1: \documentclass[11pt]{article}

2: \usepackage[colorlinks=true,

3: linkcolor=webgreen,

4: filecolor=webbrown,

5: citecolor=webgreen]{hyperref}

6: 

7: \usepackage{amssymb,psfig,epsfig}

8: \usepackage{asa}

9: 

10: \definecolor{webgreen}{rgb}{0,.5,0}

11: \definecolor{webbrown}{rgb}{.6,0,0}

12: 

13: \bibliographystyle{asa}

14: 

15: 

16: \setlength{\textwidth}{6.5in}

17: \setlength{\textheight}{9.25in}

18: \setlength{\oddsidemargin}{0in}

19: \setlength{\topmargin}{-0.25in}

20: \setlength{\headheight}{0in}

21: 

22: \newtheorem{lemma}{Lemma}

23: \newtheorem{theorem}[lemma]{Theorem}

24: \newtheorem{coro}[lemma]{Corollary}

25: \newtheorem{exam}{Example}[section]

26: \newtheorem{remark}{Remark}[section]

27: 

28: \newcommand{\eqn}[1]{(\ref{#1})}

29: \newcommand{\eeq}{\end{equation}}

30: \newcommand{\beql}[1]{\begin{equation}\label{#1}}

31: \newcommand{\bsq}{{\vrule height .9ex width .8ex depth -.1ex }}

32: \newcommand{\La}{\Lambda}

33: \newcommand{\af}{\alpha}

34: \newcommand{\sV}{{\mathcal V}}

35: \newcommand{\RR}{{\mathbb R}}

36: 

37: \makeatletter

38: \def\@sect#1#2#3#4#5#6[#7]#8{\ifnum #2>\c@secnumdepth

39:      \def\@svsec{}\else

40:      \refstepcounter{#1}\edef\@svsec{\csname the#1\endcsname.\hskip .75em }\fi

41:      \@tempskipa #5\relax

42:       \ifdim \@tempskipa>\z@

43:         \begingroup #6\relax

44:           \@hangfrom{\hskip #3\relax\@svsec}{\interlinepenalty \@M #8\par}%

45:         \endgroup

46:        \csname #1mark\endcsname{#7}\addcontentsline

47:          {toc}{#1}{\ifnum #2>\c@secnumdepth \else

48:                       \protect\numberline{\csname the#1\endcsname}\fi

49:                     #7}\else

50:         \def\@svsechd{#6\hskip #3\@svsec #8\csname #1mark\endcsname

51:                       {#7}\addcontentsline

52:                            {toc}{#1}{\ifnum #2>\c@secnumdepth \else

53:                              \protect\numberline{\csname the#1\endcsname}\fi

54:                        #7}}\fi

55:      \@xsect{#5}}

56: \def\@begintheorem#1#2{\it \trivlist \item[\hskip \labelsep{\bf #1\ #2.}]}

57: 

58: \def\section{\@startsection {section}{1}{\z@}{-3.5ex plus -1ex minus 

59:  -.2ex}{2.3ex plus .2ex}{\normalsize\bf}}

60: \makeatother

61: \makeatletter

62: \def\subsection{\@startsection {subsection}{1}{\z@}{-3.5ex plus -1ex minus

63:  -.2ex}{2.3ex plus .2ex}{\normalsize\bf}}

64: 

65: \makeatother

66: 

67: 

68: 

69: \begin{document}

70: \begin{center}

71: {\large {\bf The Lattice of $N$-Run Orthogonal Arrays}} \\

72: \vspace{1.5\baselineskip}

73: E. M. Rains and N. J. A. Sloane \\

74: \vspace*{1\baselineskip}

75: Information Sciences Research Center \\

76: AT\&T Shannon Lab \\

77: Florham Park, New Jersey 07932-0971 \\ [+.25in]

78: John Stufken \\

79: Department of Statistics \\

80: Iowa State University \\

81: Ames, IA 50011 \\

82: \vspace{1.5\baselineskip}

83: April 20, 2000 \\

84: \vspace{1.5\baselineskip}

85: {\bf ABSTRACT}

86: \vspace{.5\baselineskip}

87: \end{center}

88: \setlength{\baselineskip}{1.5\baselineskip}

89: 

90: If the number of runs in a (mixed-level) orthogonal array of strength 2 is

91: specified, what numbers of levels and factors are possible?

92: The collection of possible sets of parameters for orthogonal arrays with $N$

93: runs has a natural lattice structure,

94: induced by the ``expansive replacement'' construction method.

95: In particular the dual atoms in this lattice are the most important

96: parameter sets, since any other parameter set for an $N$-run orthogonal array can be constructed from them.

97: To get a sense for the number of dual atoms, and to begin to understand the lattice as a function

98: of $N$, we investigate the height and the size of the lattice.

99: It is shown that the height is at most $\lfloor c(N-1) \rfloor$,

100: where $c= 1.4039 \ldots$,

101: and that there is an infinite sequence of values of $N$ for which

102: this bound is attained.

103: On the other hand, the number of nodes in the lattice is bounded above

104: by a superpolynomial function of $N$ (and superpolynomial growth does occur for certain

105: sequences of values of $N$).

106: Using a new construction based on ``mixed spreads'', all parameter sets with 64 runs are determined.

107: Four of these 64-run orthogonal arrays appear to be new.

108: 

109: 

110: \clearpage

111: \section{Introduction}

112: Although mixed-level (or asymmetrical) orthogonal arrays have been 

113: the subject of a number of papers in recent years (see Chapter 9 of

114: %\cite{HSS}

115: Hedayat, Sloane and Stufken, 1999,

116: for references),

117: it seems fair to say that we know much less about them than about fixed-level orthogonal arrays 

118: (in which all factors have the same number of levels).

119: For example, there is no analogue for mixed orthogonal arrays of one of the

120: most powerful construction methods for fixed-level arrays, that based on linear codes (see Chapters 4 and 5 of Hedayat, Sloane and Stufken, 1999).

121: 

122: Again, there are many instances where the linear programming bound for fixed-level orthogonal arrays gives

123: the correct answer for the minimal number of runs needed for a specified

124: number of factors.

125: There {\em is} a linear programming bound for mixed arrays (Sloane

126: and Stufken, 1996), but it is less effective than in the fixed-level case --- it

127: ignores too much of the combinatorial nature of the problem

128: (especially when the levels involve more than one prime number),

129: and, though generally stronger than the Rao bound, 

130: does not give correct answers as often as in the fixed-level case.

131: 

132: A mixed orthogonal array

133: $OA(N, s_1^{k_1} s_2^{k_2} \cdots s_v^{k_v}, t)$ is an array of size $N \times k$,

134: where $k= k_1 + k_2 + \cdots + k_v$ is the total number of factors,

135: in which the first $k_1$ columns have symbols from $\{0,1, \ldots, s_1 -1\}$,

136: the next $k_2$ columns have symbols from $\{0,1, \ldots, s_2 -1\}$, and so on,

137: with the property that in any $N \times t$ subarray every possible  $t$-tuple of symbols occurs

138: an equal number of times as a row.

139: We usually assume $2 \le s_1 < s_2 < \cdots$ and all $k_i \ge 1$.

140: Except in Section 5, only arrays of strength 2 will be considered, and

141: we will usually omit $t$ from

142: the symbol for the array.

143: 

144: We refer to $(N, s_1^{k_1} s_2^{k_2} \ldots )$ as the

145: {\em parameter set} for an $OA(N,s_1^{k_1} s_2^{k_2} \ldots )$.

146: We also allow the parameter set $(N, 1^1 )$, corresponding

147: to the trivial array consisting of a single column of $N$ 0's.

148: In this paper we consider the question:

149: if $N$ is specified, how many different parameter sets are possible?

150: 

151: Given an array $A = OA (N, s_1^{k_1} s_2^{k_2} \ldots )$, other $N$-run

152: arrays can be obtained from it by the

153: {\em expansive replacement method}.

154: Let $S$ be one of the $s_i$ occurring in $A$, and suppose $B$ is an $OA(S,t_1^{l_1} t_2^{l_2} \ldots )$.

155: The expansive replacement method replaces a single column of $A$ at $S$

156: levels by the rows of $B$.

157: For example, if $A= OA( 16, 2^3 4^4 )$ and $B= OA(4, 2^3 )$,

158: we obtain an $OA(16, 2^6 4^3 )$.

159: If $B$ is a trivial array $OA(S, 1^1 )$, we are simply deleting one of the $S$-level factors from $A$.

160: E.g. taking $S=2$,

161: an $OA( 24, 2^{20} 4^1 )$ trivially produces an $OA(24, 2^{19} 4^1 )$.

162: The expansive replacement method also includes replacing a factor at $s$ levels by a factor at $s'$ levels, if $s'$ divides $s$.

163: For further details about the expansive

164: replacement method see Chapter 9 of Hedayat, Sloane and Stufken, 1999.

165: 

166: Let $A$ and $B$ be parameter sets for orthogonal arrays with $N$ runs.

167: We say that $B$ is {\em dominated by} $A$

168: if an orthogonal array with parameter set $B$ can be obtained from an orthogonal array with

169: parameter set $A$ by a sequence of expansive replacements.

170: 

171: Using ``dominance'' as the relation, the parameter sets for orthogonal arrays with $N$ runs form a partially ordered set, which we denote by $\Lambda_N$

172: (Hedayat, Sloane and Stufken, 1999 p. 335).

173: 

174: $\Lambda_N$ has a unique maximal element $(N, N^1 )$ (corresponding to the

175: trivial array with one factor at $N$ levels) and a unique minimal element

176: $(N, 1^1 )$.

177: It is straightforward to verify that meet $(\wedge)$ and join $(\vee )$ are

178: well-defined for this relation (we omit the proof), so $\Lambda_N$ is

179: in fact a {\em lattice} (cf. Welsh, 1976;

180: Trotter, 1995).

181: 

182: If an $OA(N, s_1^{k_1} s_2^{k_2} \ldots )$ exists,

183: then necessarily we must have:

184: \begin{itemize}

185: \item[(C1)]

186: $s_i$ divides $N$, for all $i$,

187: \item[(C2)]

188: $s_i^2$ divides $N$, if $k_i \ge 2$,

189: \item[(C3)]

190: $s_i s_j$ divides $N$, if $i \neq j$,

191: \item[(C4)]

192: the Rao bound holds:

193: \beql{EqR}

194: N-1 \ge k_1 (s_1 -1) + k_2 (s_2 -1) + \cdots \,,

195: \eeq

196: \item[(C5)]

197: the linear programming bound holds (see Sloane and Stufken, 1996).

198: \end{itemize}

199: 

200: These conditions are certainly not sufficient for an array to exist,

201: and it appears to be difficult to test if an orthogonal array does exist

202: with a putative parameter set satisfying (C1)--(C5).

203: A further difficulty is that in order to construct $\Lambda_N$ it is necessary

204: to know $\Lambda_d$ for all proper divisors $d$ of $N$.

205: 

206: To avoid these difficulties we define a second lattice, the

207: {\em idealized lattice}

208: $\Lambda'_N$:

209: this has as nodes all putative parameter sets satisfying

210: (C1) to (C4), with the dominance relation as before,

211: except that in the expansive replacement method we may now make use of any

212: of the nodes of any $\Lambda'_d$ for $d$ dividing $N$.

213: 

214: Constructing $\Lambda'_N$ is much easier than constructing

215: $\Lambda_N$, since essentially all we need to do is enumerate the solutions

216: to \eqn{EqR}.

217: Of course $\Lambda_N$ is a sublattice of $\Lambda'_N$.

218: 

219: To avoid having to repeat the adjective ``putative'',

220: from now on we will use ``parameter set'' to mean any symbol $(N, s_1^{k_1} s_2^{k_2} \ldots )$ satisfying conditions (C1) to (C4).

221: The parameter sets are precisely the nodes of $\Lambda'_N$.

222: If a parameter set is also a node of $\Lambda_N$ then it is implied that an

223: $OA(N, s_1^{k_1} s_2^{k_2} \ldots )$ does exist, i.e. that the parameter set

224: is realized by an orthogonal array.

225: 

226: It is convenient to represent $\Lambda_N$ and $\Lambda'_N$ by their Hasse diagrams (cf. Welsh, 1976, p. 45).

227: These diagrams are drawn ``from the bottom up'',

228: with $(N, 1^1)$ as the root node at the bottom (Figure \ref{fg1} shows $\La_{12}$ and $\La'_{12}$).

229: The {\em height} of a parameter set is the number of edges in the longest path

230: from that node to the root.

231: A node of height $i$ appears on the $i$th level of the diagram.

232: The height of the maximal element $(N, N^1)$ will be denoted by $ht(N)$.

233: 

234: 

235: 

236: \begin{figure}[htp]

237: \begin{center}

238: \input fg1.pstex_t

239: \end{center}

240: \caption{(Left) Lattice $\Lambda_{12}$, showing all (true)

241: parameter sets for 12-run orthogonal arrays.

242: There are 23 nodes, four dual atoms (circled) and the height is 12.

243: (Right) Idealized lattice $\Lambda'_{12}$, showing all (putative)

244: parameter sets satisfying conditions (C1)--(C4).

245: No arrays exist for the nine nodes marked $\times$.

246: There are 32 nodes, four dual atoms (circled) and again the height is 12.}

247: \label{fg1}

248: \end{figure}

249: 

250: The {\em atoms} in $\La_N$ (those nodes just above the root) are

251: precisely the parameter sets $(N, p^1)$ for the primes $p$ dividing $N$.

252: 

253: The {\em dual atoms} in $\La_N$ (those nodes just below the maximal element) are

254: especially interesting, since they dominate all other parameter sets.

255: 

256: We can now state our main results.

257: \begin{theorem}\label{th1}

258: (i) For all $N$,

259: \beql{Eq2}

260: ht(N) \le \lfloor c (N-1) \rfloor ~,

261: \eeq

262: where

263: \beql{Eq2a}

264: c ~=~ \sum_{i=0}^{\infty} \frac{1}{2^{2^i}-1} ~=~ 1.4039 \ldots ~.

265: \eeq

266: 

267: (ii) If $N= 2^{2^m}$ $(m \ge 0)$ then $ht(N) = \lfloor c (N-1) \rfloor$.

268: \end{theorem}

269: 

270: \noindent

271: 

272: Let $T(N)$ (resp. $T' (N)$) denote the total number of nodes in $\La_N$ (resp. $\La'_N$).

273: 

274: \begin{theorem}\label{th2}

275: If $N=2^{n}$,

276: \beql{Eq5}

277: \frac{1}{4} ( \log_2 N)^2 (1+o(1))

278: ~ \le~  \log_2 T(N) ~ \le~  \frac{3}{8}

279: (\log_2 N)^2 (1+o(1)) \,.

280: \eeq

281: \end{theorem}

282: 

283: \begin{theorem}\label{th3}

284: There is a constant $c_1$ such that for all $N$,

285: \beql{Eq3}

286: \ln \ln T(N) ~ \le ~ c_1 ~ \frac{\ln N}{\ln \ln N}  ~ (1+ o(1)) \,.

287: \eeq

288: \end{theorem}

289: 

290: \paragraph{Remarks.}

291: (i) The bounds in \eqn{Eq5} and \eqn{Eq3} also apply to $T' (N)$.

292: 

293: (ii) Theorem \ref{th2} shows that when $N=2^{n}$, $T(N)$ grows very roughly

294: like $N^{a \log_2 N}$, for some constant $a$ between

295: $\frac{1}{4}$ and $\frac{3}{8}$.

296: This is a

297: ``superpolynomial'' function of $N$, meaning that it grows faster than any polynomial in $N$.

298: 

299: (iii) It appears (although we have not proved this) that the upper bound

300: in \eqn{Eq3} can be achieved by taking $N$ to be a certain product of powers of the first $m$ primes,

301: where $m$ is about

302: $$\frac{1}{2e} ~ \frac{\ln N}{\ln \ln N}$$

303: (see Section 7).

304: In other words, it appears that there is an infinite sequence of values

305: of $N$ for which $T(N)$ grows {\em very} roughly like

306: $$\exp (N^{c_2 / \ln \ln N }) ~,$$

307: where $c_2$ is a constant.

308: This is again a superpolynomial function of $N$, and is now

309: close to being an exponential function, since $\ln \ln N$ grows slowly.

310: 

311: The above discussion has shown that there is an infinite sequence of values of $N$ for which 

312: the number of nodes in $\La_N$ grows superpolynomially, while the

313: height of $\La_N$ grows at most linearly.

314: It follows that the size of the largest antichain must also grow superpolynomially.

315: The data in Table \ref{ta3} suggest the following conjecture.

316: 

317: \paragraph{Conjecture.}

318: {\em There is an infinite sequence of values of $N$ for which the number

319: of dual atoms grows superpolynomially in $N$.}

320: 

321: In fact it seems likely that if $N =2^{n}$, a lower bound of the form

322: in \eqn{Eq5} (possibly with a different constant) applies to the

323: logarithm of the number of dual atoms,

324: and that for some sequence of values of $N$ a lower bound similar

325: to the upper bound on the right-hand side of \eqn{Eq3} will hold.

326: However, at present these are only conjectures.

327: 

328: In order to construct the orthogonal arrays needed to establish the lower bound in

329: Theorem \ref{th2} we make use of what we call ``mixed spreads'',

330: generalizing the notions of ``spread'' and ``partial spread'' from projective

331: geometry.

332: Arrays that can be constructed in this way we call ``geometric''.

333: Many familiar examples of orthogonal arrays, for example arrays constructed from linear codes,

334: are geometric.

335: The construction is not restricted to strength 2 (and is one of the few general

336: constructions we know of for mixed arrays of strength greater than 2).

337: The construction will be described in Section 5.

338: 

339: In Section 6 we use this construction to determine the lattice

340: $\Lambda_{64}$, and in doing so we find tight arrays with parameter sets

341: $$(64,2^5 4^{17} 8^1), (64, 4^{14} 8^3) ,

342: (64, 2^5 4^{10} 8^4), (64, 4^7 8^6) ,

343: $$

344: which appear to be new.

345: 

346: 

347: When studying parameter sets of putative orthogonal arrays with $N$ runs,

348: it is convenient to be able to say that if the number of degrees of freedom

349: of the parameter set $(N, s_1^{k_1} s_2^{k_2} \ldots )$, that is,

350: \beql{Eq61a}

351: k_1 (s_1 -1) + k_2 (s_2 -1) + \cdots ~,

352: \eeq

353: is small compared with $N-1$, then an orthogonal array certainly exists.

354: 

355: To make this precise, we define the {\em threshold function}

356: $B(N)$ to be the maximum number $b$ such that every

357: parameter set (satisfying conditions (C1) to (C4))

358: with at most $b$ degrees of freedom is realized by an orthogonal array, but some parameter set (again

359: satisfying (C1) to (C4)) with $b+1$ degrees of freedom is not realized.

360: If every parameter set satisfying (C1) to (C4) is realized, we set $B(N) = N-1$.

361: 

362: Figure \ref{fg1} shows that $B(12) =6$, since there is no $OA(12 , 2^5 3^1)$, but

363: every parameter set with at most 6 degrees of freedom is realized.

364: 

365: We are not aware of any earlier investigations of $B(N)$.

366: 

367: \begin{theorem}\label{th4}

368: If $N$ is a power of a prime then

369: $$N^{3/4} \le B(N) ~.$$

370: \end{theorem}

371: 

372: \noindent

373: In words, if the number of degrees of freedom in the parameter set

374: does not exceed $N^{3/4}$, then an orthogonal array exists.

375: This is certainly weak, but is enough to establish the lower bound of Theorem \ref{th2}.

376: It would be nice to have more precise estimates for $B(N)$.

377: 

378: A final remark.

379: We {\em could} have considered the partially ordered set whose nodes are all the

380: inequivalent orthogonal arrays with $N$ runs, rather than just their parameter sets.

381: However, the number of nodes then becomes unmanageably large, even for small values of $N$

382: (furthermore, it appears that ``meet'' and

383: ``join'' are no longer well-defined,

384: and so in general this partially ordered set would not be a lattice).

385: 

386: Consider $N=28$, for example.

387: Using Kimura's (1994a, 1994b) enumeration of the Hadamard

388: matrices of order 28, we have calculated\footnote{Using the method described on page 165 of Hedayat, Sloane and Stufken (1999).} that

389: there are precisely 7570 inequivalent $OA(28, 2^{27})$'s.

390: This would be merely a lower bound on the number of dual atoms.

391: On the other hand we know (see Table \ref{ta1}) that $\La_{28}$ has precisely four dual atoms,

392: between 47 and 55 nodes, and height 28.

393: 

394: \begin{figure}[htb]

395: \begin{center}

396: \input fg2.pstex_t

397: \end{center}

398: \caption{(a)~$\Lambda_p$ and (b)~$\Lambda_6$.}

399: \label{fg2}

400: \end{figure}

401: 

402: \section{Examples of the lattices $\Lambda_N$ and $\Lambda'_N$}

403: There are a few general cases when we can describe $\La_N$ explicitly

404: (and for which $\La'_N$ is the same as $\La_N$).

405: 

406: If $N=p$ is a prime then $\La_N = \La'_N$ has two nodes, one dual

407: atom and height 1, as shown in Fig. \ref{fg2}(a) (dual atoms are circled).

408: 

409: If $N=pq$ is the product of two distinct primes, $\La_N = \La'_N$ has five nodes,

410: one dual atom and height 3.

411: $\La_6$ is shown in Fig. \ref{fg2}(b).

412: 

413: More generally, if $N$ is the product of $u \ge 2$ distinct primes,

414: it is not difficult to show that $\La_N = \La'_N$ has $2^{u-1} -1$ dual atoms,

415: height $2u-1$, and $\beta_{u+1}$ nodes, where

416: $$\{ \beta_1, \beta_2, \beta_3, \ldots \} = \{

417: 1,2,5,15,52, \ldots \}

418: $$

419: are the Bell numbers (see Sequence \htmladdnormallink{A110}{http://www.research.att.com/cgi-bin/access.cgi/as/njas/sequences/eisA.cgi?Anum=000110} of Sloane, 1999).

420: Figure \ref{fg3} shows $\La_{30}$,

421: illustrating the case $u=3$.

422: 

423: \begin{figure}[htb]

424: \begin{center}

425: \input fg3.pstex_t

426: \end{center}

427: \caption{The lattice $\Lambda_{30}$:

428: there are $\beta_4 = 15$ nodes, three dual atoms (circled) and the height is 5.}

429: \label{fg3}

430: \end{figure}

431: 

432: If $N=p^2$ is the square of a prime, $\La_N = \La'_N$ has a single dual

433: atom, $OA(p^2, p^{p+1} )$, has height $p+2$ and contains $p+3$ nodes.

434: If $N=p^3$ ($p$ prime), $\La_N = \La'_N$ also has a single dual atom,

435: $OA(p^3, (p)^{p^2} (p^2)^1 )$, has height $p^2 + p+3$ and

436: contains

437: $2p^2 + p+4$ nodes.

438: $\La_8$ is shown in Fig. \ref{fg4}.

439: If $N= p^4$ ($p$ prime), $\La_N = \La'_N$ has two dual atoms,

440: $OA(p^4, (p^2)^{p^2 +1} )$ and

441: $OA(p^4 , (p)^{p^3} (p^3)^1)$, has height

442: $p^3 + 2p^2 + p+3$ and contains

443: $$\frac{1}{2} (p^5 + p^4 + 5p^3 + 5p^2 + 2p + 10 )$$

444: nodes.

445: 

446: 

447: For all values of $N$ mentioned so far in this section, the threshold function $B(N) = N-1$.

448: 

449: \begin{figure}[htb]

450: \begin{center}

451: \input fg4.pstex_t

452: \end{center}

453: \caption{The lattice $\Lambda_8$: 14 nodes, one dual atom, height 9.}

454: \label{fg4}

455: \end{figure}

456: 

457: If $N$ is not of one of the above forms then it seems necessary to consider each case

458: individually.

459: Table \ref{ta1} summarizes the properties of $\La_N$ for some small values of $N$.

460: Here $A(N)$ denotes the number of dual atoms in $\La_N$.

461: Most of the entries in this table can be deduced from the tables in Chapter 12 

462: of Hedayat, Sloane and Stufken (1999),

463: except for $N=32$ and 64, which are discussed in Section 6.

464: \begin{table}[htb]

465: \caption{For the lattice $\La_N$ of parameter sets of orthogonal arrays with $N$ runs,

466: the table gives the number of dual atoms $A(N)$, the height

467: $ht(N)$, the total number of nodes $T(N)$ and the threshold function $B(N)$.}

468: $$

469: \begin{array}{rrrrr}

470: \multicolumn{1}{c}{N} &

471: \multicolumn{1}{c}{A(N)} &

472: \multicolumn{1}{c}{ht(N)} &

473: \multicolumn{1}{c}{T(N)} &

474: \multicolumn{1}{c}{B(N)} \\ \hline

475: 1~~ & 0\quad & 0\quad & 1\quad & 0\quad \\

476: 2~~ & 1\quad & 1\quad & 2\quad & 1\quad \\

477: 3~~ & 1\quad & 1\quad & 2\quad & 2\quad \\

478: 4~~ & 1\quad & 4\quad & 5\quad & 3\quad \\

479: 5~~ & 1\quad & 1\quad & 2\quad & 4\quad \\

480: 6~~ & 1\quad & 3\quad & 5\quad & 5\quad \\

481: 7~~ & 1\quad & 1\quad & 2\quad & 6\quad \\

482: 8~~ & 1\quad & 9\quad & 14\quad & 7\quad \\

483: 9~~ & 1\quad & 5\quad & 6\quad & 8\quad \\

484: 10~~ & 1\quad & 3\quad & 5\quad & 9\quad \\

485: 11~~ & 1\quad & 1\quad & 2\quad & 10\quad \\

486: 12~~ & 4\quad & 12\quad & 23\quad & 6\quad \\

487: 13~~ & 1\quad & 1\quad & 2\quad & 12\quad \\

488: 14~~ & 1\quad & 3\quad & 5\quad & 13\quad \\

489: 15~~ & 1\quad & 3\quad & 5\quad & 14\quad \\

490: 16~~ & 2\quad & 21\quad & 61\quad & 15\quad \\

491: 17~~ & 1\quad & 1\quad & 2\quad & 16\quad \\

492: 18~~ & 2\quad & 10\quad & 26\quad & 15\quad \\

493: \end{array}

494: $$

495: \label{ta1}

496: \end{table}

497: 

498: \begin{table}[htb]

499: $$

500: \begin{array}{ccccc}

501: \multicolumn{5}{c}{\mbox{Table 1 (cont.)}} \\ [+.2in]

502: \multicolumn{1}{c}{N} &

503: \multicolumn{1}{c}{A(N)} &

504: \multicolumn{1}{c}{ht(N)} &

505: \multicolumn{1}{c}{T(N)} &

506: \multicolumn{1}{c}{B(N)} \\ \hline

507: 19 & 1 & 1 & 2 & 18 \\

508: 20 & 4 & 20 & 35 & 11 \\

509: 21 & 1 & 3 & 5 & 20 \\

510: 22 & 1 & 3 & 5 & 21 \\

511: 23 & 1 & 1 & 2 & 22 \\

512: 24 & 4-7 & 25 & 119-133 & 18-22 \\

513: 25 & 1 & 7 & 8 & 24 \\

514: 26 & 1 & 3 & 5 & 25 \\

515: 27 & 1 & 15 & 25 & 26 \\

516: 28 & 4 & 28 & 47-55 & 15 \\

517: 29 & 1 & 1 & 2 & 28 \\

518: 30 & 3 & 5 & 15 & 29 \\

519: 31 & 1 & 1 & 2 & 30 \\

520: 32 & 2 & 42 & 320 & 29 \\

521: 33 & 1 & 3 & 5 & 32 \\

522: 34 & 1 & 3 & 5 & 33 \\

523: 35 & 1 & 3 & 5 & 34 \\

524: \ldots & \ldots & \ldots & \ldots & \ldots \\

525: 64 & 7 & 86 & 3037 & 57 \\

526: \ldots & \ldots & \ldots & \ldots & \ldots 

527: \end{array}

528: $$

529: \end{table}

530: 

531: $N=24$ is the smallest case when we do not know the complete lattice

532: $\La_N$.

533: In $\La_{24}$ the maximal value of $k$ that occurs in each of the parameter

534: sets $2^k 3^1$, $2^k 3^1 4^1$,

535: $2^k 4^1 6^1$ and $2^k 6^1$ is presently unknown.

536: For $2^k 3^1$, for example, it is known only that an

537: $OA(24, 2^{16} 3^1 )$ exists, and an $OA(24, 2^{21} 3^1)$ is impossible by the linear programming bound.

538: The number of dual atoms is in the range 4 to 7.

539: It is possible to show that the height of $\La_{24}$ is 25, however:

540: no chain can be longer than

541: $$

542: \mbox{$24^1$ --- $2^{20} 4^1$ --- $2^{23}$ --- $2^{22}$ --- $2^{21}$ --- $\cdots$ --- $2^1$ --- $1^1$} \,.

543: $$

544: 

545: We also do not know $\La_N$ for $N=28$, 36, $\ldots$.

546: 

547: The four sequences in Table \ref{ta1} are Sequences

548: \htmladdnormallink{A39927}{http://www.research.att.com/cgi-bin/access.cgi/as/njas/sequences/eisA.cgi?Anum=039927},

549: \htmladdnormallink{A39930}{http://www.research.att.com/cgi-bin/access.cgi/as/njas/sequences/eisA.cgi?Anum=039930},

550: \htmladdnormallink{A39931}{http://www.research.att.com/cgi-bin/access.cgi/as/njas/sequences/eisA.cgi?Anum=039931}

551: and

552: \htmladdnormallink{A48893}{http://www.research.att.com/cgi-bin/access.cgi/as/njas/sequences/eisA.cgi?Anum=048893}

553: of Sloane (1999).

554: The entries in that database will be updated as further values are determined.

555: 

556: We end this section with a remark about the nature of $\La_N$ as a lattice.

557: Since not all maximal chains between two nodes need have the same length (see Fig. \ref{fg1}),

558: $\La_N$ does not in general satisfy the Jordan-Dedekind chain condition (cf. Welsh, 1976).

559: It follows that $\La_N$ is in general not distributive, not semimodular, 

560: nor is it the lattice of a matroid.

561: 

562: \section{The maximum height of $\La_N$}

563: In this section we give the proof of Theorem \ref{th1}.

564: 

565: Let $\sigma$ denote a specification $s_1^{k_1} s_2^{k_2} \ldots$ of factors at various levels,

566: leaving the number of runs unspecified.

567: Given $\sigma$, there is a smallest number of runs,

568: $N_0$ say, for which an $OA(N_0, \sigma )$ exists.

569: Let $h$ be the height of the parameter set $(N_0, \sigma )$ in $\La_{N_0}$.

570: Then if the parameter set $(N, \sigma )$ occurs in any other lattice

571: $\La_N$, it also has height $h$.

572: (E.g. the specification $\sigma = 6^1$ has height 3 in each

573: of Figs. 1, 2(b) and 3.)

574: We may therefore

575: define $ht(\sigma)$ to be $h$, independently of the number of runs.

576: 

577: The height of $\La_N$, $ht(N)$, as defined in Section 1 coincides with

578: $ht(\sigma )$ when $\sigma = N^1$.

579: The height function has the following additivity property.

580: \begin{lemma}\label{le1}

581: \beql{Eq10}

582: ht(s_1^{k_1} s_2^{k_2} \ldots ) = k_1 ht( s_1) + k_2 ht(s_2) + \cdots ~.

583: \eeq

584: \end{lemma}

585: 

586: \paragraph{Proof.}

587: If there is a single factor on the left-hand side, say $s_1 =s$,

588: $k_1 =1$, then \eqn{Eq10} just says that $ht(s^1 ) = ht(s)$, repeating the assertion made above.

589: Otherwise, more than one factor occurs in the specification $\sigma$ on the left-hand side of \eqn{Eq10}.

590: Suppose the parameter set $(N, \sigma )$ occurs as a node in some $\La_N$.

591: The portion of $\La_N$ consisting of this node and all nodes

592: dominated by it is the product in an obvious sense of $k_1$

593: copies of $\La_{s_1}$, $k_2$ copies of $\La_{s_2}$, etc.

594: The height of $(N, \sigma )$ is the sum of the heights of all these

595: sublattices, and \eqn{Eq10} follows.~~~$\bsq$

596: 

597: Lemma \ref{le1} reduces the calculation of $ht (s_1^{k_1} s_2^{k_2} \ldots )$ to the calculation of the values of $ht(s_1)$, $ht(s_2) , \ldots$.

598: To determine the latter we must consider exactly which orthogonal arrays exist with a specified number of runs.

599: For this we can apply the following lemma.

600: 

601: \begin{lemma}\label{le2}

602: \beql{Eq11}

603: ht(N) = 1+ \max \sum_i k_i ht(s_i) ~,

604: \eeq

605: where the maximum is taken over all parameter sets 

606: $(N, s_1^{k_1} s_2^{k_2} \ldots ) \neq (N,N^1)$ for which an orthogonal array exists.

607: \end{lemma}

608: 

609: \paragraph{Proof.}

610: The height of $\La_N$ is one more than the maximal height among the dual atoms.

611: \eqn{Eq11} follows by applying Lemma \ref{le1} to the parameter set of such a dual atom.~~~$\bsq$

612: 

613: We can now use linear programming to obtain an upper bound on $ht(N)$, by maximizing

614: \beql{Eq12}

615: 1+ \sum_i k_i ht (s_i)

616: \eeq

617: over all choices of $s_1$, $k_1$, $s_2$, $k_2 , \ldots$ that satisfy (C1) to (C4).

618: 

619: We first consider the case when $N=2^n$ for some $n$.

620: 

621: The case $N=64$ will illustrate the method.

622: If there is a factor $32^1$ then linear programming shows that \eqn{Eq12}

623: is maximized by $2^{32} \, 32^1$,

624: giving height 75.

625: If there is a factor $16^1$ then there is a unique parameter set that maximizes

626: \eqn{Eq12}, $4^{16} 16^1$, giving height 86.

627: Otherwise, if only factors $2^{k_1} 4^{k_2} 8^{k_3}$ occur, the height does not

628: exceed 85.

629: Since an $OA(64, 4^{16} 16^1 )$ exists, we conclude that $ht( 64) = 86$.

630: We will return to the case $N=64$ in Section 6.

631: 

632: In this way we obtain the values of $ht(N)$, $N=2^n$ shown in Table \ref{ta2}.

633: \begin{table}[htb]

634: \caption{Height of $\La_N$ for $N=2^n$.}

635: $$

636: \begin{array}{cccccc}

637: n & 1 & 2 & 3 & 4 & 5 \\

638: N=2^n & 2 & 4 & 8 & 16 & 32 \\

639: ht(N) & 1 & 4 & 9 & 21 & 42 \\

640: ht(N) /(N-1) & 1 & 1.3333 \ldots & 1.2857 \ldots & 1.4 & 1.3548 \ldots \\ [+.5in]

641: n & 6 & 7 & 8 & 9 & 10 \\

642: N=2^n & 64 & 128 & 256 & 512 & 1024 \\

643: ht(N) & 86 & 171 & 358 & 715 & 1431 \\

644: ht(N) / (N-1) & 1.3650 \ldots & 1.3465 \ldots & 1.4039 \ldots & 1.3992 \ldots & 1.3988 \ldots

645: \end{array}

646: $$

647: \label{ta2}

648: \end{table}

649: 

650: Consider the general problem of maximizing \eqn{Eq12} for $N=2^n$.

651: Comparing \eqn{EqR} and \eqn{Eq12}, we see that an $s$-level

652: factor contributes $ht(s)$ to the height but uses up $s-1$ degrees of freedom.

653: If we ignore the constraints of integrality then the value of the expression

654: in \eqn{Eq12} would be maximized

655: by a term $s^k$ where $s$ is chosen to maximize $ht(s) / (s-1)$.

656: This suggests that we should investigate this quantity in order to prove

657: Theorem \ref{th1}.

658: The data in

659: Table \ref{ta2} suggest that the ratio $ht(s) / (s-1)$ is maximized if $s$ is of the form

660: $2^{2^i}$ and is as large as possible.

661: That this is indeed so is established by the next three lemmas.

662: 

663: \begin{lemma}\label{L200}

664: Given positive real numbers $\af_r$, $\beta_r$ $(r=1,\ldots, m )$, $\gamma$,

665: the maximal value of

666: $$

667: \sum_{r=1}^m \alpha_r x_r

668: $$

669: subject to the constraints

670: $$

671: \begin{array}{c}

672: \displaystyle\sum_{r=1}^m \beta_r x_r = \gamma ~, \\[+.2in]

673: x_r \ge 0 \quad (r=1, \ldots, m )

674: \end{array}

675: $$

676: is

677: $$\gamma \max_{r=1 \ldots m} \frac{\af_r}{\beta_r} ~.

678: $$

679: \end{lemma}

680: 

681: We omit the straightforward proof.

682: \begin{lemma}\label{L201}

683: Given positive real numbers $\alpha_r$, $\beta_r$ $(r=1, \ldots, m+n )$,

684: $\gamma$, $\gamma'$ with $\gamma' \le \gamma$, the maximal

685: value of

686: \beql{Eq202}

687: \sum_{r=1}^{m+n} \alpha_r x_r

688: \eeq

689: subject to the constraints

690: \beql{Eq203A}

691: \sum_{r=1}^{m+n} \beta_r x_r = \gamma ~, 

692: \eeq

693: \beql{Eq203B}

694: \sum_{r=m+1}^{m+n} \beta_r x_r \le \gamma' ~, 

695: \eeq

696: \beql{Eq203C}

697: x_r \ge 0 \quad (r=1, \ldots, m+n)

698: \eeq

699: is

700: \beql{Eq204}

701: \max \left\{

702: \gamma \max_{r=1 \ldots m} \frac{\af_r}{\beta_r} ,~

703: (\gamma - \gamma' ) \max_{r=1 \ldots m}

704: \frac{\af_r}{\beta_r} +

705: \gamma' \max_{r = m+1 \ldots m+n} \frac{\af_r}{\beta_r} \right\}\,.

706: \eeq

707: The maximum is given by the first expression if and only if

708: $$

709: \max_{r=1 \ldots m} \frac{\af_r}{\beta_r} \ge

710: \max_{r= m+1 \ldots m+n} \frac{\af_r}{\beta_r} ~.

711: $$

712: \end{lemma}

713: 

714: \paragraph{Proof.}

715: Let $\sigma$, $0 \le \sigma \le \gamma'$, denote the value of the left-hand

716: side of \eqn{Eq203B}.

717: Then by Lemma \ref{L200} the maximal value of the sum in \eqn{Eq202} is equal to

718: $$(\gamma - \sigma ) \max_{r=1 \ldots m}

719: \frac{\af_r}{\beta_r} + \sigma

720: \max_{r= m+1 \ldots m+n} \frac{\af_r}{\beta_r} \,.

721: $$

722: This is a linear function of $\sigma$ and so its maximal value is taken

723: at one of the two endpoints, leading to \eqn{Eq204}.~~~$\bsq$

724: 

725: We can now give an upper bound on the height $ht(N)$ for $N=2^n$.

726: Let

727: $$\rho_n = \frac{ht(N)}{2^n -1} , \quad N=2^n ~.$$

728: 

729: \begin{lemma}\label{L205}

730: $$\rho_n < c \quad\mbox{for all}\quad n ~,$$

731: where $c$ is the constant

732: $$\sum_{i=0}^\infty \frac{1}{2^{2^i} -1} = 1.4039 \ldots$$

733: \end{lemma}

734: 

735: \paragraph{Proof.}

736: We use the induction hypothesis that, for $m \ge 0$,

737: \beql{Eq206}

738: \rho_{2^m} > \rho_x \quad\mbox{whenever}\quad

739: 1 \le x < 2^{m+1}, ~x \neq 2^m \,.

740: \eeq

741: This is trivially true when $m=0$.

742: We first compute $\rho_{2^m}$.

743: From the Rao-Hamming construction, for $N=2^{2^m}$, an $OA(N, (\sqrt{N})^{\sqrt{N} +1} )$ always exists, and we find that

744: $$

745: \rho_{2^m} \ge \rho_{2^{m-1}} + \frac{1}{2^{2^m} -1} ~.

746: $$

747: On the other hand, we obtain an upper bound on $\rho_{2^m}$ from the linear program:

748: choose nonnegative integers $k_1, k_2, \ldots$ so as to maximize

749: $$

750: 1 + \sum_{r=1}^{2^m -1} ht (2^r) k_{r}

751: $$

752: subject to the constraint

753: $$

754: \sum_{r=1}^{2^m -1} (2^r -1) k_{r} = 2^{2^m} -1 ~.

755: $$

756: From Lemmas \ref{le2} and \ref{L200} we have

757: \begin{eqnarray*}

758: \rho_{2^m} & \le & \frac{1}{2^{2^m} -1} + \max_{r=1 \ldots 2^m -1} \rho_r \\

759: & = & \frac{1}{2^{2^m} -1} + \rho_{2^{m-1}} ~.

760: \end{eqnarray*}

761: Since the two bounds agree,

762: \beql{Eq209}

763: \rho_{2^m} = \rho_{2^{m-1}} + \frac{1}{2^{2^m} -1} ~.

764: \eeq

765: 

766: 

767: %Provided $N=2^{2^m}$ is also of this form then we may use such an $s$,

768: %taking $s=2^{2^{m-1}} = \sqrt{N}$ and $k = \sqrt{N} +1$.

769: %In this way we obtain the following values for the height:

770: %$$

771: %\begin{array}{ccccccc}

772: %m & 0 & 1 & 2 & 3 & 4 & \cdots \\

773: %N= 2^{2^m} & 2 & 4 & 16 & 256 & 65536 & \cdots \\

774: %ht(N) & 1 & 4 & 21 & 358 & 92007 & \cdots \\

775: %\end{array}

776: %$$

777: %These values

778: %of $ht(N)$, $N=2^{2^m}$, can be achieved for $m=0,1,2, \ldots$,

779: %because an $OA(N, (\sqrt{N})^{\sqrt{N} +1} )$ exists

780: %by the Rao-Hamming construction.

781: %Writing $\af_m$ for $ht(N)$, we have $\af_0 =1$ and

782: %$$\af_m = (2^{2^{m-1}} +1) \af_{m-1} + 1, ~~m \ge 1 .$$

783: 

784: We now complete the proof of the induction step.

785: For $1 \le x < 2^m$, we have

786: $$\rho_{2^m} > \rho_{2^{m-1}} \ge \rho_x ~,$$

787: as required.

788: Suppose $2^m < x < 2^{m+1}$.

789: Then $\rho_x$ is upper-bounded by the solution to the linear program:

790: maximize

791: $$\left(1 + \sum_{r=1}^{x-1} ht (2^r ) k_{r} \right) \Bigl/ (2^x -1 )$$

792: subject to the constraints

793: \begin{eqnarray*}

794: \sum_{r=1}^{x-1} (2^r -1) k_{r} & = & 2^x -1 ~,

795: \\

796: \sum_{r=2^m}^{x-1} (2^r -1) k_{r} & \le & 2^{x-1} -1 ~.

797: \end{eqnarray*}

798: (The second constraint is implied by the requirement that there can be

799: at most one factor which has more levels than the square root of the number of runs.)

800: By Lemma \ref{L201} and induction on $x$, we obtain

801: \begin{eqnarray*}

802: \rho_x & \le & \frac{1}{2^x -1}

803: ( 1+ 2^{x-1} \rho_{2^{m-1}} + (2^{x-1} -1) \rho_{2^m} ) \\

804: & = & \rho_{2^m} - \frac{1+ 2^{x-1} - 2^{2^m}}{(2^{2^m} -1) (2^x -1)} \\

805: & < & \rho_{2^m} ~.

806: \end{eqnarray*}

807: This completes the induction step.

808: 

809: To complete the proof of Lemma \ref{L205}, by the induction hypothesis it suffices to prove that

810: $\rho_{2^m} < c$ for all $m$.

811: But from \eqn{Eq209} it follows that

812: $$\rho_{2^m} = \sum_{i=0}^m \frac{1}{2^{2^i} -1} < c ~,$$

813: and that $\rho_{2^m} \to c$ as $m \to \infty$.~~~\bsq

814: 

815: If $N$ is not a power of 2 then similar arguments show that

816: %

817: %To solve this we set $\beta_m = \af_m / (2^{2^m} -1)$ and find easily that

818: %$$\beta_m = \sum_{i=0}^m \frac{1}{2^{2^i} -1} ~,$$

819: %so $\beta_m \to c = 1.4039 \ldots$ as $m \to \infty$.

820: %Then $\af_m \le c(2^{2^m} -1)$ and $\af_m / (2^{2^m} -1) \to c$ as $m \to \infty$.

821: %

822: %If $N$ is a power of 2 but not of the form $2^{2^m}$ then

823: %$ht(N) < cN$.

824: %For other values of $N$ the same methods show that

825: the height is

826: (considerably) less than $cN$.

827: This completes the proof of Theorem \ref{th1}.

828: 

829: \section{Upper bounds on the number of parameter sets}

830: In this section we establish the upper bounds in

831: Theorems \ref{th2} and \ref{th3}.

832: We will bound $T'(N)$, the number of nodes in $\La'_N$.

833: Since $\La_N$ is a sublattice of $\La'_N$, this is also an upper bound

834: on the number of nodes in $\La_N$.

835: Suppose first that $N=2^{2r}$.

836: 

837: We start by considering parameter sets $(N, 2^{k_1} 4^{k_2} 8^{k_3} \ldots (2^r)^{k_r} )$, containing no level exceeding $\sqrt{N}$.

838: From \eqn{EqR},

839: \beql{Eq19}

840: N-1 \ge k_1 + 3k_2 + 7k_3 + \cdots +

841: (2^r -1) {k_r} ~.

842: \eeq

843: Let $\gamma$ be the number of nonnegative integer solutions

844: $(k_1, k_2, \ldots, k_N )$ to this inequality.

845: Then $\gamma / (N-1)^r$ is the Riemann sum approximating the volume

846: of the simplex bounded by the hyperplanes

847: \begin{eqnarray*}

848: && 1 \ge x_1 + 3x_2 + 7x_3 + \cdots + (2^r -1) x_r ~, \\

849: && x_1 \ge 0 , ~ x_2 \ge 0 , \ldots, x_r \ge 0

850: \end{eqnarray*}

851: in $\RR^r$.

852: For large $N$ this yields

853: \beql{Eq20}

854: \gamma = \frac{N^r}{r! \prod\limits_{i=1}^r (2^i -1)}

855: (1+o(1)) ~.

856: \eeq

857: The product in the denominator approaches $c_3 2^{r(r+1)/2}$, as $r \to \infty$, where $c_3 = 0.2887 \ldots$.

858: 

859: Now suppose the

860: parameter set contains a factor at $2^i$ levels, where $r+1 \le i \le 2r-1$.

861: There can be at most one such factor, and the number of such parameter

862: sets in each case is at most $\gamma$.

863: The total number of parameter sets is therefore at most $r \gamma$, and setting $r= \frac{1}{2} \log_2 N$ we find that 

864: $$ \log_2 (r \gamma ) \le \frac{3}{8} (\log_2 N)^2 (1+o(1)) ~.$$

865: This establishes the upper bound in Theorem \ref{th2} for $N=2^{2r}$. It also implies 

866: the upper bound for $N=2^{2r+1}$, after noting that $T'(N) \le T'(2N)$.

867: 

868: We now give a sketch of the proof of Theorem \ref{th3}, omitting many tedious

869: details.

870: To simplify the analysis we will 

871: neglect terms on the right-hand side of (\ref{EqR}) that correspond to factors 

872: with a level greater than $\sqrt{N}$.

873: Suppose first that $N$ is a large number of the form $2^{2a_1} 3^{2a_2}$.

874: Pretending for the moment that $a_1$ and $a_2$ are allowed to be real numbers, not just integers, we may consider what choice of $a_1$ and $a_2$ maximizes the number of solutions to \eqn{EqR} for a given value of $N$.

875: The number of terms on the right-hand side of \eqn{EqR} is now $(a_1 + 1)(a_2 + 1) - 1$.

876: The arguments used to establish the upper bound of Theorem \ref{th2} show that

877: the number of solutions to \eqn{EqR}

878: is maximized if $2^{2a_1}$ is approximately equal to $3^{2a_2}$.

879: 

880: Now suppose that $N$ is of the form

881: \beql{Eq7A}

882: p_1^{2a_1} ~ p_2^{2a_2} \cdots p_m^{2a_m} ~,

883: \eeq

884: where $p_1 =2$, $p_2=3, \ldots$ are the first $m$ primes.

885: We find that

886: the number of solutions to \eqn{EqR} is maximized when the numbers

887: $p_i^{2a_i}$ are all approximately equal, and we will therefore assume that $p_i^{2a_i} = N^{\frac{1}{m} (1+o(1))}$, i.e. that

888: $$a_i = \frac{1}{2m} \frac{\ln N}{\ln p_i} (1+o(1)) , \quad i =1, \ldots, m ~.$$

889: 

890: The Rao bound contains a term for every possible level

891: $$s = p_1^{i_1} ~ p_2^{i_2} \cdots p_m^{i_m}$$

892: in which $0 \le i_{\nu} \le a_\nu$, $1 \le \nu \le m$, where not

893: all the $i_\nu$ are equal to 0.

894: The number of such terms is

895: \beql{Eq30}

896: \delta  :=  (a_1 + 1) (a_2 + 1) \cdots (a_m + 1) -1 =

897: \frac{1}{(2m)^m} \frac{(\ln N)^m}{\prod\limits_{j=1}^m \ln p_j}

898: (1+ o(1)) ~.

899: \eeq

900: The product of the coefficients of all the terms on the right-hand side of the Rao bound is

901: $$\zeta := p_1^{a_1^2 a_2 \ldots a_m /2}

902: p_2^{a_1 a_2^2 a_3 \ldots a_m /2} \cdots

903: p_m^{a_1 \ldots a_{m-1} a_m^2/2} (1+ o(1)) ~.

904: $$

905: This implies

906: $$\ln \zeta = \frac{\delta}{4} \ln N (1+o(1)) ~.$$

907: Again using $\gamma$ to denote the number of solutions to the Rao inequality, we have

908: $$\gamma = \frac{N^\delta}{\delta ! \zeta} (1+ o(1)) ~,$$

909: hence

910: \begin{eqnarray*}

911: \ln \gamma & = & \frac{3}{4} \frac{(\ln N)^{m+1}}{(2m)^m (\ln m)^{m (1+o(1))}} -

912: \frac{1}{(2m)^m} \frac{(\ln N)^m}{(\ln m)^{m(1+o(1))}} (m \ln \ln N - m \ln m) \\

913: &&~~~~+~ \mbox{smaller terms} ~.

914: \end{eqnarray*}

915: This expression is maximized if we take

916: $$m = \frac{1}{2e} \frac{\ln N}{\ln \ln N} (1+o(1)) ~,$$

917: and then we find that the leading term in the expression for $\ln \gamma$ is

918: $$\frac{3}{4} \ln N N^{\frac{1}{2e} \frac{1}{\ln \ln N}} ~.$$

919: We conclude that

920: $$\ln \ln \gamma \le \frac{1}{2e} \frac{\ln N}{\ln \ln N} ~,$$

921: which establishes Theorem \ref{th3}.

922: 

923: \section{Geometric orthogonal arrays}

924: We consider subspaces $V$ of the vector space $GF(q)^n$ over $GF(q)$, where $q$ is a power of a prime.

925: By the

926: dimension of $V$, $\dim V$, we mean the vector space dimension over $GF(q)$ (rather than the projective dimension, which is one less).

927: The following

928: notion was suggested by the notions of spread and partial spread in projective geometry

929: (cf. Thas, 1995).

930: 

931: \paragraph{Definition.}

932: A {\em mixed spread of strength $t$} is a collection

933: $\sV = \{V_1, V_2 , \ldots, V_k \}$ of subspaces of $GF(q)^n$ such that for all choices

934: of $\tau \le t$ indices $i_1, i_2, \ldots, i_\tau$ 

935: (with $1 \le i_1 < i_2 < \cdots < i_\tau \le k$) the dimension

936: of the span of $V_{i_1}, \ldots, V_{i_\tau}$ is equal to $\dim V_{i_1} + \cdots + \dim V_{i_\tau}$.

937: 

938: An equivalent condition is that the span of $V_{i_1}, \ldots, V_{i_\tau}$ is

939: the direct sum $V_{i_1} \oplus \cdots \oplus V_{i_\tau}$

940: for all choices of $\tau \le t$ indices $i_1, \ldots, i_\tau$ 

941: with $1 \le i_1 < \cdots < i_\tau \le k$.

942: 

943: Any collection $\sV$ of subspaces has strength 1.

944: $\sV$ has strength 2 if and only if every pair $V_i$, $V_j \in \sV$, $i \neq j$, intersect just in the zero vector.

945: $\sV$ has strength 3 if and only if it has strength 2 and for any triple of distinct

946: subspaces each one meets the span of the other two just in the zero vector.

947: 

948: If $V$ is a $d$-dimensional subspace of $GF(q)^n$ we denote by $V^\ast$ the dual space, the

949: space of linear functionals on $V$

950: (see for example Hoffman and Kunze, 1961),

951: and we fix a labeling $f_0 , f_1, \ldots, f_{q^d -1}$ for the elements of $V^\ast$.

952: 

953: Given a mixed spread of strength $t$, $\sV= \{ V_1, V_2, \ldots, V_k \}$, where the $V_i$ are 

954: subspaces of $GF(q)^n$, we obtain an orthogonal array $OA(\sV)$ with $q^n$ runs as follows.

955: The columns of the array are labeled $V_1, V_2, \ldots, V_k$ and the rows are labeled

956: by the linear functionals $f \in (GF(q)^n )^\ast$.

957: If $f$ restricted to $V_i$, $f |_{V_i}$, is the $j$th linear functional in $V_i^\ast$,

958: the $(f, V_i)$-th entry in the array is $j$.

959: The symbols in column $i$ are therefore taken from $\{0,1, \ldots, q^{\dim V_i} -1 \}$.

960: 

961: We will say that an orthogonal array constructed in this way is {\em geometric}.

962: 

963: \begin{theorem}\label{P1}

964: The orthogonal array $OA(\sV)$ has strength $t$ if and only if the mixed spread $\sV$ has strength $t$.

965: \end{theorem}

966: 

967: \paragraph{Proof.}

968: Suppose $OA(\sV)$ has strength $t$.

969: Consider for example the first $t$ columns.

970: In the projection of the array onto these columns we see

971: $$\prod_{i=1}^t q^{\dim V_i} = q^{\sum\limits_{i=1}^t \dim V_i}$$

972: different $t$-tuples of symbols.

973: Since these depend only on the restrictions of the $f \in (GF(q)^n)^\ast$ to the span of

974: $V_1, \ldots, V_t$, the dimension of that space must be at least the sum of the dimensions of $V_1, \ldots, V_t$, and clearly it cannot have a higher dimension.

975: So $\sV$ is a mixed spread of strength $t$.

976: 

977: Conversely, suppose $\sV$ is a mixed spread of strength $t$.  We can write

978: $$GF(q)^n = V_{1} \oplus \cdots \oplus V_{t} \oplus X$$

979: where $X$ is the complementary space to the $V_i$.

980: Since the dual of a direct sum is canonically isomorphic to the direct sum of

981: the duals, we immediately find that as we run through the

982: linear functionals on $GF(q)^n$, every tuple

983: $(f|_{V_1} , \ldots f|_{V_t} , f|_X )$ 

984: of restrictions occurs precisely once.

985: Ignoring the last component, we see that every tuple

986: $(f|_{V_1} , \ldots f|_{V_t} )$ occurs precisely $|X|$ times.

987: Hence $OA(\sV)$ has strength $t$.~~~$\bsq$

988: 

989: \begin{lemma}\label{P5}

990: Any geometric array of strength $2$ can always be extended to a tight array (i.e. one meeting the Rao bound) by adding $q$-level factors.

991: \end{lemma}

992: 

993: \paragraph{Proof.}

994: We simply group any unused points into 1-dimensional subspaces.~~~$\bsq$

995: 

996: \paragraph{Examples.} (i)

997: The 1-dimensional subspaces of $GF(q)^n$ form a mixed spread of strength 2.

998: The corresponding array is the familiar

999: $$OA(q^n , q^k ) , \quad k = (q^n -1) /(q-1) ~,$$

1000: of the Rao-Hamming construction.

1001: 

1002: (ii) More generally, a classical $a$-spread in $PG(b,q)$ is a mixed spread of strength 2 in our sense.

1003: This is a set of subspaces of $PG(b,q)$ of projective dimension $a$ which

1004: partitions $PG(b,q)$ (Thas, 1995), and exists if and only if $a+1$ divides $b+1$.

1005: From Theorem \ref{P1} we obtain an

1006: $$OA(q^{b+1}, (q^{a+1} )^k ), \quad

1007: k = (q^{b+1} -1)/(q^{a+1} -1) \,,

1008: $$

1009: which of course is also given by the Rao-Hamming construction.

1010: 

1011: We could also have obtained example (ii) directly from example (i), by remarking that a mixed

1012: spread of strength $t$ over $GF(q)$, $q =p^\beta$, is also a mixed spread

1013: of strength $t$ over $GF(q')$, $q' = p^\af$, if $q'$ divides $q$.

1014: The dimensions of the subspaces are multiplied by $\beta / \af$.

1015: 

1016: (iii) Provided $a \ge b/2$, there exists a mixed spread of strength 2

1017: in $GF(q)^b$ consisting of a single subspace $GF(q)^a$ and a partitioning of the remaining points into $q^a$ subspaces $GF(q)^{b-a}$.

1018: This can be proved directly, or alternatively is equivalent to Lemma 2.1

1019: of Eisfeld, Storme and Sziklai (1999).

1020: From Theorem \ref{P1} we obtain a geometric

1021: $$OA(q^b, (q^{b-a} )^{q^a} (q^a)^1 )$$

1022: whenever $a \ge b/2$.

1023: Orthogonal arrays with these parameters were already known from the difference scheme

1024: construction (Hedayat, Sloane and Stufken, 1999, Example 9.19), but the present

1025: construction also shows that they are geometric.

1026: 

1027: (iv) The classical ``partial $a$-spread'' constructed in Lemma 2.2 of Eisfeld, Storme and Sziklai (1999)

1028: translates in our language into a mixed spread of strength 2 consisting of $k$ $b$-dimensional

1029: subspaces $(b \ge 2)$ of $GF(q)^n$, where $n = ib+r$, $0 \le r< b$, and

1030: $$k =q^r \frac{q^{ib} -1}{q^b -1} - q^r +1 ~.$$

1031: This produces a geometric $OA(q^n, (q^b)^k)$

1032: (again arrays with these parameters were known from

1033: the difference scheme construction),

1034: which by Lemma \ref{P5} can be extended to a tight

1035: \beql{Eq80}

1036: OA(q^n, (q^1)^l (q^b)^k ) ~,

1037: \eeq

1038: where $l=q^b (q^r -1)/(q-1)$.

1039: The orthogonal arrays constructed by Wu (1989) are a special case of \eqn{Eq80},

1040: but in general these arrays may be new.

1041: 

1042: (v) Generalizing examples (i) and (ii), any orthogonal array formed from the codewords of

1043: a projective linear code (one for which the columns of a generator matrix are nonzero

1044: and projectively distinct) is geometric.

1045: 

1046: (vi) The $OA(256, 2^{16} )$ of strength 5 formed from the Nordstrom-Robinson code (see Hedayat, Sloane and Stufken, 1999, Section 5.10) is not geometric, and no geometric

1047: $OA(256, 2^{16} )$ of strength 5 exists.

1048: 

1049: We shall see other examples in Section 6.

1050: 

1051: \paragraph{Remarks.}

1052: An unmixed geometric orthogonal array is always linear, in the sense of Hedayat, Sloane and Stufken (1999), Chapter 3.

1053: In general a mixed geometric orthogonal array is additive but not necessarily linear\footnote{For the distinction between additive and linear sets in the

1054: context of coding theory see Calderbank et~al. (1998).}

1055: over each of the fields involved.

1056: 

1057: If the strength is 2, the number of degrees of freedom in the parameter set for $OA(\sV)$ is equal to the total number of nonzero points in all the subspaces $V_i$.

1058: 

1059: Finally, the following is a recipe for constructing the orthogonal

1060: array from a mixed spread 

1061: $\sV = \{V_1, V_2 , \ldots, V_k \}$ of subspaces of $GF(q)^n$

1062: in the case when $q$ is a prime.

1063: Let $v_{1}^{(i)}, \ldots, v_{d_i}^{(i)}$ be a basis for $V_i$,

1064: where $d_i = \dim V_i$, $1 \le i \le k$.

1065: Let $w_0, \ldots, w_{q^n - 1}$ be the vectors of $GF(q)^n$.

1066: Then the $i$th entry of the $j$th row of the orthogonal array,

1067: for $1 \le i \le k$, $0 \le j \le q^n - 1$, 

1068: is the number 

1069: $$ \sum_{r=1}^{d_i} w_j \cdot v_r^{(i)} ~ q^{r - 1} ~.$$

1070: (This is a number in the range $\{0, \ldots, q^{d_i}-1\}.)$

1071: 

1072: \section{If the number of runs is a power of 2}

1073: In this section we consider the case $N=2^n$, $n=1,2, \ldots$.

1074: We have already discussed $ht(N)$ in Section 3 (see Table \ref{ta2}).

1075: With the assistance of Michele Colgan, we used a computer to determine the number of dual atoms $A' (N)$ and the total number of nodes $T' (N)$ in the idealized lattice $\La'_N$ for $n \le 9$.

1076: The results are shown in the second and third columns of Table \ref{ta3}.

1077: Note in particular the extremely rapid growth from $N=256$ to $N=512$.

1078: We regard this as convincing evidence that when $N=2^n$, $A'(N)$ (and therefore

1079: presumably $A(N)$) grows faster than any polynomial in $N$.

1080: \begin{table}[htb]

1081: \caption{Dual atoms and total number of nodes in idealized lattice $\La'_N$

1082: $(A'(N)$ and $T' (N))$ and in lattice $\La_N$ $(A(N)$ and $T(N))$.}

1083: $$

1084: \begin{array}{rrrrr}

1085: \multicolumn{1}{c}{N} &

1086: \multicolumn{1}{c}{A'(N)} &

1087: \multicolumn{1}{c}{T'(N)} &

1088: \multicolumn{1}{c}{A(N)} &

1089: \multicolumn{1}{c}{T(N)}

1090: \\ [+.1in]

1091: 1 & 0 & 1 & 0 & 1 \\

1092: 2 & 1 & 2 & 1 & 2 \\

1093: 4 & 1 & 5 & 1 & 5 \\

1094: 8 & 1 & 14 & 1 & 14 \\

1095: 16 & 2 & 61 & 2 & 61 \\

1096: 32 & 3 & 322 & 2 & 320 \\

1097: 64 & 11 & 3058 & 7 & 3037 \\

1098: 128 & 21 & 33364 \\

1099: 256 & 72 & 789085 \\

1100: 512 & 144521 & 18614215

1101: \end{array}

1102: $$

1103: \label{ta3}

1104: \end{table}

1105: 

1106: As to the lattice $\La_N$ itself, for $n \le 4$ this is covered by the results in Section 2.

1107: For $N=32$ there are precisely two parameter sets in $\La'_{32}$ which do not exist, $(32, 4^{10} )$ and $(32, 2^1 4^{10} )$.

1108: These can be ruled out either by the linear programming bound or by the Bose-Bush bound (Hedayat, Sloane and Stufken,

1109: 1999, Theorem 2.8).

1110: All other parameter sets in $\La'_{32}$ are realized.

1111: It follows that $\La_{32}$ contains exactly two dual atoms, $OA(32, 2^{16} 16^1 )$ and $OA(32, 4^8 8^1 )$.

1112: 

1113: Before considering $\La_{64}$ we give a lemma that will be used to construct new arrays.

1114: 

1115: \begin{lemma}\label{P2}

1116: Suppose $V_1$, $V_2$, $V_3$ are three $r$-dimensional subspaces of $GF(2)^{2r}$ such that

1117: $V_i \cap V_j = \{0\}$, $i \neq j$.

1118: Then their union can be replaced by $2^r -1$ two-dimensional subspaces,

1119: any pair of which meet just in the zero vector.

1120: \end{lemma}

1121: 

1122: {\em Proof.}

1123: Since $V_1 \cap V_2 = \{ 0\}$, $V_1$ and $V_2$ span the space

1124: $GF(2)^{2r}$.

1125: Let $\pi_1$, $\pi_2$ be the associated projection maps from $GF(2)^{2r}$ to $V_1$, $V_2$ respectively.

1126: Then $i_1 = \pi_1 \bigl|_{V_3} : V_3 \to V_1$ and $i_2 = \pi_2 \bigl|_{V_{3}} : V_3 \to V_2$ are both

1127: isomorphisms.

1128: It follows that $V_3$ is the set

1129: $$\{ v+ i(v) : v ~\mbox{in}~

1130: V_1 \} , ~~i=i_2 i_1^{-1} ~.

1131: $$

1132: But then we need simply take the planes $\{0, v, i(v), v+i(v) \}$ for $v \in V_1$ to establish

1133: the lemma.~~~$\bsq$

1134: 

1135: The lemma implies that if a geometric $OA(2^{2r}, \ldots )$ exists then so does the array

1136: obtained by replacing $(2^r)^3$ in the parameter set by $4^k$, $k=2^r -1$.

1137: In particular, in a geometric $OA(64, \ldots )$ we can replace $8^3$ by $4^7$.

1138: 

1139: \begin{theorem}\label{P3}

1140: The lattice $\La_{64}$ contains precisely seven dual atoms, with parameter sets

1141: \begin{eqnarray}\label{Eq69}

1142: && (64, 2^5 4^{17} 8^1), (64, 4^{14} 8^3 ) , (64, 2^5 4^{10} 8^4) , \nonumber \\

1143: && (64, 4^7 8^6 ), (64, 8^9), (64, 4^{16} 16^1 ) , (64, 2^{32} 32^1 ) \,.

1144: \end{eqnarray}

1145: A geometric orthogonal array exists for each of these parameter sets.

1146: \end{theorem}

1147: 

1148: \paragraph{Proof.}

1149: As an intermediate step, we use

1150: mixed spreads of strength 2

1151: to construct orthogonal arrays with the following parameter sets:

1152: \begin{eqnarray}\label{Eq70}

1153: && (64, 4^{21} ),

1154: (64, 2^5 4^{17} 8^1),

1155: (64, 2^4 4^{15} 8^2),

1156: (64, 4^{14} 8^3) , \nonumber \\

1157: && (64, 2^5 4^{10}  8^4) ,

1158: (64, 2^4 4^8 8^5) , (64, 4^7 8^6) , (64, 2^8 4^2 8^7) , \nonumber \\

1159: && (64, 2^4 4^1 8^8),

1160: (64, 8^9) , (64, 4^{16} 16^1),

1161: (64, 2^{32} 32^1 ) \,.

1162: \end{eqnarray}

1163: On the other hand, linear programming shows that orthogonal arrays do not exist with

1164: parameter sets

1165: \beql{Eq71}

1166: (64, 4^{18} 8^1), (64, 4^{16} 8^2) , (64, 4^{11} 8^4), (64, 4^9 8^5),

1167: (64, 4^3 8^7), (64, 4^2 8^8) \,.

1168: \eeq

1169: We then check that every parameter set with 64 runs either dominates one of \eqn{Eq71}

1170: (and so is not realized), or is dominated by one of \eqn{Eq70} (and is realized).

1171: Furthermore, the parameter sets in \eqn{Eq69} dominate all of \eqn{Eq70}.

1172: 

1173: It remains to construct the arrays mentioned in \eqn{Eq69}.

1174: The last two follow from Example (iii) of Section 5.

1175: Also $(64, 8^9)$ comes from Example (i), and $(64, 4^7 8^6)$ and

1176: $(64, 4^{14} 8^3)$ follow from Lemma \ref{P2}.

1177: 

1178: To construct an $OA(64, 2^5 4^{10} 8^4)$ we proceed as follows.

1179: We begin by constructing an explicit example of an $OA( 64, 8^9)$ from Theorem \ref{P1}, by using an extended Reed-Solomon code of length 9, dimension 2

1180: and minimal distance 8 over $GF(8)$.

1181: This gives a decomposition of $GF(2)^6$ into 9 copies of $GF(2)^3$

1182: meeting only in the zero vector.

1183: These 9 subspaces are spanned by

1184: the following nine triples of columns:

1185: \beql{Eq72}

1186: \begin{array}{ccccccccc}

1187: 0 & I & I & I & I & I & I & I & I \\

1188: I & 0 & I & A & A^2 & A^3 & A^4 & A^5 & A^6

1189: \end{array}

1190: \eeq

1191: where

1192: $$0 = \left[ \begin{array}{ccc}

1193: 0 & 0 & 0 \\

1194: 0 & 0 & 0 \\

1195: 0 & 0 & 0

1196: \end{array}

1197: \right], \quad

1198: I = \left[

1199: \begin{array}{ccc}

1200: 1 & 0 & 0 \\

1201: 0 & 1 & 0 \\

1202: 0 & 0 & 1

1203: \end{array}

1204: \right] , \quad

1205: A = \left[

1206: \begin{array}{ccc}

1207: 0 & 1 & 0 \\

1208: 1 & 0 & 1 \\

1209: 0 & 1 & 1

1210: \end{array}

1211: \right] ,

1212: $$

1213: and $A^7 = I$.

1214: 

1215: We may replace the first four subspaces and the last subspace (which together

1216: contain 35 nonzero points) by ten two-dimensional

1217: subspaces with five single points left over.

1218: One choice for the ten two-dimensional subspaces

1219: is shown in Table \ref{ta4}.

1220: \begin{table}[htb]

1221: \caption{Ten pairs of columns each spanning a two-dimensional

1222: subspace of $GF(2)^6$.}

1223: $$

1224: \begin{array}{c@{\,}c}

1225: 0 & 1 \\

1226: 0 & 0 \\

1227: 0 & 1 \\

1228: 0 & 0 \\

1229: 1 & 0 \\

1230: 1 & 1 \\

1231: \end{array}

1232: \qquad

1233: \begin{array}{c@{\,}c}

1234: 0 & 1 \\

1235: 0 & 1 \\

1236: 0 & 0 \\

1237: 1 & 1 \\

1238: 0 & 1 \\

1239: 1 & 0 \\

1240: \end{array}

1241: \qquad

1242: \begin{array}{c@{\,}c}

1243: 0 & 0 \\

1244: 0 & 1 \\

1245: 1 & 0 \\

1246: 0 & 0 \\

1247: 1 & 1 \\

1248: 1 & 0 \\

1249: \end{array}

1250: \qquad

1251: \begin{array}{c@{\,}c}

1252: 0 & 0 \\

1253: 0 & 0 \\

1254: 0 & 1 \\

1255: 0 & 0 \\

1256: 0 & 0 \\

1257: 1 & 0 \\

1258: \end{array}

1259: \qquad

1260: \begin{array}{c@{\,}c}

1261: 0 & 1 \\ 0 & 1 \\ 0 & 0 \\ 1 & 0 \\ 1 & 0 \\ 1 & 0 \\

1262: \end{array}

1263: \qquad

1264: \begin{array}{c@{\,}c}

1265: 0 & 1 \\ 1 & 1 \\ 1 & 1 \\ 0 & 1 \\ 0 & 1 \\ 0 & 1 \\

1266: \end{array}

1267: \qquad

1268: \begin{array}{c@{\,}c}

1269: 0 & 0 \\ 0 & 1 \\ 0 & 0 \\ 1 & 0 \\ 0 & 0 \\ 0 & 0 \\

1270: \end{array}

1271: \qquad

1272: \begin{array}{c@{\,}c}

1273: 0 & 1 \\ 0 & 0 \\ 0 & 0 \\ 1 & 0 \\ 1 & 1 \\ 0 & 0 \\

1274: \end{array}

1275: \qquad

1276: \begin{array}{c@{\,}c}

1277: 1 & 0 \\ 0 & 1 \\ 0 & 1 \\ 0 & 1 \\ 0 & 1 \\ 0 & 0 \\

1278: \end{array}

1279: \qquad

1280: \begin{array}{c@{\,}c}

1281: 0 & 1 \\ 1 & 1 \\ 0 & 1 \\ 1 & 0 \\ 0 & 0 \\ 1 & 0

1282: \end{array}

1283: $$

1284: \label{ta4}

1285: \end{table}

1286: This gives a mixed spread of strength 2 consisting of four 3-dimensional subspaces,

1287: ten two-dimensional subspace and five points,

1288: and so by Theorem \ref{P1} corresponds to an $OA(64, 2^5 4^{10} 8^4 )$.

1289: Finally, Lemma \ref{P2} produces an $OA(64, 2^5 4^{17} 8^1 )$.~~~$\bsq$

1290: 

1291: \paragraph{Remark.}

1292: Geometric orthogonal arrays with parameter sets of the form

1293: $(64, \ldots 8^k \ldots )$ involve selecting $k$ disjoint (except for the zero vector) copies of $GF(2)^3$ inside $GF(2)^6$.

1294: It is simpler to work projectively, and then we must choose $k$ disjoint copies of $PG(2,2)$ inside $PG(5,2)$.

1295: Equation \eqn{Eq72} then gives a decomposition of $PG(5,2)$ into nine copies of $PG(2,2)$.

1296: 

1297: With the help of Magma (Bosma and Cannon, 1995;

1298: Bosma, Cannon and Mathews, 1994;

1299: Bosma, Cannon and Playoust, 1997),

1300: we showed that if $1 \le k \le 4$ there is a unique way to choose $k$ disjoint

1301: $PG(2,2)$'s in $PG(5,2)$, and these are equivalent to a subset of \eqn{Eq72}.

1302: For $k=5$, there are precisely two ways, one of which is equivalent

1303: to a subset of \eqn{Eq72}

1304: while the other contains no $PG(2,2)$ in its complement, and so cannot be extended to $k=6$.

1305: For $k=6, \ldots, 9$, there is again a unique way to choose $k$ disjoint planes.

1306: In particular the decomposition into nine planes shown in \eqn{Eq72} is also unique.

1307: 

1308: An example of a maximal set of five $PG(2,2)$'s in $PG(5,2)$ is shown in Table \ref{ta5}.

1309: This corresponds to a geometric $OA(64, 8^5)$ that cannot be extended to a geometric $OA(64, 8^6)$.

1310: It would be interesting to determine if it can be extended to {\em any} $OA(64, 8^6)$.

1311: \begin{table}[htb]

1312: \caption{A set of five disjoint $PG(2,2)$'s in $PG(5,2)$ that is not contained in a set of six.  Each triple of columns spans one of the subspaces.}

1313: 

1314: $$

1315: \begin{array}{c@{\,}c@{\,}c}

1316: 1 & 0 & 0 \\

1317: 0 & 1 & 0 \\

1318: 0 & 0 & 1 \\

1319: 0 & 0 & 0 \\

1320: 0 & 0 & 0 \\

1321: 0 & 0 & 0 \\

1322: \end{array}

1323: \qquad

1324: \begin{array}{c@{\,}c@{\,}c}

1325: 0 & 0 & 0 \\

1326: 0 & 0 & 0 \\

1327: 0 & 0 & 0 \\

1328: 1 & 0 & 0 \\

1329: 0 & 1 & 0 \\

1330: 0 & 0 & 1 \\

1331: \end{array}

1332: \qquad

1333: \begin{array}{c@{\,}c@{\,}c}

1334: 1 & 0 & 0 \\

1335: 0 & 1 & 0 \\

1336: 0 & 0 & 1 \\

1337: 1 & 0 & 0 \\

1338: 0 & 1 & 0 \\

1339: 0 & 0 & 1 \\

1340: \end{array}

1341: \qquad

1342: \begin{array}{c@{\,}c@{\,}c}

1343: 1 & 0 & 0 \\

1344: 0 & 1 & 0 \\

1345: 0 & 0 & 1 \\

1346: 0 & 1 & 0 \\

1347: 1 & 0 & 1 \\

1348: 0 & 1 & 1 \\

1349: \end{array}

1350: \qquad

1351: \begin{array}{c@{\,}c@{\,}c}

1352: 1 & 0 & 0 \\

1353: 0 & 1 & 0 \\

1354: 0 & 0 & 1 \\

1355: 1 & 1 & 1 \\

1356: 1 & 1 & 0 \\

1357: 1 & 0 & 0 \\

1358: \end{array}

1359: $$

1360: \label{ta5}

1361: \end{table}

1362: 

1363: %where $W$ is the span of $V_1, \ldots, V_t$ and $X$ is the complementary space.

1364: 

1365: \section{The existence of orthogonal arrays with certain parameter sets}

1366: In this section we prove Theorem \ref{th4}, the lower bound in Theorem \ref{th2}, and also give some other conditions which are sufficient to guarantee that a parameter set can be realized by an orthogonal array.

1367: 

1368: \begin{lemma}\label{le6}

1369: Suppose $N=p^m$ is a power of a prime and $(N, s_1^{k_1} s_2^{k_2} \ldots )$

1370: is a parameter set with $k= \Sigma_i k_i$ factors.

1371: If $k \le p^{\lfloor (m+1)/2 \rfloor} +1$ then this parameter set is realized by a geometric orthogonal array.

1372: \end{lemma}

1373: 

1374: \paragraph{Proof.}

1375: Suppose first that the parameter set contains a factor with $s=p^n > \sqrt{N}$ levels.

1376: If $m$ is even then a geometric $OA(p^m , (p^{m-n} )^{p^n} (p^n)^1)$ exists

1377: by Section 5, and $p^{m-n}$ is the largest number of levels other factors can have if there is an $s$-level factor.

1378: Since there are $p^n$ factors with $p^{m-n}$ levels, the existence of any array with one

1379: $s$-level factor and at most $p^{m/2}$ factors with $\le p^{m-n}$ levels follows immediately.

1380: The case that $m$ is odd follows similarly.

1381: 

1382: We now assume that all $s_i \le \sqrt{N}$.

1383: If $m$ is even, $N=p^{2r}$, then a geometric

1384: \beql{Eq61}

1385: OA (p^{2r} , (p^r )^{ p^r +1})

1386: \eeq

1387: exists by Section 5.

1388: Any parameter set with all $s_i \le p^r$ and $k \le p^r +1$ is dominated by \eqn{Eq61} and so is realized.

1389: If $m$ is odd, $N= p^{2r+1}$, then a geometric

1390: \beql{Eq62}

1391: OA(p^{2r+1}, (p^r)^{p^{r+1} +1} )

1392: \eeq

1393: also exists by Section 5.

1394: Any parameter set with all $s_i \le p^r$ and $k \le p^{r+1} +1$ is dominated by \eqn{Eq62} and so is also realized.~~~$\bsq$

1395: 

1396: Since the number of factors in a parameter set is less than or equal to the number of degrees of freedom

1397: \eqn{Eq61a}, Lemma \ref{le6} immediately implies that any

1398: parameter set

1399: $(N=p^m, s_1^{k_1} s_2^{k_2} \ldots )$ with at most

1400: $p^{\lfloor (m+1)/2 \rfloor} +1$ degrees of freedom is realized by an orthogonal array.

1401: However, Theorem \ref{th4} is much stronger.

1402: 

1403: \paragraph{Proof of Theorem \ref{th4}.}

1404: We will show that any parameter set

1405: $(p^m, p^{k_1} (p^2)^{k_2} (p^3)^{k_3} \ldots )$ satisfying

1406: \beql{Eq40}

1407: \sum_{i \ge 1} k_i (p^i -1) \le p^{3m/4}

1408: \eeq

1409: is realized by a geometric orthogonal array, where $p$ is any prime.

1410: To simplify the notation we assume $m=4r$

1411: is a multiple of 4.

1412: The arguments in the other three cases require only minor modifications and are left to the reader.

1413: 

1414: From \eqn{Eq40} we have

1415: \beql{Eq41}

1416: k_{r+1} + k_{r+2} + \cdots + k_{4r-1} \le p^{2r} +1 ~,

1417: \eeq

1418: and so by Lemma \ref{le6} a geometric

1419: $$OA( p^{4r} , (p^{r+1})^{k_{r+1}} (p^{r+2})^{k_{r+2}} \ldots

1420: (p^{4r-1} )^{k_{4r-1}} )

1421: $$

1422: exists.

1423: 

1424: We now proceed by induction.

1425: Let $H_n$ be the hypothesis that every parameter set

1426: \beql{Eq50}

1427: (p^{4r} , p^{b_1} (p^2)^{b_2} \ldots

1428: (p^r)^{b_r} (p^{r+1} )^{k_{r+1}} \ldots

1429: (p^{4r-1} )^{k_{4r-1}} )

1430: \eeq

1431: with

1432: $b_i \le k_i$ for $1 \le i \le r$ and

1433: $$b_1 + b_2 + \cdots + b_r = n$$

1434: can be realized by a geometric orthogonal array constructed using disjoint

1435: subspaces of $PG(4r-1, p)$.

1436: We have shown that $H_0$ holds.

1437: Suppose $H_n$ holds with

1438: $$n < k_1 + k_2 + \cdots + k_r ~.$$

1439: We will show that we can increase $b_r$ by 1 and still realize the

1440: parameter set, thus establishing $H_{n+1}$.

1441: 

1442: To show this, note that the number of projectively

1443: distinct nonzero points in all the subspaces in \eqn{Eq50} is at most

1444: $$\sum_{i=1}^{4r-1} k_i \frac{p^i -1}{p-1} -1 \le \frac{p^{3r}}{p-1} -1 ~.$$

1445: However, by Theorem 1 of (Thas, 1995), Section 7,

1446: if a subset of $PG (4r-1, p)$ contains fewer than

1447: $$(p^{3r+1} -1)/ (p-1) ~~\mbox{points} ~,

1448: $$

1449: there is a subspace $PG(r-1, p)$ disjoint from it.

1450: Since

1451: $$\frac{p^{3r}}{p-1} -1 < \frac{p^{3r+1}-1}{p-1} ~,$$

1452: such a subspace exists and we can use it to augment $b_r$ by 1.

1453: 

1454: By induction, we can realize the parameter set $(p^{4r}, p^{k_1} (p^2)^{k_2} \ldots (p^{4r-1})^{k_{4r-1}} )$, as required.~~~$\bsq$

1455: 

1456: \paragraph{Proof of lower bound of Theorem \ref{th2}.}

1457: First suppose $N = 2^{2r}$.

1458: From Theorem \ref{th4}, every parameter set

1459: $$(2^{2r} , 2^{k_1} 4^{k_2} 8^{k_3} \ldots (2^r )^{k_r} )$$

1460: with at most $2^{3r/2}$ degrees of freedom can be realized.

1461: The lower bound of Theorem \ref{th2} now follows in the same way that

1462: we proved the upper bound in Section 4.

1463: If $N = 2^{2r+1}$ we use the previous case

1464: together with $T(N) \ge T(N/2)$.~~~$\bsq$

1465: 

1466: It would be nice to have analogues of Theorem \ref{th4} and Lemma \ref{le6} when $N$ is not a prime power!

1467: 

1468: \subsection*{Acknowledgments}

1469: We thank Michele Colgan for computing the properties of the lattices

1470: $\La'_N$ shown in Table \ref{ta3}.

1471: The research of John Stufken

1472: was supported by NSF grant DMS-9803684.

1473: 

1474: %\begin{thebibliography}{99}

1475: %

1476: %\bibitem[HSS]

1477: %Hedayat, A. S., Sloane, N. J. A. and Stufken, J. (1999).

1478: %{\em Orthogonal Arrays: Theory and Applications.}

1479: %New York: Springer-Verlag.

1480: %\end{thebibliography}

1481: 

1482: \clearpage

1483: \section*{References}

1484: 

1485: 

1486: \begin{description}

1487: 

1488: \item[~~~]

1489: Bosma, W. and Cannon, J. (1995).

1490: {\em Handbook of Magma Functions},

1491: Sydney.

1492:  

1493: \item[~~~]

1494: Bosma, W., Cannon, J. and Mathews, G. (1994).

1495: Programming with algebraic structures:

1496: Design of the Magma language, in

1497: {\em Proceedings of the 1994 International Symposium on Symbolic and Algebraic Computation},

1498: M. Giesbrecht, Ed.,

1499: Association for Computing Machinery, 52--57.

1500:  

1501: \item[~~~]

1502: Bosma, W., Cannon, J.  and Playoust, C. (1997).

1503: The Magma algebra system I:

1504: The user language.

1505: {\em J. Symb. Comp.},

1506: {\bf 24}, 235--265.

1507: 

1508: \item[~~~]

1509: Calderbank, A. R., Rains, E. M., Shor, P. W. and Sloane, N. J. A. (1998).

1510: \htmladdnormallink{Quantum error correction via codes over $GF(4)$}{http://www.research.att.com/~njas/doc/qc2.pdf}.

1511: {\em IEEE Trans. Information Theory}, {\bf 44}, 1369--1387.

1512: 

1513:  

1514: \item[~~~]

1515: Eisfeld, J., Storme, L. and Sziklai, P. (1999).

1516: Maximal partial spreads in finite projective spaces, preprint.

1517: 

1518: \item[~~~]

1519: Hedayat, A. S., Sloane, N. J. A. and Stufken, J. (1999).

1520: {\em \htmladdnormallink{Orthogonal Arrays: Theory and Applications}{http://www.research.att.com/~njas/doc/OA.html}.}

1521: New York: Springer-Verlag.

1522: 

1523: 

1524: \item[~~~]

1525: Hoffman, K. and Kunze, R. (1961).

1526: {\em Linear Algebra},

1527: Engelewood Cliffs, NJ: Prentice-Hall.

1528: 

1529: \item[~~~]

1530: Kimura, H. (1994a).

1531: Classification of Hadamard matrices of order 28 with Hall sets.

1532: {\em Discrete Math.},

1533: {\bf 128},

1534: 257--268.

1535: 

1536: \item[~~~]

1537: Kimura, H. (1994b).

1538: Classification of Hadamard matrices of order 28.

1539: {\em Discrete Math.},

1540: {\bf 133},

1541: 171--180.

1542: 

1543: \item[~~~]

1544: Sloane, N. J. A. (1999).

1545: The On-Line Encyclopedia of Integer Sequences.

1546: Published electronically at \htmladdnormallink{http://www.research.att.com/$\sim$njas/sequences/}{http://www.research.att.com/~njas/sequences/}.

1547: 

1548: \item[~~~]

1549: Sloane, N. J. A. and Stufken, J. (1996).

1550: \htmladdnormallink{A linear programming bound for orthogonal arrays with mixed levels}{http://www.research.att.com/~njas/doc/mixed.pdf}.

1551: {\em J. Statist. Plann. Infer.},

1552: {\bf 56}, 295--305.

1553: 

1554: \item[~~~]

1555: Thas, J. A. (1995).

1556: Projective geometries over a finite field,

1557: Chapter 7 of {\em Handbook of Incidence Geometry}, edited by

1558: Buekenhout, F. Amsterdam: North-Holland.

1559: 

1560: \item[~~~]

1561: Trotter, W. F. (1995).

1562: Partially ordered sets,

1563: Chapter 8 of {\em Handbook of Combinatorics}, edited by Graham, R. L.,

1564: Gr\"{o}tschel, M. and Lov\'{a}sz, L. Amsterdam: North-Holland;

1565: Cambridge, MA: MIT Press.

1566: 

1567: \item[~~~]

1568: Welsh, D. J. A. (1976).

1569: {\em Matroid Theory.}

1570: London: Academic Press.

1571: 

1572: \item[~~~]

1573: Wu, C. F. J. (1989).

1574: Construction of $2^m 4^n$ designs via a grouping scheme.

1575: {\em Annals Statist.}, {\bf 17}, 1880--1885.

1576: \end{description}

1577: 

1578: 

1579: \end{document}

1580: