corresponding to the polynomial ft is also maximal in these cases.
Theorem 2 thus permits the selection of pairs of maximal sequences
with known bound on the crosscorrelation function. This result
is of practical importance since, for example, for n = 13 there are
630 maximal sequences and there exist pairs of these sequences
620
IEEE
whose correlation values are as high as B = 703 while Theorem 2
guarantees the selection of pairs of sequences such that 101 5 129.
This result is a special case of the more general theorem stated
in the following which has been obtained independently by Goldfsl
and Kasami,f41 and is related to the weight distribution of errorcorrecting codes.
where
TRANSACTIONS
ON
INFORMATION
v(g
[ v(O),V(l), * . *
,v(+
b(i) = 5!((~zL+1)i)
where 01is a primitive 2  1 root of unity (n odd), I is any integer
such that (I, k) = 1, and T is the trace of GF(2). Then e(a, b) (n) =
 1 when u(7) = 0 and
i 2 ( (L+1)2+
qu, b)(T) =
1)
when
or
(n+1)/2
 1)
(2
a(~) = 1.
z21 + 1.
t ; 1; 01.
n1,2 ones
x2 + 1
= Icm (fl, fz, . . . f.I)
then.fv I gol,
Proof:
= u (by Lemma 2) = 2  1
>2n1 > v  1 implies fU / g,l (by Lemma 1)
mf.
Let
fU and f. are factors of g+l where u and a are as in Lemma 2
and
mfs
2  1 root, of unity in a splitting field of
0,
1)
Lemma S
Lemma 1
Let 01be any primitive
 2)v(?z 
Again it is clear that any cyclic permutation of [v(O), . . . v(n  l)]
will result in a larger integer and hence mf, = e. A similar argument
holds when a = 2n1  1  2(n/2).
l
The proof of this theorem is contained in Gold151.
In the remainder of this section, we proceed to the proof of
Theorem 2 by means of a series of lemmas.
1967
= [l, l,>;. * l,,
n1,2 ones
Let a and b be maximal linear sequences given by
and
OCTOBER
fj n+1
__( 2 > . . . v(n
Theorem
a(i) = T(ori)
THEORY,
= v (by Lemma 2) > B  1 impliesf,
/ g,,l.
Lemma 4
**I+ 1
gk= lcm7fl,fz, **. fk)
where fi is the minimal polynomial of (Y<.Let f be an irreducible
polynomial of degree n. Let A, be the conjugate class of roots off.
Let ml = min {i 1 tii P Af), the class leader of A,. Then rnt > k
implies f is a factor of gk.
Proof: f irreducible of degree n implies f ( zinl + 1 implies f 1 gk
lcm Vi, f2, . *. fkl, and rnf > k implies ft j Icm (fi . . . fk) implies
f 1 gk.
Lemma .%?
Let 01be any primitive
2%  1 root of unity in a splitting field of
X
21 +
1.
Let u = 2*l  1. Let
v=
2l 
1  2
for
n odd
for
n even.
Let ,fc denote the minimal polynomial
Then a, b e V(gh) implies ]@(a, b) j < 2  1  21~.
Proof: a, b e V(gk) implies a + b e V(gk) implies Ijo + b] 1 > k by
Theorem 1. Since 1 is not a primitive root of unity, 1 + z is clearly a
factor of g, and hence V( 1 + z) C V(gk). Thus the constant sequence
.of ones is a member of V(gk), and hence V(gk) is closed with respect
to the operation of complementation, i.e., h f V(gk) implies h c V(hd
where A(i) = 1 + h(i).
Thus a, b e V(gk) implies a + b z V(gk) implies a + b c V( gk)
implies ]G]]
= 2n  1  ]]a + bl! > k implies k < jja + bl/ <
2%  1  k implies ]0(a, b)l < (2n I)
 2k.
Proof of Theorem 2: Let p = a2. Then 0 is clearly a primitive
2  1 root of unity and
o1 =
1 y1
y/2
Let fU be the minimal polynomial of 0%. Let f% be the minimal
polynomial of au. Then mf, = u and rnf, = a.
Proof: 01r and a* belong to the same conjugate class of GF(2) if,
and only if, there exists an integer k such that (Y = (01~)~if, and
only if, T = s.2k modulo 2%  1 if, and only if, there exists a cyclic
permutation p such that [r(O), r(l), . . . r(n  l)] = [s(p(O)),
4PU)) . . . s (p(n  l))] where
nl
n1
r = c ~(42
i=o
and
s = c ~(232~.
i=
Now u = 2%  1 = c::h
u(i) 2; where [u(O), u(l), . . . u(n  l)] =
[l 1 **. 1 01. Clearly any permutation of [l 1 . 1 0] corresponds
to a larger integer and hence mf, = U. Now
2) = 2nl
y12
n1
=
c
i=o
v(4y
of ai. Let
a2(2n1
at zz
1)
p
+ )
1
ia(2n+
L w+)2
_ 1 _ pw7
+ 1
(for
n odd) =
for
n even;
a*[2
= p _ 1 _ 21/2 = p
J
.
= pl _ 1 _ y/z 1
\ a2[24 _ 1 _ y/y
Now fl is the minimal polynomial of 01 = flu and ft is the minimal
polynomial of at = &,. By Lemma 2 ml, = u > v  1 and mf, =
u > v  1. Thus by Lemma 1 fi andft are factors of gol. Hence,
V(fd C V(g&
and v(fd C I%,&
Thus o e V(fd and b p V(ft)
implies a and b E (S.I). Thus by Lemma 4
Ie(a, b)/ < 2 
7
2  1 
q2:
(2*  1  qy
_

2
2
1  2(v  1) =
_

y19
29
=
=
+ 3
2(%+2)/z + 3.
2(*+1)/z
621
CORRESPONDENCE
Since the value of the crosscorrelation function is always odd we
have
le(a, b)I 5 2n+12 + 1
for
n odd
le(a, b) 1 2 2r+22 + 1
for
n even.
For 13stage shift registers there are pairs of maximal sequences
with crosscorrelation peaks as high as O(T) = 703 while proper
selection of shift registers in accordance with the above theorem
guarantees sequences whose cross correlation satisfies the inequality
ItI
1 < 23+)2 + 1 = 129.
W e note further that for purely random sequences of length 23 1 = 8191 we would expect the crosscorrelation function to exceed
A 180 for 5 percent of the correlation values and hence
2a=22
linear sequences chosen in accordance with our technique perform
bett,er with respect to their crosscorrelation properties than purely
random sequences.
IV.
CONSTRUCTION
OF ENCODING
FAMILIES
In this section we show how to provide large families of encoding
sequences each of period 2n  1 and such that the crosscorrelation
function of any pair of sequences of the family has a crosscorrelation
function 0 which satisfies the inequality
le(T)l 5 2(n+z)z + 1.
In a spread spectrum multiplexing application such families form
ideal codes which minimizes interlink interference. Instead of
having each communication link employ a different maximal
sequence we assign to each link a member of the encoding family
to be constructed below. These are nonmaximal linear sequences,
and hence their autocorrelation function will not be twovalued;
however, the outofphase value of the autocorrelation function
will satisfy the above inequality. Thus by slightly relaxing the
conditions on the autocorrelation function we obtain a family of
encoding sequences with the high crosscorrelation peaks eliminated.
The procedure for generating these encoding families is embodied
in the following theorem.
Theorem
Let fi and ft be a preferred pair of primitive polynomials of
degree 12 whose corresponding shift registers generate maximal
linear sequences of period 2  1 and whose crosscorrelation fmlction
e satisfies the inequality.
jet< t =
2(n+1z + 1
for
n odd
2(n+2)2 + 1
for
n even
Z* + 211 + ~12 + ~14 and the corresponding 14stage shift register
will generate 129 different linear sequences of period 127. The
crosscorrelation function B of any pair of such sequences will
satisfy the inequality /e(7)] 5 17.
ROBERT
REFERENCES
111 N. Zierler,
Linear
recurring
sequences,
J. SIAM,
vol. 7, March
1959.
@I R. Gold,
Characteristic
linear
sequences
and
their
ooset
functions,
(accepted
for publication
J. Sot. Ind. ii&.
Math.,
May
1965).
[$I W. W. Peterson,
Error Correcting
Codes. New York:
Wiley,
1961.
141 T. Kasami,
Weight
distribution
formula
for some class of cyclic
codes,
University
of Illinois?
Urbana,
Rept. R265, April 1966.
[sl R. Gold, Maxunal
recursive
sequences
with 3valued
recursive
crosscorrelation functions
(submitted
for publication,
January
1967).
Average DigitError
Codes
Probability
INTRODUCTION
Let
Then the shift register corresponding to the product polynomial
fi.ft will generate 2* + 1 different sequences each period 2n  1
and such that the crosscorrelation function 0 of any pair of such
sequences satisfies the above inequality.
Proof: u E V(fl.ft) = V(fd + V(fJ implies
a = b +
c where
b E V(fi) and c E V(fJ. Period (b + c) = lcm (period b, period cl =
2  1. Since degree fi.ft = 2n, there are (azm l)/(an  1) = 2 i1 essentially different sequences in V(fi .f t). Finally a, b P V(fl .f t)
implies
a = a, + a,
i b = b, +
b,
where al, bl E V(fi), and at, bt z V(ft). \@(a, b)\ = IB(al I bl, at +
bt) 1 5 t by Theorem 2 since al + bl E V(fi) and UI $ b, E V(f t).
Thus, by way of illustration, if we consider the pair of polynomials,
fi(x) = 1 + z + x2 + x3 + x7 and f dz) = 1 + x I x2 i x3 I x4 Ix5 + x7 then the product polynomial isfi(x) fz(x) = 1 I 2 I x6 f
After Decoding Random
W h e n a transmitted code word is decoded incorrectly due to
channel errors, it does not follow that all of the digits in the decoded
word are incorrect. One way of getting an estimate of the actual
number of digit errors to be expected is to determine the average
over a whole class of codes of the digiterror probability and to
compare this with the worderror probability averaged over the
same class. In this correspondence the ratio of these two averages
is determined for random block codes and it is shown that, at fixed
code rate and channelerror probability, t,he ratio approaches a
nonzero limit with increasing code length; the value of this limit
is given as a function of code rate and channelerror probability.
It is pointed out that the result can be extended to random linear
codes and to the information digits of random parity check codes
is also given. A more detailed derivation for these two linear cases
is given elsewhere.
It should be stressed at the outset that both the average worderror probability and the average digiterror probability can be
strongly influenced (particularly at small channel probabilities)
by a very few codes with unusually large word and digiterror
probabilities. Hence, it is unlikely that there is any easy extension of
these results to statements about the distribution of the ratio of the
two probabilities over the classes of codes considered.
NOTATION
n # mod 4.
GOLD
Magnavox Research Laboratories
Torrance, Calif. 90503
n =
R =
K =
p =
AND
ASSUMPTIONS
code length
code rate
2R
channelerror probability
H(t) =  t logz t  (1  t) log, (1  t)
pR = smaller solution of R + H&)
= 1
PC
PR/(l
2pR
2PR)
pw = average worderror probability
pD = average digiterror probability
E = expected value of.
The
means
f(t)ldt)
The
as724
symbol N used in equations of the form f(t) N g(t) as t + 03
that the ratio of the two functions approaches unity:
f 1.
expression o(n) used in equations of the form f(n) = o(n)
~0 meansf(n)/n+Oasn+
m.
Manuscript
received
December
12, 1966; revised
May
1 J. N. Pierce,
Average
digit error probability
after
codes,
Air Force
Cambridge
Research
Labs.,
Bedford,
October
1966.
22, 1967.
decoding
Mass..
random
linear
Kept.
66695,
Lebih dari sekadar dokumen.
Temukan segala yang ditawarkan Scribd, termasuk buku dan buku audio dari penerbitpenerbit terkemuka.
Batalkan kapan saja.