1: \documentstyle[aps,prl,twocolumn,epsfig]{revtex}
2: \newcommand{\be}{\begin{equation}}
3: \newcommand{\ee}{\end{equation}}
4: \newcommand{\erf}{\mbox{erf}}
5: \begin{document}
6: \twocolumn[\hsize\textwidth\columnwidth\hsize\csname@twocolumnfalse\endcsname
7: \title {Low autocorrelated multi-phase sequences}
8: \author{ Liat Ein-Dor$^1$,Ido Kanter$^1$,
9: and Wolfgang Kinzel$^2$}
10: \address{
11: $^1$ Department of Physics, Bar-Ilan University\\
12: \mbox{~} Ramat-Gan, 52900 Israel \\
13: $^2$ Institut f\"{u}r Theoretische Physik, Universit\"{a}t W\"{u}rzburg\\
14: \mbox{~} Am Hubland, D-97074 W\"{u}rzburg, Germany
15: }
16: \maketitle
17:
18: \begin{abstract}
19: The interplay between the ground state energy of the generalized
20: Bernasconi model to multi-phase,
21: and the minimal value of the maximal autocorrelation function,
22: $C_{max}=\max_K{|C_K|}$, $K=1,..N-1$, is examined
23: analytically and the main results are:
24: %liat
25: (a) The minimal value of $\min_N{C_{max}}$ is $0.435\sqrt{N}$
26: significantly smaller than the typical value for random sequences
27: $O(\sqrt{\log{N}}\sqrt{N})$.
28: (b) $\min_N{C_{max}}$ over all sequences of length $N$ is
29: obtained in an energy which
30: is about $30\%$ above the ground-state energy of the generalized
31: Bernasconi model,
32: independent of the number of phases $m$.
33: (c) The maximal merit factor $F_{max}$ grows linearly with $m$.
34: (d) For a given $N$, $\min_N{C_{max}}\sim\sqrt{N/m}$
35: indicating that for $m=N$,
36: $\min_N{C_{max}}=1$, i.e. a Barker code exits.
37: The analytical results are confirmed by simulations.
38:
39: \end{abstract}
40: \pacs{PACS numbers: 05.20.-y, 87.10.+e}
41: ]
42: In many applications of communication science \cite{Schroder,Shapiro},
43: as well as a variety of other fields, it is necessary to find
44: sequences with low autocorrelation functions.
45: Some of these applications utilize the pulse compression feature
46: of the low autocorrelated sequences to obtain high resolution in
47: radars and sonars. In other applications, the shifts of
48: such periodic sequences can be used to identify users in multi-user
49: systems.
50: Due to their importance, low autocorrelated sequences
51: have evoked a wide spread interest accompanied
52: by the development of
53: various methods for constructing such sequences \cite{Schroder}.
54:
55: In order to construct a binary sequence $S=(s_1,...,s_N)$ with low off
56: peak autocorrelations, one has to define which quantity has to
57: be minimized.
58: The different applications divide the low autocorrelated sequences
59: into two types in correspondence with the quantity which they minimize.
60: The first kind of low autocorrelated sequences minimizes
61: the Hamiltonian of the Bernasconi model
62: \cite{Bernasconi,Krauth,Mezard,Parisi} which is given by:
63:
64: \be
65: H=\sum_{K=1}^{N-1}C_{K}^2
66: \label {energy}
67: \ee
68: where for the case of non-periodic boundary conditions, which is at
69: the center of our study,
70: \be
71: C_{K}=\sum_{j=1}^{N-K}s_js_{j+K},
72: \label {CK}
73: \ee
74: and for the periodic case
75: $C_{K}=\sum_{j=1}^{N}s_js_{((j+K-1)\bmod N)+1}$.
76: Note that for random sequences the average value of $H$
77: in the non-periodic case is $N^2/2$,
78: whereas for low autocorrelated sequences this energy is reduced
79: by a merit factor $F>1$ to $N^2/2F$.
80: However, there are applications for which
81: the maximal off peak autocorrelation, $C_{max}=\max_K{|C_K|}$,
82: has to be minimized. The second kind of low autocorrelated sequences
83: are the solutions of this minimization problem.
84: %liat
85: Note that for random sequences $C_{max}$ is typically
86: $O(\sqrt{\log{N}}\sqrt{N})$.
87: For a sequence of length $N$,
88: the maximal possible ratio between the peak, $|C_0|=N$,
89: and the maximal off-peak autocorrelations,
90: $|C_K|$ with $K=1,...,N-1$ is $N/1$.
91: The only known binary sequences with this ratio are the Barker codes of length
92: $2$, $3$, $4$, $5$, $7$, $11$ and $13$. Obviously, the Barker sequences,
93: when they exist, furnish a minimum for the two minimization problems.
94: However, Turyn \cite{Turyn} has shown that no other
95: binary codes such as this exist for any length less than $144$ or for odd
96: length greater than $13$.
97: An exact solution for these two minimization problems
98: is known only for systems (up to $N=59$)
99: which are small enough
100: to permit an effective exhaustive search \cite{Mertens1,Mertens2}.
101: Extrapolation of the ground state energies which were
102: found for small systems using an exhaustive search
103: indicates
104: $F_{max}=\lim_{N\to\infty}\frac{N^2}{2H}=8.5$ \cite{Mertens1}. Moreover,
105: $F_{max}$ was conjectured by Golay
106: \cite{golay1,golay2} to be bounded
107: from above by $12.324$.
108:
109: The following questions regarding low autocorrelated sequences
110: are still open and are at the center of our study: (a) Do
111: the sequences that minimize ${C_{max}}$ minimize
112: the energy of the Bernasconi model as well?
113: (b) In case that the two minimization problems are not equivalent, how
114: far are the energy values of the sequences which minimize ${C_{max}}$
115: from the values of the ground state energy?
116: (c) Exhaustive search of sequences with length up to $N=59$ show that
117: the degeneracy of the ground state is of O(1) and is bounded from
118: above by $8$. Is the degeneracy of the states which minimize
119: $C_{max}$ is of O(1) as well?
120: (d) In this context, which of the two minimal quantities is easier
121: to approach?
122: Additional interesting questions arise when multi-phase
123: sequences, whose terms are complex $m$-th roots of $1$
124: for $m>2$, are considered. For such sequences
125: \be
126: s_l=\exp(i\frac{2\pi}{m}l)
127: \label {s_l}
128: \ee
129: where l=1,...,m and the correlations are defined as
130: \be
131: C_{K}=\sum_{j=1}^{N-K}s_j{s^*_{j+K}}.
132: \ee
133: where $s^*_{j+K}$ is the conjugate of $s_{j+K}$.
134: The information about multi-phase low autocorrelated sequences
135: is more limited, mainly since the configuration space grows
136: like $m^N$ which makes the exhaustive search ineffective even
137: for very small sequences. Hence, not much is known
138: about the influence of
139: the number of phases both on the
140: maximal value of $F$ and on the minimal value of $C_{max}$.
141: Is it possible to attain better solutions by increasing the
142: number of phases?
143:
144: In this study we suggest an analytical technique to calculate the
145: quantities of interest, namely, $\max_{N}F$ and $\min_{N}{C_{max}}$
146: for both binary and multi-phase sequences.
147: An obvious lower bound of $\min_{N,F}{C_{max}}$ over
148: all sequences of length $N$ and a merit factor
149: $F$ is obtained by assuming that all $|C_K|$ equal $\sqrt{N/(2F)}$.
150: Nevertheless it is tempting to investigate the relations between $F$
151: and $\min_{N,F}{C_{max}}$ and to measure more accurately
152: the optimal value $\min_{N}{C_{max}}$ over all sequences of length $N$
153: similarly to the aforementioned upper bound given for $F_{max}$ \cite{golay1}.
154: For our analytical study we consider the non-periodic
155: Bernasconi model. We approximate the sequences to be random
156: and the autocorrelation functions to be independent variables
157: with the following Gaussian distribution
158: \be
159: P(C_{K})=\frac{1}{\sqrt{2\pi(N-K)}}\exp(\frac{-C_{K}^2}{2(N-K)}).
160: \label {P(C)}
161: \ee
162: Using Eq. (\ref {P(C)}), the probability $P_F(C_{max})$ that the
163: autocorrelations of a sequence with a merit factor
164: $F=N^2/2H$ have an upper bound $C_{max}$ such that $|C_K|\le {C_{max}}$
165: is given by
166: \be
167: P_F(C_{max})=\int_{-C_{max}}^{C_{max}}\prod_K{dC_K}
168: P(C_{K})\delta(\frac{1}{N}\sum_{K=1}^{N-1}{C_K^2-
169: \frac{N}{2F}}) .
170: \label {P_F}
171: \ee
172: Since there are $2^N$ distinct binary sequences of length $N$,
173: it is necessary that $P_F\ge{2^{-N}}$ in order for a sequence with the
174: corresponding features, namely a merit factor $F$
175: and an upper bound $C_{max}$,
176: to exist. Hence, by equating $P_F(C_{max})$ to $2^{-N}$ one can
177: find the minimal upper bound $\min_{N,F}{C_{max}}$ among the upper bounds
178: of all sequences with a merit factor $F$.
179: Moreover, the Gaussian distributions of $C_K$ imply that
180: $\min_{N,F}{C_{max}}$ must be of $O(\sqrt{N})$ in order for
181: $P_F(C_{max})$ to be equal
182: $2^{-N}$. Assigning
183: $\min_{N,F}{C_{max}}=B(F)\sqrt{N}$ and inserting the integral representation
184: of
185: $\delta$ function in
186: Eq. (\ref {P_F}) , the saddle point method can be used to obtain the
187: following set of equations
188: \begin{eqnarray}
189: \frac{\lambda}{2F}-\frac{1}{4\lambda}((1-2\lambda)\ln(1-2\lambda)+2\lambda)
190: \nonumber \\
191: -\int_0^{1}\ln{\erf(B(F)g(y,\lambda))}dy=\ln{2},
192: \label {A_min1}
193: \end{eqnarray}
194:
195:
196: \begin{eqnarray}
197: \frac{1}{2F}-\frac{1}{4\lambda^2}(\ln(1-2\lambda)+2\lambda)+
198: \nonumber \\
199: B(F)\sqrt{\frac{2}{\pi}}\int_0^{1}{{\exp(-B(F)^2
200: g(y,\lambda)^2)}\over{
201: \erf(B(F)g(y,\lambda))\sqrt{2}g(y,\lambda)}}dy=0.
202: \label {A_min2}
203: \end{eqnarray}
204: \noindent
205: where $g(y,\lambda)=\sqrt{\frac{1-2\lambda(1-y)}{2(1-y)}}$.
206: Solving numerically this set of equations, it turns out that a solution
207: exists only for a bounded region of $F$, where $.215\leq F\leq 12.324$.
208: These two limit values of $F$ are alternatively obtained by
209: taking the limits of the integration in Eq. (\ref {P_F})
210: to infinity.
211: The existence of a lower bound is a consequence of the fact that as the energy
212: becomes smaller the allowed values of the correlations decrease
213: and the probability for such sequences (Eq. (\ref {P_F}))
214: decreases as well. As was pointed out above, when this probability drops
215: below $2^{-N}$, there are no more sequences with the required energy.
216: Similarly, for high energy values, high values of correlations are
217: required and the corresponding probabilities become smaller again.
218: The analysis of $B(F)$ inside the allowed region reveals
219: that the minimal value of $B(F_c)=0.435$ is obtained for $F_c=8.839$
220: whereas
221: the maximal merit factor $F=12.324$
222: corresponds to a higher value of $B(F)=1$. Thus, the
223: merit factor of sequences which minimizes $C_{max}$ is around $71\%$
224: lower then $F_{max}$.
225: Exhaustive search results in the region $N\in [15,28]$
226: indicate that the average ratio between
227: the two merit factors, $<<<F(\min{C_{max}})>>/F_{max}>_N$,
228: is $\sim 0.6771$ where $<<$ $>>$ denotes the average
229: over all sequences which minimize ${C_{max}}$ whereas $<$ $>_N$ denotes
230: the average over the different sequences length $N$.
231: The deviation from the analytical result $\sim 0.717$
232: is attributed
233: to finite size effects.
234: Fig. \ref{fig1} shows the behavior of $B(F)$ derived from the numerical
235: solution of Eqs. (\ref {A_min1}) and (\ref {A_min2}).
236: The behavior of $P_F(C_{max})$ divides this graph into three regimes.
237: Starting with the large $F$ regime, there is only
238: one sequence whose $C_{max}$ equals $B(F)\sqrt{N}$.
239: This feature holds until $F_c=8.839$ for which
240: the minimal $B(F)=0.435$ is obtained.
241: Further decrease of $F$ results in an increasing number
242: of sequences
243: with an upper bound of $B(F)\sqrt{N}$. This increment terminates at
244: $F=0.8$ and for smaller values of $F$ the number of sequences
245: reduces to $1$ as $F$ approaches $0.215$.
246: In order to compare the analytical behaviour of $B$
247: to that obtained from simulations, it is necessary to find a way
248: to circumvent the mismatch between the analytical values of $F$ and
249: those obtained from simulations.
250: The rescaling of $F$ by $F_{max}$ yields a parameter $F/F_{max}$
251: whose range of values, $[0,1]$, does not depend on a specific realization.
252: This allows a comparison between analytical results and simulations.
253: Results of an exhaustive search over sequences of length $N=32$
254: show that the behavior of $B$ as a function of $F/F_{max}$ resembles
255: the analytical prediction (Fig. \ref{fig2}).
256:
257: Moreover, the existence
258: of the three aforementioned regions
259: of $P_F(C_{max})$ is confirmed by counting the number of sequences with
260: the minimal upper bound $B$ for each of the micro-canonical ensembles
261: with a merit factor $F$ and length $N=32$.
262: The histogram is demonstrated in the inset of Fig. \ref{fig2}, where
263: the merit factor is rescaled to $F/F_{max}$.
264: In the small $F$ regime, the average number of sequences increases
265: with $F$ until $F/F_{max}\sim 0.2$ from which the number
266: of sequences decreases up to the plateau which starts at
267: $F/F_{max}\sim 0.83$.
268:
269: \begin{figure}
270: %\centerline{\epsfxsize=3.25in \epsffile{Bernfig1.eps}}
271: \narrowtext \epsfxsize=3in
272: \hskip 0.5in
273: \epsfig{width=0.9\columnwidth,file=Bernfig1.eps}
274: \vskip 0.15in
275: \caption{$B(F)$ vs. $F$. This graph is composed of three regions
276: corresponding to the three different behaviours of $P_F(C_{max})$.
277: In the dashed region $0.215\leq F\leq 0.8$,
278: the fraction of sequences $P_F(C_{max})$
279: is getting larger as $F$
280: increases, while in the dotted-line region, $0.8<F<F_c$,
281: this fraction decreases
282: to $2^{-N}$ as $F$ approaches $F_c$.
283: For $F>F_c$ (solid-line)
284: the fraction of sequences
285: with the minimal value
286: $B(F)\sqrt{N}$ is constant and equals $2^{-N}$.}
287: \label {fig1}
288: \end{figure}
289:
290: %\vspace {0.1cm}
291: \begin{figure}
292: \narrowtext \epsfxsize=3in
293: \hskip 0.5in
294: \epsfig{width=0.9\columnwidth,file=Bernfig2.eps}
295: \vskip 0.1in
296: \caption{$B$ vs. $F/F_{max}$. The solid line represents
297: the analytical results while the circles represent the exhaustive search
298: results for $N=32$. Note that the analytical $F_{max}$ is $12.324$ whereas
299: for $N=32$ $F_{max}=8$.
300: Inset: a histogram describing the logarithm of the number of sequences with
301: the minimal upper bound $B$ for each of the micro-canonical ensembles
302: with a merit factor $F$ as was derived from an exhaustive search
303: for $N=32$.}
304: \label {fig2}
305: \end{figure}
306:
307:
308:
309: Now we turn to study the influence of the number of phases both on the
310: maximal value of $F$ and on the minimal value of $C_{max}$.
311: We assign $D_K$ and $E_K$ to be independent Gaussian variables which
312: represent the real and the imaginary parts of $C_K$
313: respectively. In case that $m=2^p$ where $p$ is an integer number greater
314: than $1$,
315: these two parts have the same probability
316: distribution
317: \be
318: P(D_{K})=\frac{1}{\sqrt{\pi(N-K)}}\exp(\frac{-D_{K}^2}{(N-K)}).
319: \label {P_m(D)}
320: \ee
321: Under the same assumptions that have been used in the derivation
322: of the binary case,
323: Eq. (\ref {P_F}) becomes
324: \begin{eqnarray}
325: P^m_F(C_{max})=\int_{-C_{max}}^{C_{max}}\prod_K{dD_KdE_K}
326: P(D_{K})P(E_{K})
327: \nonumber \\
328: \delta(\frac{1}{N}\sum_{K=1}^N{(D_K^2+E_K^2)-\frac{N}{2F}}).
329: \label {P_F(B)}
330: \end{eqnarray}
331: Similarly to the binary case,
332: if $P^m_F(C_{max})=m^{-N}$
333: then $B(F)\sqrt{N}$
334: is the minimal value of $C_{max}$ for which
335: there is a sequence with a merit factor $F$.
336: The same procedure that has been used to calculate $B(F)$
337: for the binary case results in
338: the following set of equations
339: \begin{eqnarray}
340: \frac{\lambda}{2F}-\frac{1}{\lambda}((1-\lambda)\ln(1-\lambda)+\lambda)
341: \nonumber \\
342: -2\int_0^{1}\ln{\erf(B(F)g(y,\lambda))}dy&=&\ln (m).
343: \label {B_min1}
344: \end{eqnarray}
345:
346: \begin{eqnarray}
347: \frac{1}{2F}-\frac{1}{\lambda^2}(\ln(1-\lambda)+\lambda)+
348: \nonumber \\
349: \frac{2B(F)}{\sqrt{\pi}}\int_0^{1}{{
350: \exp(-B(F)^2g(y,\lambda)^2)}\over{
351: \erf(B(F)g(y,\lambda))g(y,\lambda)}}dy=0,
352: \label {B_min2}
353: \end{eqnarray}
354: where $g(y,\lambda)=\sqrt{\frac{1-\lambda(1-y)}{(1-y)}}$.
355: Solving numerically these two equations,
356: the behaviors of $\min_N{C_{max}}/\sqrt{N}$ and $F_{max}$ as a function of $m$
357: are obtained. Fig. \ref{fig3} shows that $\min_N{C_{max}}/\sqrt{N}$
358: is proportional to $1/\sqrt{m}$ with slight deviations for small $m$.
359: This relation between $B(F_c)=\min_N{C_{max}}/\sqrt{N}$ and $m$ implies
360: that such a relation holds for $B(F_{max})$ as well.
361: Since the variance of $C_K^2$ equals $N-K$,
362: their typical values drop linearly with $K$.
363: Hence, $H=N^2/2F$ is approximated by a sum of an algebric series of $N-1$
364: terms, $C_K^2$,
365: with an upper bound, $C_{max}^2$, of $O(N/m)$.
366: Assuming that $C_K^2$ is homogeneously distributed
367: between $0$ to $C_{max}^2$ yields a linear increment of
368: $F_{max}$ as a function of $m$ in agreement with the numerical solution
369: of $F_{max}$ which is depicted in the inset of Fig. \ref{fig45}.
370: The results of $F_{max}$ and $B(F_c)=\min_N{C_{max}}/\sqrt{N}$
371: as a function of $m$ show that
372: the solutions of the two minimization problems are improved by
373: increasing the number of phases $m$.
374:
375: \vspace{4mm}
376: \begin{figure}
377: %\centerline{\epsfxsize=3.25in \epsffile{Bernfig3.eps}}
378: %\vspace{0.4cm}
379: \narrowtext \epsfxsize=3in
380: \hskip 0.5in
381: \epsfig{width=0.9\columnwidth,file=Bernfig3.eps}
382: \vskip 0.15in
383: \caption{
384: $\min_N{C_{max}}/\sqrt{N}$ vs. $m$ as was
385: obtained from the numerical solution of Eqs.
386: (\ref {B_min1}) and (\ref {B_min2}).
387: The solid line is the least square fit $0.54/\sqrt{m}$.}
388: \label {fig3}
389: \end{figure}
390:
391: \vspace*{-3mm}
392: \begin{figure}
393: %\epsfxsize= 0.49\textwidth
394: %\epsffile{Berninset45.eps}
395: \narrowtext \epsfxsize=3in
396: \hskip 0.5in
397: \epsfig{width=0.9\columnwidth,file=Berninset45.eps}
398: \vskip 0.15in
399: \caption{Results for $\min_{N}{C_{max}}$
400: as a function of $m$ for sequences of length $N=32$.
401: The filled circles stand for the simulated annealing results.
402: The deviation from the analytical results (solid-line) is attributed to
403: finite size effects and to the suboptimal solution
404: obtained in our limited
405: running times of the simulations.
406: Inset: Analytical results for $F_{max}$ as a function of $m$
407: which were obtained from the numerical solution of Eqs. (\ref {B_min1})
408: and (\ref {B_min2}). The solid line is the least square fit, $3.7m$.}
409: \label {fig45}
410: \end{figure}
411:
412: These results raise the
413: question whether it is possible to increase $m$ such that
414: $\min_N{C_{max}}$
415: becomes 1. An asymptotic expansion of Eqs. (\ref {B_min1})
416: and (\ref {B_min2}) reveals that for $m=N$, $\min_N{C_{max}}=1$.
417: Note that $C_{N-1}$ is always 1 and therefore $\min_N{C_{max}}=1$
418: holds for the entire regime $m>N$.
419:
420: %$C_K$ is a sum of $N-K$ terms of magnitude $1$ with $m$
421: % possible directions in the complex space. Since $m=2^p$,
422: % each of the $m$ possible vectors has an inverse vector
423: %in its opposite direction. Hence when $N-K$ is odd,
424: % $C_K$ is bound from below by $1$ corresponding to optimal cases
425: % in which pairs of terms in
426: %opposite directions are cancelled and one single term is left.
427: %Consequently, the result $\min_N{C_{max}}=1$ holds for the entire
428: %regime $m>N$.
429: We used the simulated annealing
430: method for sequences of lengths $N=32$
431: with different number of phases $m$, to find $\min_N{C_{max}}$.
432: The relatively small sequences size was chosen to
433: enable an appropriate scan of the configuration space
434: in reasonable computational time.
435: The results are exhibited in Fig. \ref{fig45},
436: and support the anticipated $\sqrt{N/m}$ behavior of $\min_N{C_{max}}$.
437: For $m=32$, there is a deviation of $\min_N{C_{max}}$
438: from the analytical prediction
439: $\min_N{C_{max}}=1$, probably
440: since the simulated annealing method yields only
441: suboptimal solutions.
442:
443: Finally, we would like to examine the simulated annealing method in
444: light of the results of this study. Simulations show that
445: for the same running times, the simulated
446: annealing method yields ${C_{max}}$ which is closer to its
447: minimal value than $\sum_{K=1}^{N-1}C_K^2$.
448: This can be explained by the larger degeneracy of ${C_{max}}$
449: compared with that of $\sum_{K=1}^{N-1}C_K^2$.
450: Moreover, it turns out that in order to minimize ${C_{max}}$,
451: it is preferable to start the searching process with the minimization
452: of the energy function, $\sum_{K=1}^{N-1}C_K^2$,
453: and then replace it with ${C_{max}}$ \cite {unpublished}. In this way
454: the system avoids the plateaus which characterize the landscape
455: of $C_{max}$ in the configuration space.
456: However, future research is necessary to find out how the
457: results of this study can be applied to further improve the
458: searching processes
459: of low autocorrelated sequences.
460:
461: Critical comments on the manuscript by R. Metzler are acknowledged.
462: The research is supported in part by the Israel Academy of science.
463:
464: \vspace*{-0.5cm}
465: \begin{thebibliography}{10}
466: \vspace*{-1cm}
467: \bibitem{Schroder}
468: M. R. Schroeder 1984 {\em Number Theory in Science and Communication}
469: (Berlin: Springer)
470:
471: \bibitem{Shapiro}
472: I.I.Shapiro, G. H. Pettengill, M. E. Ash,
473: M. L. Stone, W. B. Smith W B, R. P. Ingalls and
474: R. A. Brockelman, Phys. Rev. Lett. {\bf 20}, 1265 (1968)
475:
476: \bibitem{Bernasconi}
477: J. Bernasconi, J. Physique {\bf 48}, 559 (1927).
478:
479: \bibitem{Krauth}
480: W. Krauth and M. Mezard, Z. Physik B {\bf 97}, 127 (1995).
481:
482: \bibitem{Mezard}
483: J. P. Bouchaud and M. Mezard, J. Phys. I France {\bf 4}, 1109 (1994).
484:
485: \bibitem{Parisi}
486: E. Marinari, G. Parisi and F. Ritort, J. Phys. A {\bf 27}, 7615 (1994).
487:
488: \bibitem{Turyn}
489: R. Turyn, Sylvania Electronic Systems, Final Rept. Contract,
490: AF19(604)-5473, January 29, (1960).
491:
492: \bibitem{Mertens1}
493: S. Mertens, J. Phys. A {\bf 29}, 473 (1996).
494:
495: \bibitem{Mertens2}
496: S. Mertens, C. Bessenrodt J. Phys. A {\bf 31}, 3731 (1998).
497:
498: \bibitem{golay1}
499: M. J. E. Golay,
500: {\em IEEE Trans. Inf. Theory} {\bf IT-28}, 543 (1982).
501:
502: \bibitem{golay2}
503: M. J. E. Golay,
504: {\em IEEE Trans. Inf. Theory} {\bf IT-29}, 934 (1983).
505:
506: \bibitem{unpublished}
507: L. Ein-Dor (unpublished)
508:
509: %\bibitem{Schroeder:Zahlentheorie}
510: %M. Schroeder, {\em Number Theory in Science and Communication}
511: % (Springer-Verlag, Berlin, 1984).
512:
513: \end{thebibliography}
514:
515: \end{document}
516:
517:
518: