1: \documentclass[11pt]{article}

2: 

3: \usepackage{amsmath}

4: \usepackage{amssymb}

5: \usepackage{amsthm}

6: \setlength{\textwidth}{6in} \setlength{\textheight}{8in}

7: \setlength{\evensidemargin}{0in} \setlength{\oddsidemargin}{0in}

8: \setlength{\footskip}{.5in} \setlength{\topmargin}{0in}

9: \setlength{\parskip}{8pt} \baselineskip 24pt

10: 

11: \usepackage[dvips]{graphicx}

12: \usepackage{color}

13: \usepackage{epsfig}

14: \usepackage[mathscr]{eucal}

15: \newtheorem{theorem}{Theorem}

16: \renewcommand{\thetheorem}{\arabic{section}.\arabic{theorem}}

17: \newtheorem{prop}[theorem]{Proposition}

18: \newtheorem{lemma}[theorem]{Lemma}

19: \newtheorem{cor}[theorem]{Corollary}

20: \newtheorem*{claim}{Claim}

21: \newtheorem{definition}[theorem]{Definition}

22: \newenvironment{poc}{\noindent \emph{Proof of Claim.}}{\vspace*{0.5\baselineskip}}

23: \renewcommand{\qedsymbol}{\vrule height12pt width7pt

24: depth0pt}

25: 

26: \allowdisplaybreaks

27: 

28: \newcommand{\cl}[2]{\ensuremath{\langle {#1} | {#2} \rangle}}

29: 

30: \def\ispreprint{}

31: 

32: \ifx\ispreprint\undefined

33: \def\baselinestretch{2.5}

34: \else

35: \fi

36: 

37: \author{Joshua Cooper\\ \small Department of Mathematics \\ \small University of California, San Diego, La Jolla, CA \\ \small \texttt{jcooper@math.ucsd.edu}}

38: \date{\today}

39: 

40: \begin{document}

41: \title{{\huge \bf Quasirandom Permutations}}

42: \ifx\ispreprint\undefined

43: \author{{\Large Joshua N.\ Cooper}\\

44: {\it Department of Mathematics, University of California}\\

45: {\it at San Diego, La Jolla, California}\\

46: E-mail: jcooper@math.ucsd.edu\\

47: Proposed running head: Quasirandom Permutations}

48: \else

49: \author{{\Large Joshua N.\ Cooper}\\

50: {\it Department of Mathematics, University of California}\\

51: {\it at San Diego, La Jolla, California}\\

52: E-mail: jcooper@math.ucsd.edu}

53: \fi

54: 

55: \date{July 1, 2002}

56: 

57: \maketitle

58: 

59: \ifx\ispreprint\undefined \pagebreak \else \fi

60: \renewcommand{\thesection}{\Roman{section}.}

61: \maketitle

62: 

63: \begin{abstract}

64: Chung and Graham [\ref{CG2}] define quasirandom subsets of

65: $\mathbb{Z}_n$ to be those with any one of a large collection of

66: equivalent random-like properties.  We weaken their definition and

67: call a subset of $\mathbb{Z}_n$ $\epsilon$-balanced if its

68: discrepancy on each interval is bounded by $\epsilon n$.  A

69: quasirandom permutation, then, is one which maps each interval to

70: a highly balanced set.  In the spirit of previous studies of

71: quasirandomness, we exhibit several random-like properties which

72: are equivalent to this one, including the property of containing

73: (approximately) the expected number of subsequences of each

74: order-type.  We provide a few applications of these results, present a construction for a family of

75: strongly quasirandom permutations, and prove that this construction

76: is essentially optimal, using a result of W. Schmidt on the

77: discrepancy of sequences of real numbers.

78: \end{abstract}

79: 

80: \ifx\ispreprint\undefined \indent {\it Key words:} quasirandom;

81: permutation; discrepancy; van der Corput

82: \\

83: \noindent {\it Correspondence Data:}

84: 

85: \indent Joshua N. Cooper \vspace{-.25in}

86: 

87: \indent UCSD Department of Mathematics \vspace{-.25in}

88: 

89: \indent 9500 Gilman Drive 0112 \vspace{-.25in}

90: 

91: \indent La Jolla, CA 92093-0112 \vspace{-.25in}

92: 

93: \indent jcooper@math.ucsd.edu

94: 

95: \pagebreak \else \fi

96: 

97: \section{Introduction}

98: 

99: In recent years, combinatorialists have been investigating several

100: realms of random-like -- ``quasirandom'' -- objects.  For a given

101: probability space $\mathcal X$, the basic idea is to choose some

102: collection of properties that large objects in $\mathcal X$ have

103: almost surely, and define a sequence $\{X_i\}_{i=1}^{\infty}

104: \subset \mathcal X$ to be {\it quasirandom} if $X_i$ has these

105: properties in the limit.  Often, this approach amounts to choosing

106: some random variables $\eta_j$ defined on $\mathcal X$ which tend

107: to their expected values almost surely as $|X| \rightarrow

108: \infty$, and defining $X_i$ to be quasirandom when

109: $(\eta_0(X_i),\eta_1(X_i),\ldots) \rightarrow

110: (\mathbb{E}\eta_0,\mathbb{E}\eta_1,\ldots)$ sufficiently quickly.

111: The resulting definitions are explored by finding many such

112: collections of properties and showing that quasirandomness with

113: respect to any one of them is equivalent to all the rest -- often

114: rather surprisingly, since the properties may appear completely

115: unrelated to one another. Quasirandom graphs, hypergraphs, set

116: systems, subsets of $\mathbb{Z}_n$, and tournaments have all been

117: examined in this way.  Quasirandom families of permutations have

118: been defined in [\ref{MMW1}], and Gowers [\ref{G1}] has used a careful

119: quantitative analysis of strongly quasirandom

120: (``$\alpha$-uniform'', in his terminology) subsets of

121: $\mathbb{Z}_n$ as an integral component of his remarkable new

122: proof of Szemer\'edi's Theorem.  Quasirandom objects

123: also have applications in algorithms as deterministic substitutes

124: for randomly generated objects, in addition to their purely

125: theoretical uses.  In fact, specific types of random-like

126: permutations have been used already in a number of contexts.

127: Lagarias [\ref{L1}] constructed random-like permutations of a

128: $d$-dimensional array of cells in order to solve a practical

129: memory-mapping problem, and Alon [\ref{A1}] used ``pseudo-random''

130: permutations to improve on the best known deterministic

131: maximum-flow algorithm of Goldberg and Tarjan.  Quasirandom sequences of reals are also fundamental to the extensively studied ``quasi-Monte Carlo'' methods of numerical analysis ([\ref{N1}]). In this paper, we

132: add (individual) permutations to the growing list of objects for

133: which a formal notion of quasirandomness has been defined.

134: 

135: In Section 2, we discuss the concept of $\epsilon$-balance, which

136: weakens the quasirandomness of Chung and Graham.  It is shown to

137: be equivalent to several ``types'' of quasirandomness for subsets

138: of $\mathbb{Z}_n$, including an infinite family of eigenvalue

139: bounds.  Section 3 is an excursion into the realm of subsequence

140: statistics of permutations, a subject that has generated a good

141: deal of interest recently (e.g., [\ref{AF1}] and [\ref{B1}]) --

142: and whose roots go back at least as far as 1935 ([\ref{ES1}]).  In

143: Section 4, quasirandom permutations are defined as those which map

144: intervals to uniformly balanced sets, and we prove that this

145: definition is equivalent to several other random-like conditions.  Two applications are given, including a proof that random permutations have small discrepancy.

146: Section 5 contains a construction for a family of strongly

147: quasirandom permutations that generalize the classical van der Corput sequences.  We show that this construction is

148: essentially optimal, using a result of Schmidt on the discrepancy

149: of sequences of real numbers.  Finally, Section 6 concludes with

150: some open problems and directions for future work.

151: 

152: \bigskip

153: \section{Balanced Sets}

154: \setcounter{theorem}{0}

155: 

156: Throughout the following, we consider permutations, i.e., elements

157: of $S_n$, as actions on $\mathbb{Z}_n$ as well as sequences of

158: numbers $(\sigma(0), \sigma(1), \ldots , \sigma(n-1))$ (``one-line

159: notation''). When an ordering on $\mathbb{Z}_n$ is used, we mean

160: the one inherited from $[0,n-1] \subset \mathbb{Z}$.  If $f_i$,

161: $i=1,2$, is a function from a totally ordered set $A$ to a totally

162: ordered set $B_i$, we say that $f_1$ and $f_2$ are

163: \textit{isomorphic} (and write $f_1 \sim f_2$) if, for any $a_1,

164: a_2 \in A$, $f_1(a_1) < f_1(a_2)$ iff $f_2(a_1) < f_2(a_2)$.  Note

165: that this definition still makes sense when $f_1$ and $f_2$ are

166: defined on different sets $A_1$ and $A_2$, so long as $|A_1| =

167: |A_2|$ is finite and we identify them via the unique

168: order-isomorphism between them. Then, if $\sigma \in S_n$ and

169: $\tau \in S_m$, $m \leq n$, we say that $\tau$ occurs in $\sigma$

170: at the set $A = \{ a_i \}_{i=1}^{m} \subset \mathbb{Z}_n$ whenever

171: $\sigma|_A \sim \tau$.  For each $A \subset \mathbb{Z}_n$ and

172: permutation $\tau$, we write ${\bf X}^{\tau}_A(\sigma)$ for the

173: indicator random variable of the event that $\tau$ occurs in

174: $\sigma$ at $A$, and we write ${\bf X}^{\tau}(\sigma)$ for the

175: random variable that counts the number of occurrences of $\tau$ in

176: $\sigma$, i.e., ${\bf X}^{\tau}(\sigma) = \sum_A {\bf

177: X}^{\tau}_A(\sigma)$ where $A$ ranges over all subsets of

178: $\mathbb{Z}_n$ of cardinality $m$.

179: 

180: For any subset $S \subset \mathbb{Z}_n$ (or $S \subset

181: \mathbb{Z}$), there is a minimal representation of $S$ as a union

182: of intervals.  We call these intervals the \textit{components} of

183: $S$ and denote the number of them by $c(S)$.  Also, we adopt the

184: convention that the symbols for a set and the characteristic

185: function of that set be the same, so, for example, $S(x) = 1$ if

186: $x \in S$ and $S(x) = 0$ if $x \not \in S$.  Finally, for any

187: function from $\mathbb{Z}_n$ to $\mathbb{C}$, we write

188: $\tilde{f}(k)$ for the $k^{\mbox{\scriptsize th}}$ Fourier

189: coefficient of $f$, defined by

190: $$

191: \tilde{f}(k) = \sum_{x \in \mathbb{Z}_n} f(x) e^{-2 \pi ikx/n}.

192: $$

193: A well known alternative definition of the Fourier coefficients of

194: a set $S$ is the spectrum of the circulant matrix $M_S$ whose

195: $(i,j)$ entry is $S(i+j)$.

196: 

197: \bigskip

198: 

199: One would expect random permutations to ``jumble'' the elements on

200: which it acts, i.e., there should be no correlation between

201: proximity in $\mathbb{Z}_n$ and proximity in the image.  We can

202: measure proximity by means of intervals: the elements of a small

203: interval are all ``close'' to one another.  Thus, if we define an

204: interval of $\mathbb{Z}_n$ to be the projection of any interval of

205: $\mathbb{Z}$, a permutation $\sigma \in S_n$ will be called

206: ``quasirandom'' if the intersection of any interval $I$ with the

207: image of any other interval $J$ under $\sigma$ has cardinality

208: approximately $|I||J|/n$, i.e., no interval contains much more or

209: less of the image of any other interval than one would expect if

210: $\sigma$ were chosen randomly.

211: 

212: Thus, for any two sets $S, T \subset \mathbb{Z}_n$ we define the

213: {\it discrepancy} of $S$ in $T$ as

214: $$

215: D_T(S) = \left | | S \cap T | - \frac{|S||T|}{n} \right |.

216: $$

217: Note that we may apply this definition to multisets $S$ and $T$,

218: and that it is symmetric in its arguments.  Before proceeding, we

219: present a simple lemma to the effect that $D$ is subadditive:

220: 

221: \begin{lemma} \label{subadd} If $S = A \cup B$, $A$ and $B$ disjoint, then $D_S(T) \leq D_A(T) + D_B(T)$.  If $T = C \cup D$, $C$ and $D$ disjoint, then $D_S(T) \leq D_S(C) + D_S(D)$.  That is, $D$ is subadditive in both of its arguments.

222: \end{lemma}

223: \begin{proof}

224: By the triangle inequality, we have

225: \begin{align*}

226: D_T(S) & = \left | \, | \, S \cap T | - \frac{|S||T|}{n} \right | \\

227: & = \left | \, | \, A \cap T | + | \, B \cap T | - \frac{|A||T|}{n} - \frac{|B||T|}{n} \right | \\

228: & \leq \left | D_T(A) \right | + \left | D_T(B) \right |.

229: \end{align*}

230: The other statement follows by symmetry.

231: \end{proof}

232: 

233: Define $D(S)$ to be the maximum of $D_J(S)$, taken over all

234: intervals $J \subset \mathbb{Z}_n$, and call a set $S \subset

235: \mathbb{Z}_n$ $\epsilon${\it -balanced} if $D(S) < \epsilon n$.

236: This definition of quasirandomness is implied by that of Chung and

237: Graham [\ref{CG2}], according to the next proposition.

238: 

239: \begin{prop} \label{weak} If, for all $T \subset \mathbb{Z}_n$ and all but $\epsilon n$ of $x \in \mathbb{Z}_n$, $D_{T+x}(S) < \epsilon n$, then, for all intervals $J \subset \mathbb{Z}_n$, $D_J(S) < 2 \epsilon n$.

240: \end{prop}

241: \begin{proof}

242: Suppose there exists an interval $J \subset \mathbb{Z}_n$ such that

243: $$

244: \left | \, |S \cap J| - \frac{|S||J|}{n} \right | \geq 2 \epsilon

245: n.

246: $$

247: Then, for all $x \in \mathbb{Z}_n$,

248: $$

249: \left | \, |S \cap (J+x)| - \frac{|S||J+x|}{n} \right | \geq \left

250: | \, |S \cap J| - \frac{|S||J|}{n} \right | - |x|.

251: $$

252: Therefore, for each $x$ with $|x| \leq \epsilon n$, $D_{J+x}(S)

253: \geq \epsilon n$.  Since there are at least $\epsilon n$ such

254: $x$'s, setting $T = J$ contradicts the hypothesis of the

255: proposition.

256: \end{proof}

257: 

258: It is easy to see that the set $S = \{2x \, | \, 0 \leq x \leq n-1

259: \} \subset \mathbb{Z}_{2n}$ is, for any $\epsilon > 0$ and

260: sufficiently large $n$, $\epsilon$-bounded.  However, $S \cap

261: (S+t)$ does not have cardinality approximately $|S|^2/n$ for

262: almost all $t$, i.e., it violates ``weak translation''. Therefore,

263: $\epsilon$-boundedness is strictly weaker than quasirandomness in

264: the sense of [\ref{CG2}].

265: 

266: We use the convention that when ``little oh'' notation is used,

267: convergence in $n$ alone is intended.  (That is, the convergence

268: is uniform in any other quantities involved.)  The following is

269: the main result of this section.

270: 

271: \begin{theorem} \label{balance} For $r \in \mathbb{Z}_n$, we define $|r|$ to be the absolute value of the unique representative of $\, r$ from the interval $(-n/2,n/2]$.  Then, for any sequence of subsets $S \subset \mathbb{Z}_n$ and choice of $\alpha > 0$, the following are equivalent: \\ [.06in]

272: 

273: \noindent \begin{tabular}{p{.4in}p{4.6in}}

274: {\bf [B]}  & (Balance) $D(S) = o(n)$. \\

275: \end{tabular}

276: 

277: \noindent \begin{tabular}{p{.4in}p{4.6in}}

278: {\bf [PB]} & (Piecewise Balance) For any subset $T \subset \mathbb{Z}_n$, $D_T(S) = o(nc(T))$, where $c(T)$ denotes the number of components of $T$. \\

279: \end{tabular}

280: 

281: \noindent \begin{tabular}{p{.4in}p{4.6in}}

282: {\bf [MB]} & (Multiple Balance) Let $kS$ denote the multiset $\{ks|s \in S\}$.  Then, for any $k \in \mathbb{Z}_n \! \setminus \! \{0\}$, $D(kS) = o(n|k|)$. \\

283: \end{tabular}

284: 

285: \noindent \begin{tabular}{p{.4in}p{4.6in}}

286: {\bf [E($\frac{1}{2}$)]} & (Eigenvalue Bound $\frac{1}{2}$) For all nonzero $k \in \mathbb{Z}_n$, $\tilde{S}(k) = o(n|k|^{1/2})$. \\

287: \end{tabular}

288: 

289: \noindent \begin{tabular}{p{.4in}p{4.6in}}

290: {\bf [E($\alpha$)]} & (Eigenvalue Bound $\alpha$) For all nonzero $k \in \mathbb{Z}_n$, $\tilde{S}(k) = o(n|k|^{\alpha})$. \\

291: \end{tabular}

292: 

293: \noindent \begin{tabular}{p{.4in}p{4.6in}}

294: {\bf [S]} & (Sum) $ \sum_{r \neq 0} \left( |\tilde{S}(k)|/|k| \right )^2 = o(n^2). $ \\

295: \end{tabular}

296: 

297: \noindent \begin{tabular}{p{.4in}p{4.6in}} {\bf [T]} &

298: (Translation) For any interval $J$,

299: $$ \sum_{k \in \mathbb{Z}_n} \left ( |S \cap (J+k)| - \frac{|S||J|}{n} \right )^2 = o(n^3). $$

300: \end{tabular}

301: \end{theorem}

302: 

303: \ifx\ispreprint\undefined

304: \else

305: \begin{center}

306: \begin{figure}[ht]

307: \hspace{.7in} \input{figa.pstex_t} \caption{Diagram of

308: implications for Theorem \ref{balance}.}

309: \end{figure}

310: \end{center}

311: \fi

312: 

313: We will show that {\bf [B]} $\Rightarrow$ {\bf [PB]} $\Rightarrow$

314: {\bf [MB]} $\Rightarrow$ {\bf [E($\frac{1}{2}$)]} $\Rightarrow$

315: {\bf [E($\alpha$)]} $\Rightarrow$ {\bf [S]} $\Rightarrow$ {\bf

316: [T]} $\Rightarrow$ {\bf [B]}.  In each case, a statement involving

317: some $\epsilon$ is shown to imply the next for some $f(\epsilon)$,

318: where $f$ is a function which tends to zero as its argument does.

319: For example, Proposition \ref{division} below states that if

320: $D_T(S) < \epsilon n c(T)$ for all $T$, then $D_T(kS) < 2 \epsilon

321: n |k|$ for all k, so that $f(\epsilon)=2\epsilon$.  It appears to

322: be theoretically useful to track what happens to $\epsilon$ as we

323: pass through each implication -- see, for example, [\ref{G1}].

324: Thus, we include Figure 1 as an accompaniment to Theorem

325: \ref{balance}.  (Note that, by the proof of Proposition

326: \ref{MB2EA}, Figure 1 is only valid for $\epsilon < \pi / 8$,

327: though this is hardly a significant restriction.)  The shortcut

328: edge from {\bf [E($\frac{1}{2}$)]} to {\bf [S]} is given to

329: illustrate the (best possible) choice of $\alpha = 1/4$ in {\bf

330: [E($\alpha$)]}, and the circular arrow represents one complete

331: traversal of the cycle of implications, including the shortcut

332: edge.

333: 

334: Theorem \ref{balance} is proven in pieces, beginning with the

335: following proposition.

336: 

337: \begin{prop} {\bf [B]} $\Rightarrow$ {\bf [PB]} $\Rightarrow$ {\bf [MB]}. \label{division}

338: \end{prop}

339: \begin{proof}

340: Suppose that $D(S) < \epsilon n$.  Then, by Lemma \ref{subadd},

341: for any $T$, $D_T(S) \leq \sum D_{T_i}(S)$, where the sum is over

342: the components of $T$.  Thus, $D_T(S) < \epsilon n \, c(T)$, and

343: {\bf [B]} $\Rightarrow$ {\bf [PB]}.

344: 

345: Now, suppose {\bf [PB]} holds for $S$.  Note that, for a given $k

346: \in \mathbb{Z}_n \setminus \{0\}$ and interval $J$, the set

347: $J^\prime$ of elements $x \in \mathbb{Z}_n$ such that $kx \in J$

348: has at most $|k|$ components.  Let $J_i$ be the set of integer

349: points (viewed as elements of $\mathbb{Z}_n$) lying in $[a/k,b/k]

350: + in/k$, so that $J^\prime = \bigcup_i J_i$.  Then the cardinality

351: of $J_i$ is off from $|J|/k$ by at most $1$.  By {\bf [PB]} and

352: the triangle inequality,

353: \begin{align*}

354: D_J(kS) &= \left | \, |k(S-t) \cap J| - \frac{|I||J|}{n} \right | \\

355: & \leq \left | \, \sum_i |S \cap J_i| - \sum_i \frac{|I||J_i|}{n} \right | + \frac{|I|}{n} \left | \, \sum_i |J_i| - |J| \right | \\

356: & < \epsilon n |k| + |k| \frac{|I|}{n} \leq 2 \epsilon n |k|.

357: \end{align*}

358: since, trivially, $\epsilon \geq n^{-1}$.

359: \end{proof}

360: 

361: Now, we wish to show that Multiple Boundedness implies the first

362: eigenvalue bound.  The basic idea is to imbed the elements of $S$

363: into the unit circle via the exponential map, and then show that a

364: great deal of cancellation happens because of the relatively

365: uniform distribution of elements of $S$.

366: 

367: \begin{prop} {\bf [MB]} $\Rightarrow$ {\bf [E($\frac{1}{2}$)]}. \label{MB2EA}

368: \end{prop}

369: \begin{proof}

370: Let $\omega = e^{2 \pi i / n}$ and $J_m^j = [\frac{nj}{m},\frac{n(j+1)}{m})$, and let $\epsilon$ be the bound on $(nk)^{-1} D(k S)$.  Recall that $\epsilon \geq n^{-1}$.  First we prove the following:

371: \begin{claim} Let $m$ and $j$ be positive integers with $0 \leq j < m$, and $m \geq 2$.  If we define the multiset $S_j = k S \cap J_m^j$ and let $\gamma_j = \omega^{-n (j+1/2)/m}$, then

372: $$

373: \left | \, \sum_{x \in S_j} \omega^{-x} - \frac{|S|}{m} \gamma_j

374: \right | < \frac{\pi |S|}{m^2} + 2 \epsilon |k|n

375: $$

376: \end{claim}

377: 

378: \begin{poc} We may write the left-hand side of the above expression as

379: \begin{align*}

380: \left | \, \sum_{x \in S_j} \omega^{-x} - \frac{|S|}{m} \gamma_j \right | &= \left | \, \sum_{x \in S_j} (\omega^{-x} - \frac{|S|}{m|S_j|} \gamma_j ) \right | \\

381: & \leq \frac{|S|}{m|S_j|} \left | \, \sum_{x \in S_j} (\omega^{-x} - \gamma_j ) \right | + \left | \, \sum_{x \in S_j} \omega^{-x} (1 - \frac{|S|}{m|S_j|}) \right | \\

382: & \leq \frac{|S|}{m|S_j|} \sum_{x \in S_j} \left | \, (\omega^{-x}

383: - \gamma_j ) \right | + \sum_{x \in S_j} \left | \, \omega^{-x} (1

384: - \frac{|S|}{m|S_j|}) \right |

385: \end{align*}

386: Now, for $x \in S_j$,

387: $$

388: | \, \omega^{-x} - \gamma_j | \leq | \, \omega^{-n j / m} -

389: \omega^{-n (j+1/2) / m} | \leq \frac{n/2}{m} \cdot \frac{2 \pi}{n}

390: = \frac{\pi}{m}

391: $$

392: Plugging this expression in and applying {\bf [MB]}, we have

393: \begin{align*}

394: \left | \, \sum_{x \in S_j} \omega^{-x} - \frac{|S|}{m} \gamma_j \right | & \leq \frac{|S|}{m|S_j|} \cdot |S_j| \cdot \frac{\pi}{m} + \left | |S_j| - \frac{|S|}{m} \right | \\

395: & \leq \frac{\pi |S|}{m^2} + \left | |kS \cap J_m^j| - \frac{|I||J_m^j|}{n} \right | + \frac{|I|}{n} \left | \frac{n}{m} - |J_m^j| \right | \\

396: & < \frac{\pi |S|}{m^2} + \epsilon |k| n + \frac{|I|}{n} \\

397: & \leq \frac{\pi |S|}{m^2} + 2 \epsilon |k| n

398: \end{align*}

399: thus, proving the claim.

400: \end{poc}

401: 

402: \medskip

403: 

404: If we sum over all $j \in [0, m-1)$,

405: \begin{align*}

406: \left | \, \sum_{x \in S} \omega^{-kx} \right | &= \left | \, \sum_{j=0}^{m-1} \sum_{x \in S_j} \omega^{-x} \right | \\

407: & \leq \left | \, \sum_{j=0}^{m-1} \frac{|S|}{m} \gamma_j \right | + \sum_{j=0}^{m-1} \left | \, \sum_{x \in S_j} \omega^{-x} - \frac{|S|}{m} \gamma_j \right | \\

408: & < 0 + \frac{\pi |S|}{m} + 2 \epsilon |k| nm

409: \end{align*}

410: if we assume that $m \geq 2$.  Thus, letting $m = \left \lfloor

411: \left ( \frac{\pi |S|}{2\epsilon |k|n} \right )^{1/2} \right

412: \rfloor$, we have

413: $$

414: \left | \, \sum_{x \in S} \omega^{-kx} \right | < \sqrt{18 \pi

415: \epsilon n|k||S|} \leq n \sqrt{18 \pi \epsilon |k|}

416: $$

417: unless $m < 2$, i.e., $ \epsilon > \frac{\pi}{8} $, which is

418: eventually impossible, given {\bf [MB]}.  We may therefore

419: conclude that $| \tilde{S}(k) | = o(n|k|^{1/2})$.

420: \end{proof}

421: 

422: A small improvement to the constant in the bound above is possible

423: by letting $m$ be rational, instead of integral. However, doing so

424: adds some complexity to the proof without making any significant

425: improvements.

426: 

427: Before we proceed with the next implication, the following lemma

428: will be necessary.  It implies, surprisingly, that {\bf

429: [E($\alpha$)]} is equivalent to {\bf [E($\beta$)]} for all

430: $\alpha$ and $\beta$.

431: 

432: \begin{lemma} \label{EA2EB} {\bf [E($\alpha$)]} implies {\bf [E($\beta$)]} for any $\alpha, \beta > 0$.

433: \end{lemma}

434: \begin{proof}

435: Let $M = \left \lceil \frac{\alpha}{\beta} \right \rceil$, and

436: assume {\bf [E($\alpha$)]}.  Then

437: $$

438: \left | \tilde{S}(k) \right |^{M} = \left | \sum_{t \in

439: \mathbb{Z}_n} S(t) \omega^{-kt} \right |^{M} = \left |

440: \sum_{t_1,\ldots,t_{M}} \left [ \prod_{j=1}^{M} S(t_j) \right ]

441: \omega^{-k \sum_{i=1}^M t_i} \right |

442: $$

443: Letting $u = \sum_{i=2}^M t_i$, we have

444: \begin{align*}

445: \left | \tilde{S}(k) \right |^{M} &= \left | \sum_{t_2,\ldots,t_{M}} \left [ \prod_{j=2}^{M} S(t_j) \right ] \sum_{t_1} S(t_1) \omega^{-k (t_1 + u)} \right | \\

446: & \leq \sum_{t_2,\ldots,t_{M}} \left [ \prod_{j=2}^{M} S(t_j) \right ] \left | \sum_{t_1} S(t_1) \omega^{-kt_1} \right | \\

447: &= \sum_{t_2,\ldots,t_{M}} \left [ \prod_{j=2}^{M} S(t_j) \right ] \left | \tilde{S}(k) \right | \\

448: &< \sum_{t_2,\ldots,t_{M}} \left [ \prod_{j=2}^{M} S(t_j) \right ] \epsilon n |k|^\alpha \\

449: &= |S|^{M-1} \epsilon n |k|^\alpha \leq \epsilon n^{M} |k|^\alpha.

450: \end{align*}

451: Thus, taking the $M^{\mbox{\scriptsize th}}$ root of both sides,

452: we have

453: $$

454: \left | \tilde{S}(k) \right | < \epsilon^{1/M} n |k|^{\alpha/M}

455: \leq \epsilon^{\left \lceil \alpha/\beta \right \rceil^{-1}} n

456: |k|^{\beta}.

457: $$

458: \end{proof}

459: 

460: The following corollary is actually what is needed for Theorem

461: \ref{balance}.

462: 

463: \begin{cor} {\bf [E($\frac{1}{2}$)]} $\Rightarrow$ {\bf [E($\alpha$)]}.

464: \end{cor}

465: 

466: Note that, to proceed with the next proposition, $\alpha = 1/2$

467: would not quite be enough -- we {\it have} to reduce it by a bit

468: with Proposition \ref{EA2EB}.

469: 

470: \begin{prop} {\bf [E($\alpha$)]} $\Rightarrow$ {\bf [S]}.

471: \end{prop}

472: \begin{proof}

473: By Proposition \ref{EA2EB}, we know that $|\tilde{S}(k)| <

474: \epsilon^{\left \lceil 4\alpha \right \rceil^{-1}} n |k|^{1/4}$

475: for all $k \neq 0$. Then

476: $$

477: \sum_{r \neq 0} \left( \frac{|\tilde{S}(k)|}{|k|} \right )^2 <

478: \sum_{k \neq 0} \left( \frac{\epsilon^{\left \lceil 4\alpha \right

479: \rceil^{-1}} n |k|^{1/4}}{|k|} \right )^2 \leq \epsilon^{2 \left

480: \lceil 4\alpha \right \rceil^{-1}} n^2 \sum_{k \neq 0} |k|^{-3/2}

481: < 6 \epsilon^{2 \left \lceil 4\alpha \right \rceil^{-1}} n^2.

482: $$

483: where we have used the approximation $|\zeta(s)| < (Re(s)-1)^{-1}

484: + 1$ for $s$ with $Re(s) > 1$.

485: \end{proof}

486: 

487: We now write a cyclic sum in terms of Fourier coefficients. A proof of the following standard lemma is included for the sake of completeness.

488: 

489: \begin{lemma} \label{intb} If $J$ is an interval of $\mathbb{Z}_n$, then $\tilde{J}(k) \leq \frac{n}{2|k|}$.

490: \end{lemma}

491: \begin{proof}

492: We may write the magnitude of the $k^{\mbox{\scriptsize th}}$

493: Fourier coefficient of $J = [a+1,a+M]$ as

494: \begin{align*}

495: |\tilde{J}(k)| &= | \sum_x J(x) \omega^{-kx} | = | \sum_{x=a}^b \omega^{-kx} | = | \sum_{x=1}^M \omega^{-kx} | \\

496: & = \frac{|\omega^{-kM} - 1|}{|\omega^{-k} - 1|} \leq \frac{2}{4

497: |k| / n} = \frac{n}{2|k|}

498: \end{align*}

499: since $|e^{i \theta} - 1| \geq \frac{2 |\theta|}{\pi}$ for all

500: $\theta$.

501: \end{proof}

502: 

503: \begin{prop} {\bf [S]} $\Rightarrow$ {\bf [T]}.

504: \end{prop}

505: \begin{proof}

506: Assume that $\sum_{k \neq 0} \left( \frac{|\tilde{S}(k)|}{|k|}

507: \right )^2 < \epsilon n^2$.  We may write the ``translation'' sum

508: as

509: \begin{equation} \label{eqa1}

510: \sum_{k \in \mathbb{Z}_n} \left ( |S \cap (J+k)| -

511: \frac{|S||J|}{n} \right )^2 = \sum_{k \in \mathbb{Z}_n} |S \cap

512: (J+k)|^2 - \frac{|S|^2|J|^2}{n}

513: \end{equation}

514: Recall that $M_S$ is the $n \times n$ matrix whose $(i,j)$ entry

515: is $S(i+j)$.  Letting {\boldmath $v$} be the vector

516: $(J(0),J(1),\ldots)$, we find that $M_S \mbox{\boldmath $v$}$ is

517: the vector whose $k^{\mbox{\scriptsize th}}$ entry is $|I \cap

518: (J+k)|$.  Therefore, letting $\phi_k =

519: (1,\omega^{k},\omega^{2k},\ldots)$ be the $k^{\mbox{\scriptsize

520: th}}$ eigenvector of $M_S$,

521: \begin{align*}

522: \sum_{k \in \mathbb{Z}_n} |S \cap (J+k)|^2 &= | M_S \mbox{\boldmath $v$} |^2 = | M_S \sum_k \frac{<\mbox{\boldmath $v$},\phi_k>}{|\phi_k|^2}\,\phi_k |^2 \\

523: &= \sum_k | \tilde{S}(k)^2 \frac{<\mbox{\boldmath $v$},\phi_k>^2}{|\phi_k|^2} | \\

524: &= \sum_{k \neq 0} \left | \tilde{S}(k)

525: \frac{\tilde{J}(-k)}{\sqrt{n}} \right |^2 + \frac{|S|^2|J|^2}{n}

526: \end{align*}

527: Applying this equality, property {\bf [S]}, and Lemma \ref{intb}

528: to Equation \ref{eqa1},

529: \begin{align*}

530: \sum_{k \in \mathbb{Z}_n} \left ( |S \cap (J+k)| - \frac{|S||J|}{n} \right )^2 &= \sum_{k \neq 0} \left | \tilde{S}(k) \frac{\tilde{J}(-k)}{\sqrt{n}} \right |^2 \\

531: & \leq \frac{n}{4} \sum_{k \neq 0} \left | \frac{\tilde{S}(k)}{|k|} \right |^2 \\

532: & < \frac{\epsilon^2}{4} n^3.

533: \end{align*}

534: \end{proof}

535: 

536: To complete the circle of implications and finish the proof of

537: Theorem \ref{balance}, we show that $\epsilon$-boundedness is

538: implied by the ``translation'' property.

539: 

540: \begin{prop} {\bf [T]} $\Rightarrow$ {\bf [B]}

541: \end{prop}

542: \begin{proof}

543: Suppose that, for some interval $J \subset \mathbb{Z}_n$,

544: $$

545: \left | |S \cap J| - \frac{|S||J|}{n} \right | \geq 2

546: \epsilon^{1/3} n

547: $$

548: Then, following the line of argument given for Proposition

549: \ref{weak}, we may conclude that

550: $$

551: \left | |S \cap (J+k)| - \frac{|S||J|}{n} \right | \geq

552: \epsilon^{1/3} n

553: $$

554: whenever $|k| \leq \epsilon^{1/3} n$.  Since there are at least

555: $\epsilon^{1/3} n $ such $k$'s,

556: $$

557: \sum_k \left ||S \cap (J+k)| - \frac{|S||J|}{n} \right |^2 \geq

558: \epsilon^{1/3} n \cdot \epsilon^{2/3} n^2 = \epsilon n^3

559: $$

560: contradicting {\bf [T]}.

561: \end{proof}

562: 

563: \bigskip

564: \section{Statistics of Sub-Permutations}

565: \setcounter{theorem}{0}

566: 

567: Before we formally define quasirandom permutations, an excursion

568: into the realm of subsequence statistics is necessary. We wish to

569: relate ${\bf X}^{\tau}(\sigma)$, for $\tau \in S_m$, to the

570: quantities ${\bf X}^{\tau^\prime}(\sigma)$, with $\tau^\prime \in

571: S_{m+1}$, by counting occurrences of $\tau$ inside each occurrence

572: of $\tau^\prime$.  Define ${\bf v}_m(\sigma) \in \mathbb{Z}^{S_m}$

573: to be the vector whose $\tau$ component is ${\bf

574: X}^{\tau}(\sigma)$, and write ${\bf \tilde{v}}_m(\sigma)$ for the

575: vector

576: $$

577: {\bf v}_m(\sigma) - \mathbb{E}{\bf v}_m = {\bf v}_m(\sigma) -

578: \frac{\bf \hat{1}}{m!} \binom{n}{m}

579: $$

580: Also, let $B_m$ be the $m!$ by $(m+1)!$ matrix whose

581: $(\tau,\tau^\prime)$ entry is ${\bf X}^{\tau}(\tau^\prime)$ for

582: $\tau \in S_m$ and $\tau^\prime \in S_{m+1}$, and define $A_m$ to

583: be $B^*_m B_m$.

584: 

585: \begin{prop} For any $\sigma \in S_m$,

586: \begin{equation} \label{mainform}

587: (n-m)^2 {| {\bf \tilde{v}}_m(\sigma) |}^2 = {{\bf

588: \tilde{v}}_{m+1}(\sigma)}^* A_m {\bf \tilde{v}}_{m+1}(\sigma)

589: \end{equation}

590: \end{prop}

591: \begin{proof}

592: Let $\Gamma$ be the set of pairs $(U,V)$, with $U \subset V

593: \subset \mathbb{Z}_n$, $|U|=m$, $|V|=m+1$, and $\sigma_U \sim

594: \tau$.  Then, conditioning on the order-type of $U$ yields

595: $$

596: |\Gamma| = \sum_{\tau^\prime \in S_{m+1}} {\bf

597: X}^{\tau}(\tau^\prime) {\bf X}^{\tau^\prime}(\sigma)

598: $$

599: because each set $U$ contributes ${\bf X}^{\tau}(\sigma|_U)$ to

600: $|\Gamma|$.  If we instead condition on $V$ itself, then

601: $$

602: |\Gamma| = (n - m) {\bf X}^{\tau}(\sigma)

603: $$

604: because each subset $V$ is contained in exactly $n-m$ supersets

605: $U$.  We may therefore write

606: $$

607: \sum_{\tau^\prime \in S_{m+1}} {\bf X}^{\tau}(\tau^\prime) {\bf

608: X}^{\tau^\prime}(\sigma) = (n - m) {\bf X}^{\tau}(\sigma),

609: $$

610: i.e., $(n-m) {\bf v}_m(\sigma) = B_m {\bf v}_{m+1}(\sigma)$.  The

611: desired result then follows by linearity of expectation.

612: \end{proof}

613: 

614: Now that a numerical relationship between subsequences of length

615: $m+1$ and subsequences of length $m$ has been established, we need

616: a bound on the eigenvalues of $A_m$.

617: 

618: \begin{prop} \label{propsums} The following hold for all $m \geq 1$:

619: \begin{enumerate}

620: \item The column sums of $B_m$ are equal to $m+1$.

621: \item The row sums of $B_m$ are equal to $(m+1)^2$.

622: \item The row (and column) sums of $A_m$ are equal to $(m+1)^3$.

623: \end{enumerate}

624: \end{prop}

625: \begin{proof}

626: The proofs are all straightforward manipulations.

627: 

628: \begin{enumerate}

629: \item Let $b_{\tau \tau^\prime}$ denote the $(\tau,\tau^\prime)$ entry of the matrix $B_m$.  Then, denoting the set of subsets of $\mathbb{Z}_{m+1}$ of cardinality $m$ by $P_m$,

630: \begin{align*}

631: \sum_{\tau \in S_{m}} b_{\tau \tau^\prime} & = \sum_{\tau \in S_{m}} {\bf X}^{\tau}(\tau^\prime) = \sum_{\tau \in S_{m}} \sum_{A \in \mathcal{P}_m} {\bf X}^{\tau}_A(\tau^\prime) \\

632: & = \sum_{A \in \mathcal{P}_m} \sum_{\tau \in S_{m}} {\bf

633: X}^{\tau}_A(\tau^\prime) = \sum_{A \in \mathcal{P}_m} 1 = m+1.

634: \end{align*}

635: \item Note that, for a set $A \in P_m$, a permutation $\tau^\prime \in S_{m+1}$ is uniquely determined by its restriction to $A$.  Therefore, for a given $\tau \in S_m$, the number of $\tau^\prime \in S_{m+1}$ such that $\tau^\prime|_A \sim \tau$ is equal to the number of possible sets $\tau^\prime(A)$, i.e.,

636: $$

637: \sum_{\tau^\prime \in S_{m+1}} {\bf X}^{\tau}_A(\tau^\prime) =

638: m+1.

639: $$

640: Now, we sum over the first index:

641: \begin{align*}

642: \sum_{\tau^\prime \in S_{m+1}} b_{\tau \tau^\prime} & = \sum_{\tau^\prime \in S_{m+1}} {\bf X}^{\tau}(\tau^\prime) = \sum_{\tau^\prime \in S_{m+1}} \sum_{A \in \mathcal{P}_m} {\bf X}^{\tau}_A(\tau^\prime) \\

643: & = \sum_{A \in \mathcal{P}_m} \sum_{\tau^\prime \in S_{m+1}} {\bf

644: X}^{\tau}_A(\tau^\prime) = \sum_{A \in \mathcal{P}_m} (m+1) =

645: (m+1)^2.

646: \end{align*}

647: \item Let $a_{\tau \tau^\prime}$ denote the $(\tau,\tau^\prime)$ entry of the matrix $A_m$.  Since $A_m$ is symmetric, we need only show the result for column sums.

648: \begin{align*}

649: \sum_{\tau \in S_{m}} a_{\tau \tau^\prime} & = \sum_{\tau \in S_{m}} \sum_{\tau^{\prime \prime} \in S_{m+1}} b_{\tau \tau^{\prime \prime}} b_{\tau^\prime \tau^{\prime \prime}} = \sum_{\tau^{\prime \prime} \in S_{m+1}} (\sum_{\tau \in S_{m}} b_{\tau \tau^{\prime \prime}}) b_{\tau^\prime \tau^{\prime \prime}} \\

650: & = \sum_{\tau^{\prime \prime} \in S_{m+1}} (m+1) b_{\tau^\prime

651: \tau^{\prime \prime}} = (m+1)^3.

652: \end{align*}

653: where the third equality follows from part (1) and the fourth from

654: part (2).

655: \end{enumerate}

656: \end{proof}

657: 

658: \begin{cor} \label{eigcor} ${\bf \hat{1}}_m$ is an eigenvector of $A_m$ with eigenvalue $(m+1)^3$.

659: \end{cor}

660: 

661: \begin{proof}

662: By Proposition \ref{propsums},

663: \begin{align*}

664: A_m {\bf \hat{1}}_m = (m+1)^3 {\bf \hat{1}}_m.

665: \end{align*}

666: \end{proof}

667: 

668: \begin{prop} $(m+1)^3$ is the largest eigenvalue of $A_m$.

669: \end{prop}

670: 

671: \begin{proof}

672: By Corollary \ref{eigcor} and the Perron-Frobenius Theorem, we

673: need only show that $A_m$ is irreducible.  Consider the weighted

674: bipartite graph $G_m$ on the sets $S_m$ and $S_{m+1}$, where

675: $\sigma \in S_{m}$ is connected by an edge to $\tau \in S_{m+1}$

676: with weight ${\bf X}^\sigma (\tau)$.  (In particular, there is an

677: edge connecting $\sigma$ to $\tau$ iff $\sigma$ occurs in $\tau$.)

678: Then the adjacency matrix of $G_m$ is $B_m$, and the entries of

679: $A_m$ represent sums of weighted length-$2$ paths from $S_{m+1}$

680: to itself.  If $G_m$ is connected, then $A_m$ is irreducible.  To

681: establish connectivity, we show that there is a path from every

682: permutation $\tau \in S_{m+1}$ to the identity element of

683: $S_{m+1}$ in $G_m$.

684: 

685: \begin{claim} For each $k$, $0 \leq k \leq m$, there is a path in $G_m$ from each permutation $\tau \in S_{m+1}$ to some permutation $\tau^\prime \in S_{m+1}$ such that $\tau^\prime (i) = i$ whenever $0 \leq i \leq k$.

686: \end{claim}

687: 

688: \begin{poc}

689: We proceed by induction.  Suppose the claim is true for $k$, and

690: let $\tau$ be any element of $S_{m+1}$.  The inductive hypothesis

691: supplies us with a path from $\tau$ to a $\tau^\prime$ such that

692: $\tau^\prime (i) = i$ whenever $1 \leq i \leq k$.  Let $\sigma$ be

693: the unique permutation in $S_m$ such that $\sigma \sim \tau^\prime

694: |_{\mathbb{Z}_m \setminus \{k+1\}}$.  Then define $\tau^{\prime

695: \prime}$ as follows:

696: $$

697:   \tau^{\prime \prime}(i) = \left\{

698:                \begin{array}{ll}

699:                i & \text{if $0 \leq i \leq k+1$}, \\

700:                \sigma(i-1) + 1 & \text{if $k+1 < i \leq m$}

701:                \end{array}

702:                \right.

703: $$

704: It is easy to check that $\tau^{\prime \prime}$ is actually an

705: element of $S_{m+1}$.  Furthermore, $\tau^\prime|_{\mathbb{Z}_m

706: \setminus \{k+1\}} \sim \sigma \sim \tau^{\prime \prime}

707: |_{\mathbb{Z}_m \setminus \{k+1\}}$, so there is a path in $G_m$

708: from $\tau$ to a permutation which agrees with the identity on

709: $[k+1]$.

710: \end{poc}

711: 

712: Since every vertex of $S_{m}$ is connected to \emph{some} vertex

713: of $S_{m+1}$, and they are all connected to the identity, $G_m$ is

714: connected and $A_m$ is irreducible.

715: \end{proof}

716: 

717: We apply this result and the Courant-Fischer Theorem to Equation

718: (\ref{mainform}):

719: $$

720: (n-m)^2 {| {\bf \tilde{v}}_m(\sigma) |}^2 \leq \lambda_{max}(A_m)

721: | {\bf \tilde{v}}_{m+1}(\sigma) |^2 = (m+1)^3 | {\bf

722: \tilde{v}}_{m+1}(\sigma) |^2

723: $$

724: Thus, we have

725: \begin{cor} \label{down} For any $\sigma \in S_n$,

726: $$

727: | {\bf \tilde{v}}_{m}(\sigma) |^2 \leq \frac{(m+1)^3}{(n-m)^2} {|

728: {\bf \tilde{v}}_{m+1}(\sigma) |^2 }

729: $$

730: \end{cor}

731: 

732: Although Corollary \ref{down} is all we will need to use later, we

733: include here a short proof that $B_m$ has maximal rank.

734: 

735: \begin{definition}

736: For a permutation $\sigma \in S_{m}$, write $\sigma^\circ$ for the

737: permutation in $S_{m+1}$ such that $\sigma^\circ(0) = 0$ and

738: $\sigma^\circ(i) = \sigma(i-1) + 1$ for $1 \leq i \leq m-1$.

739: \end{definition}

740: 

741: Recall that, for two permutations $\sigma$ and $\sigma^\prime$ on

742: the totally ordered sets $S$ and $S^\prime$, respectively, where

743: $S$ and $S^\prime$ have the same cardinality, we may compare

744: $\sigma$ and $\sigma^\prime$ in the lexicographic order by

745: identifying the $i^{th}$ elements of $S$ and $S^\prime$ for each

746: $i$.

747: 

748: \begin{theorem} \label{firstoccur}

749: For a given $\sigma \in S_{m}$ the lexicographically least

750: permutation $\tau \in S_{m+1}$ such that ${\bf X}^{\sigma}(\tau)$

751: is nonzero is $\tau = \sigma^\circ$.

752: \end{theorem}

753: \begin{proof}

754: First, note that ${\bf X}^{\sigma}(\sigma^\circ) > 0$, since

755: $\sigma \sim \sigma^\circ |_S$, where $S = {\mathbb{Z}_{m+1}

756: \setminus \{0\}}$.  To show that ${\bf X}^{\sigma}(\tau) = 0$ for

757: all $\tau < \sigma^\circ$, we proceed by induction.  For $m=1$,

758: the result is trivial: $(0)$ occurs in $(01)$, and in no earlier

759: permutation, since $(01)$ is lexicographically first.  Now suppose

760: the result is true for $m-1$, but there is some $\sigma \in S_{m}$

761: and  $\tau \in S_{m+1}$ such that ${\bf X}^{\sigma}(\tau) \neq 0$

762: and $\tau < \sigma^\circ$.  We know that $\tau(0)$ must equal $0$,

763: since $\sigma^\circ$ does, and $\tau$ precedes $\sigma^\circ$.

764: Therefore, $\tau < \sigma^\circ \Rightarrow \tau |_S <

765: \sigma^\circ |_S$.  Since $\sigma^\circ |_S$ is isomorphic to

766: $\sigma$, $\tau |_S$ is lexicographically precedent to $\sigma$,

767: so $\sigma \not \sim \tau |_S$.  Thus, $\sigma \sim \tau |_T$ for

768: some $T \neq S$, i.e., a $T \in P_m$ which includes $0$.

769: Restricting both of these permutations to all but the first

770: element, we have $\sigma|_U \sim \tau|_{T \cap S} \Rightarrow {\bf

771: X}^{\sigma |_U}(\tau |_S) > 0$.  On the other hand, $\sigma \sim

772: \tau |_T \Rightarrow \sigma(0)=0$, since $0 \in T$, so the fact

773: that $\tau|_S < \sigma^\circ |_S \sim \sigma \sim

774: (\sigma|_U)^\circ$ implies that $\sigma|_U$ occurs in a

775: permutation lexicographically precedent to $(\sigma|_U)^\circ$,

776: contradicting the inductive hypothesis.

777: \end{proof}

778: 

779: \begin{cor} \label{bcor}

780: $\text{rank}(B_m) = m!$.

781: \end{cor}

782: 

783: \begin{proof}

784: If we order the columns and rows of $B_m$ by the lexicographic

785: order on their indices, then Theorem \ref{firstoccur} implies that

786: the first nonzero entry in the $i^{th}$ row occurs in the $i^{th}$

787: column, since $\sigma < \tau$ iff $\sigma^\circ < \tau^\circ$, and

788: the permutations of the form $\sigma^\circ$ precede all others.

789: \end{proof}

790: 

791: \bigskip

792: \section{Quasirandom Permutations}

793: \setcounter{theorem}{0}

794: 

795: In this section, we discuss several equivalent formulations of

796: quasirandom permutations.  The central definition is, roughly,

797: that a quasirandom permutation is one which sends each interval to

798: a highly balanced set.  Thus, we will write $D(\sigma)$ for

799: $\max(D_J(\sigma(I))$, where the maximum is taken over all

800: intervals $I$ and $J$, and a sequence of permutations ${\sigma_j}$

801: will be called {\it quasirandom} if $D(\sigma_j) = o(n)$.  The

802: following is the main result of this section.

803: 

804: \begin{theorem} \label{qrp} For any sequence of permutations $\sigma \in S_n$ and integer $m \geq 2$ with $n > m$, the following are equivalent: \\ [.06in]

805: \noindent \begin{tabular}{p{.4in}p{4.7in}}

806: {\bf [UB]} & (Uniform Balance) $D(\sigma) = o(n)$. \\

807: \end{tabular}

808: 

809: \noindent \begin{tabular}{p{.4in}p{4.7in}} {\bf [SP]} &

810: (Separability) For any intervals $I,J,K,K^\prime \subset

811: \mathbb{Z}_n$,

812: $$

813: \left | \sum_{x \in K \cap \sigma^{-1}(K^\prime)} I(x)

814: J(\sigma(x)) - \frac{1}{n} \sum_{x \in K,y \in K^\prime} I(x)J(y)

815: \right | = o(n)

816: $$ \\

817: \end{tabular}

818: 

819: \noindent \begin{tabular}{p{.4in}p{4.7in}} {\bf [mS]} &

820: (m-Subsequences) For any permutation $\tau \in S_m$ and intervals

821: $I,J \subset \mathbb{Z}_n$,

822: $$

823: {\bf X}^\tau(\sigma |_{I \cap \sigma^{-1}(J)}) = \frac{1}{m!}

824: \binom{|\sigma(I) \cap J|}{m} + o(n^m).

825: $$ \\

826: \end{tabular}

827: 

828: \noindent \begin{tabular}{p{.4in}p{4.7in}} {\bf [2S]} &

829: (2-Subsequences) For any intervals $I,J \subset \mathbb{Z}_n$,

830: $$

831: {\bf X}^{(01)}(\sigma |_{I \cap \sigma^{-1}(J)}) - {\bf

832: X}^{(10)}(\sigma |_{I \cap \sigma^{-1}(J)}) = o(n^2).

833: $$ \\

834: \end{tabular}

835: 

836: \end{theorem}

837: 

838: It follows immediately that these conditions are also equivalent

839: to each interpretation of the statement ``For all intervals $J

840: \subset \mathbb{Z}_n$, $\sigma(J)$ is $\epsilon$-balanced'' given

841: by the equivalences of Theorem \ref{balance}.  Thus, we have a

842: total of ten equivalent quasirandom properties: seven arising as

843: ``uniformly convergent'' versions of the properties in Theorem

844: \ref{balance} and three new ones, which are included with uniform

845: balance in Figure 2.

846: 

847: \ifx\ispreprint\undefined

848: \else

849: \begin{center}

850: \begin{figure}[ht]

851: \hspace{1.25in} \input{figb.pstex_t} \caption{Diagram of

852: implications for Theorem \ref{qrp}.}

853: \end{figure}

854: \end{center}

855: \fi

856: 

857: Again, we prove the theorem piece by piece, keeping track of

858: $\epsilon$ as we go.  The next result states that, if uniform

859: balance is obeyed, then the variable $x$ and its image under

860: $\sigma$ are nearly independent.

861: 

862: \begin{prop} {\bf [UB]} $\Leftrightarrow$ {\bf [SP]}.

863: \end{prop}

864: \begin{proof}

865: {\bf [UB]} holds iff, for all intervals $I,J,K,K^\prime \subset

866: \mathbb{Z}_n$,

867: $$

868: \left | |\sigma(I \cap K) \cap (J \cap K^\prime)| - \frac{1}{n} |I

869: \cap K| |J \cap K^\prime| \right | < \epsilon n.

870: $$

871: But this quantity is equal to

872: $$

873: \left | \sum_{x \in K \cap \sigma^{-1}(K^\prime)} I(x)

874: J(\sigma(x)) - \frac{1}{n} \sum_{x \in K,y \in K^\prime} I(x) J(y)

875: \right |

876: $$

877: so that {\bf [UB]} is equivalent to {\bf [SP]}.

878: \end{proof}

879: 

880: Now, we show that the separability achieved in the last

881: proposition is sufficient to imply that subsequences happen at the

882: ``right'' rate (i.e., what one would expect of truly random

883: permutations) on certain sets of indices.  A computational lemma

884: will greatly simplify the proof.

885: 

886: \begin{lemma} \label{prodest} If, for each $j$ with $1 \leq j \leq k$, $|a_j|<n^{c_j}$, $\epsilon_j > n^{-1}$, and

887: $$

888: |x_j - a_j| < \epsilon_j n^{c_j-1}

889: $$

890: then

891: $$

892: \left | \prod_{j=1}^k x_j - \prod_{j=1}^k a_j \right | < 3^{k-1}

893: n^{\sum_{j=1}^k c_j - k} \prod_{j=1}^k \epsilon_j.

894: $$

895: \end{lemma}

896: \begin{proof}

897: We show the result for $k=2$, and the general case follows by a

898: simple induction.  Thus,

899: \begin{align*}

900: \left | x_1 x_2 - a_1 a_2 \right | & \leq \left | x_1 - a_1 \right | \cdot \left | x_2 - a_2 \right | + a_1  \left | x_2 - a_2 \right | + a_2 \left | x_1 - a_1 \right | \\

901: &< \epsilon_1 \epsilon_2 n^{c_1 + c_2 - 2} + \epsilon_1 n^{c_1 + c_2 - 1} + \epsilon_2 n^{c_1 + c_2 - 1} \\

902: &< 3 \epsilon_1 \epsilon_2 n^{c_1 + c_2 - 2}.

903: \end{align*}

904: \end{proof}

905: 

906: \begin{prop} {\bf [SP]} $\Rightarrow$ {\bf [mS]}.

907: \end{prop}

908: \begin{proof}

909: Let $I,J$ be intervals, and let $K = {I \cap \sigma^{-1}(J)}$.

910: Note that we may write the number of ``occurrences'' of $\tau \in

911: S_m$ in $\sigma |_K$ as

912: $$

913: {\bf X}^\tau(\sigma |_K) = \sum_{x_1,\cdots,x_m \in K}

914: \prod_{i=0}^m \left ( \chi(x_i < x_{i+1})

915: \chi(\sigma(x_{\tau^{-1}(j)}) < \sigma(x_{\tau^{-1}(j+1)})) \right

916: )

917: $$

918: In the interest of notational compactness, we will denote

919: $\chi(x<y)$ by $\cl{x}{y}$, and define $\cl{x}{y} = 1$ if either

920: $x$ or $y$ is undefined.  Furthermore, for any subset $A \subset

921: [m]$, we will denote the following expression

922: $$

923: \sum_{\{x_i\}_{i \not \in A} \subset K} \, \sum_{\{x_k\}_{k \in A}

924: \subset I} \sum_{\{x^{\prime}_k\}_{k \in A} \subset J} \left (

925: \prod_{j=0}^m \cl{x_j}{x_{j+1}}

926: \cl{x^A_{\tau^{-1}(j)}}{x^A_{\tau^{-1}(j+1)}} \right )

927: $$

928: by $\Sigma(A)$, where $x^A_j$ means $\sigma(x_j)$ for $j \not \in

929: A$, and $x^{\prime}_k$ for $j \in A$.  Thus, ${\bf X}^\tau(\sigma

930: |_K) = \Sigma(\emptyset)$.  The proof will now proceed by

931: induction on the subsets of $[m]$, ordered by inclusion.

932: 

933: Suppose $A \subset B \subset [m]$, with $B \setminus A = \{s\}$,

934: and assume that

935: \begin{equation} \label{indstep}

936: \left | {\bf X}^\tau(\sigma |_K) - n^{-|A|} \Sigma(A) \right | <

937: |A| \epsilon n^m

938: \end{equation}

939: By {\bf [SP]}, we know that, for any $a,b,c,d \in \mathbb{Z}_n$,

940: the quantity

941: $$

942: \left | \sum_{x_s \in K} \cl{a}{x_s} \cl{x_s}{b}

943: \cl{c}{\sigma(x_s)} \cl{\sigma(x_s)}{d} - \frac{1}{n} \sum_{x_s

944: \in I, x_s^\prime \in J} \cl{a}{x_s} \cl{x_s}{b}

945: \cl{c}{x^\prime_s} \cl{x^\prime_s}{d} \right |

946: $$

947: is bounded above by $\epsilon n$.  Then, substituting $a=x_{s-1}$,

948: $b=x_{s+1}$, $c=x^A_{\tau^{-1}(\tau(s)-1)}$, and

949: $d=x^A_{\tau^{-1}(\tau(s)+1)}$ to account for all the terms

950: containing $x_s$ in the product portion of the expression

951: $\Sigma(A)$, we have (after a very messy but otherwise

952: straightforward calculation),

953: $$

954: \left | \Sigma(A) - n^{-1} \Sigma(B) \right | < \epsilon n

955: |K|^{m-|B|} |I|^{|A|} |J|^{|A|} \leq \epsilon n^{m + |A|}

956: $$

957: Applying this to the inductive hypothesis with the aid of the

958: triangle inequality yields

959: \begin{align*}

960: \left | {\bf X}^\tau(\sigma |_K) - n^{-|B|} \Sigma(B) \right | & \leq \left | {\bf X}^\tau(\sigma |_K) - n^{-|A|} \Sigma(A) \right | + n^{-|A|} \left | \Sigma(A) - n^{-1} \Sigma(B) \right | \\

961: & < |A| \epsilon n^m + n^{-|A|} \epsilon n^{m+|A|} = |B| \epsilon

962: n^m.

963: \end{align*}

964: Therefore, (\ref{indstep}) is true for all $A \subset [m]$.  In

965: particular, it is true for $A = [m]$, so that

966: $$

967: \left | {\bf X}^\tau(\sigma |_K) - n^{-m} \Sigma([m]) \right | < m

968: \epsilon n^m

969: $$

970: Since we have

971: \begin{align}

972: \nonumber \left | \sum_{\{x_j\} \subset I} \prod_{j=0}^m \cl{x_j}{x_{j+1}} - \frac{|I|^m}{m!} \right | & = \left | \binom{|I|}{m} - \frac{|I|^m}{m!} \right | \leq \frac{(|I|+m)^m - |I|^m}{m!} \\

973: \nonumber & = \frac{1}{m!} \sum_{k=1}^{m} \binom{m}{k} |I|^{m-k} m^k \leq \frac{|I|^{m-1}}{m!} \sum_{k=0}^{m} \binom{m}{k} m^k \\

974: & = \frac{(1+m)^m}{m!} |I|^{m-1} \label{approxbin}

975: \end{align}

976: and also

977: $$

978: \left | \sum_{\{x_j\} \subset J} \prod_{j=0}^m \cl{x_j}{x_{j+1}} -

979: \frac{|J|^m}{m!} \right | \leq \frac{(1+m)^m}{m!} |J|^{m-1}

980: $$

981: we may conclude that

982: $$

983: \left | \Sigma([m]) - \frac{|I|^m|J|^m}{m!^2} \right | < 3

984: n^{2m-1} \frac{(1+m)^{2m}}{m!^2}

985: $$

986: by Lemma \ref{prodest}.  Thus,

987: \begin{align*}

988: \left | {\bf X}^\tau(\sigma |_K) - \frac{1}{m!^2} \left ( \frac{|I||J|}{n} \right )^m \right | & \leq \left | {\bf X}^\tau(\sigma |_K) - \frac{1}{n^m} \Sigma([m]) \right | \\

989: & + \frac{1}{n^m} \left | \Sigma([m]) - \frac{|I|^m|J|^m}{m!^2} \right | \\

990: & < m \epsilon n^m + 3 n^{m-1} \frac{(1+m)^{2m}}{m!^2} \\

991: & < 4 \epsilon n^m \frac{(1+m)^{2m}}{m!^2}

992: \end{align*}

993: But, by {\bf [UB]} (which is equivalent to {\bf [SP]}), and Lemma

994: \ref{prodest}

995: $$

996: \left | \left ( \frac{|I||J|}{n} \right )^m - |\sigma(I) \cap J|^m \right | < 3^{m-1} \epsilon^m n^m \\

997: $$

998: Since $m \geq 2$, this gives

999: \begin{align*}

1000: \left | {\bf X}^\tau(\sigma |_K) - \frac{1}{m!^2} |\sigma(I) \cap J|^m \right | &< n^m \frac{1}{m!^2} \left (3^{m-1} \epsilon^m + 4 \epsilon (1+m)^{2m} \right ) \\

1001: & < n^m \frac{1}{m!^2} \left ((1+m)^{2m} \epsilon + 4 \epsilon (1+m)^{2m} \right) \\

1002: & = \frac{5 \epsilon (1+m)^{2m}}{m!^2} n^m

1003: \end{align*}

1004: Finally, the fact that $\epsilon \geq n^{-1}$ implies, as in

1005: (\ref{approxbin}),

1006: $$

1007: \left | \binom{|\sigma(I) \cap J|}{m} - \frac{|\sigma(I) \cap

1008: J|^m}{m!} \right | \leq \epsilon \frac{(1+m)^m}{m!} n^m

1009: $$

1010: so we may conclude

1011: $$

1012: \left | {\bf X}^\tau(\sigma |_K) - \frac{1}{m!} \binom{|\sigma(I)

1013: \cap J|}{m} \right | < \frac{6 \epsilon (1+m)^{2m}}{m!^2} n^m < 4

1014: e^{2m} \epsilon n^m

1015: $$

1016: where we have used the Stirling approximation $m! > \sqrt{2 \pi m}

1017: \left (\frac{m}{e} \right)^m$.

1018: \end{proof}

1019: 

1020: Now, we use the results of the previous section to show that {\bf

1021: [mS]} implies {\bf [2S]}.

1022: 

1023: \begin{prop} {\bf [mS]} $\Rightarrow$ {\bf [2S]}.

1024: \end{prop}

1025: \begin{proof}

1026: Let $K=\sigma(I) \cap J$ for some intervals $I,J \in

1027: \mathbb{Z}_n$.  We may assume that $n \geq 2m$, so that $n-k>n/2$

1028: for all $k<m$.  Therefore, by Corollary \ref{down}, we may write

1029: $$

1030: | {\bf \tilde{v}}_{m-1}(\sigma|_K) |^2 \leq \frac{m^3}{(n-m+1)^2}

1031: {| {\bf \tilde{v}}_{m}(\sigma|_K) | } < \frac{4 m^3}{n^2} m!

1032: \max_{\tau \in S_m} \left | X^\tau(\sigma|_K) - \binom{|K|}{m}

1033: \right |^2

1034: $$

1035: Iterating this process $m-2$ times, we find

1036: \begin{align*}

1037: | {\bf \tilde{v}}_{m-1}(\sigma) |^2 &< \frac{4^{m-5} m!^{m+1}}{n^{2m-4}} \max_{\tau \in S_m} \left | X^\tau(\sigma|_K) - \binom{|K|}{m} \right | \\

1038: & < \frac{4^{m-2} m!^{m+1}}{n^{2m-4}} \epsilon^2 n^{2m} = 4^{m-5}

1039: m!^{m+1} \epsilon^2 n^4.

1040: \end{align*}

1041: Let the quantity $d$ be defined by

1042: $$

1043: d = \left | X^{(01)}(\sigma|_K) - \binom{|K|}{2} \right | = \left

1044: | X^{(10)}(\sigma|_K) - \binom{|K|}{2} \right |.

1045: $$

1046: Then $| {\bf \tilde{v}}_{2}(\sigma) |^2 = 2d^2$, so $d < 2^{m -

1047: 11/2} m!^{(m+1)/2} \epsilon n^2$, and

1048: $$

1049: \left | X^{(01)}(\sigma|_K) - X^{(10)}(\sigma|_K) \right | < 2d <

1050: 2^m m!^{(m+1)/2} \epsilon n^2

1051: $$

1052: which implies {\bf [2S]}.

1053: \end{proof}

1054: 

1055: In what follows, we denote the complement of a set $S \in

1056: \mathbb{Z}_n$ by $\bar{S}$, and we denote by $S^*$ its projection

1057: onto $[0,n-1]$.  Also, call an interval $I \subset \mathbb{Z}_n$

1058: ``contiguous'' if $I^*$ is an interval, ``terminal'' if $\bar{I}$

1059: is contiguous, ``initial'' if it is terminal and contains $0$, and

1060: ``final'' if it is terminal and contains $n-1$.

1061: 

1062: \begin{prop} {\bf [2S]} $\Rightarrow$ {\bf [UB]}.

1063: \end{prop}

1064: \begin{proof}

1065: Suppose $\sigma$ satisfies {\bf [2S]} but not {\bf [UB]}. We claim

1066: that, for infinitely many $n$ and some $\epsilon > 0$, there are

1067: intervals $I, J \subset \mathbb{Z}_n$ with $I$ and $J$ initial,

1068: and $D_J(\sigma(I))$ at least $27\epsilon n/2$.  Since {\bf [UB]}

1069: is not true for $\sigma$, we may choose $\epsilon$ so that there

1070: are proper subintervals $I,J \subset \mathbb{Z}_n$ with

1071: $D_{J}(\sigma(I)) \geq 54 \epsilon n$.  Suppose $J$ is not

1072: contiguous.  Then $\bar{J}$ is contiguous, and, since

1073: $$

1074: \left ( \left | \sigma(I) \cap J  \, \right | - \frac{|I||J|}{n}

1075: \right ) + \left ( \left | \sigma(I) \cap \bar{J} \, \right | -

1076: \frac{|I||\bar{J}|}{n} \right ) = \left | \sigma(I) \right | -

1077: \frac{|I|n}{n} = 0

1078: $$

1079: we may replace $J$ with $\bar{J}$ and retain the property that

1080: $D_{J}(\sigma(I)) \geq 54 \epsilon n$.  Now, suppose $J$ is not

1081: terminal.  Let $J^\prime$ be a component of $(\bar{J})^*$.

1082: $J^\prime$ is terminal because $J$ is contiguous, and, since we

1083: have

1084: $$

1085: \left | \sigma(I) \cap (J \cup J^\prime)  \, \right | -

1086: \frac{|I||J \cup J^\prime |}{n} = \left ( \left | \sigma(I) \cap J

1087: \, \right | - \frac{|I||J|}{n} \right ) + \left ( \left |

1088: \sigma(I) \cap J^\prime \, \right | - \frac{|I||J^\prime|}{n}

1089: \right )

1090: $$

1091: either $D_{J \cup J^\prime}(\sigma(I)) \geq 27 \epsilon n$ or

1092: $D_{J^\prime}(\sigma(I)) \geq 27 \epsilon n$.  Thus, we may assume

1093: that $J$ is terminal (since $J^\prime$ and $J \cup J^\prime$ are),

1094: and $D_{J}(\sigma(I)) \geq 27 \epsilon n$.  If $J$ is final,

1095: taking its complement makes it initial without disturbing the

1096: discrepancy.  Apply the same process to $I$ to ensure that it is

1097: initial, with the penalty that now

1098: \begin{equation} \label{tst}

1099: D_{J}(\sigma(I)) \geq 27 \epsilon n/2.

1100: \end{equation}

1101: 

1102: For ease of notation, we will let

1103: \begin{eqnarray*}

1104: A & = I \cap \sigma^{-1}(J) & \hspace{.3in} a=|A| \\

1105: B & = I \cap \sigma^{-1}(\bar{J}) & \hspace{.3in} b=|B| \\

1106: C & = \bar{I} \cap \sigma^{-1}(J) & \hspace{.3in} c=|C| \\

1107: D & = \bar{I} \cap \sigma^{-1}(\bar{J}) & \hspace{.3in} d=|D|

1108: \end{eqnarray*}

1109: For subsets $S,T \subset \mathbb{Z}_n$, let $\partial_\sigma

1110: (S,T)$ denote the number of pairs $(x,y) \in S \times T$ such that

1111: $x < y$ and $\sigma(x) < \sigma(y)$.  Then

1112: \begin{align*}

1113: {\bf X}^{(01)}(\sigma) & = {\bf X}^{(01)}(\sigma|_I) + {\bf X}^{(01)}(\sigma|_{\bar{I}}) + \partial_\sigma(I, \bar{I}) \\

1114: & = {\bf X}^{(01)}(\sigma|_I) + {\bf X}^{(01)}(\sigma|_{\bar{I}}) + \partial_\sigma(A, C) \\

1115: & \, \, \, \, \, + \partial_\sigma(A, D) + \partial_\sigma(B, C) +

1116: \partial_\sigma(B, D)

1117: \end{align*}

1118: Now, $\partial_\sigma(B, C) = 0$ and $\partial_\sigma(A, D) = ad$,

1119: since every element of $J$ is less than every element of

1120: $\bar{J}$, and every element of $I$ is less than every element of

1121: $\bar{I}$.  Also,

1122: \begin{align*}

1123: \partial_\sigma(A, C) & = {\bf X}^{(01)}(\sigma|_{\sigma^{-1}(J)}) - {\bf X}^{(01)}(\sigma|_A) - {\bf X}^{(01)}(\sigma|_C) \\

1124: \partial_\sigma(B, D) & = {\bf X}^{(01)}(\sigma|_{\sigma^{-1}(\bar{J})}) - {\bf X}^{(01)}(\sigma|_B) - {\bf X}^{(01)}(\sigma|_D)

1125: \end{align*}

1126: Thus, we have

1127: \begin{align*}

1128: {\bf X}^{(01)}(\sigma) & = {\bf X}^{(01)}(\sigma|_I) + {\bf X}^{(01)}(\sigma|_{\bar{I}}) + {\bf X}^{(01)}(\sigma|_{\sigma^{-1}(J)}) - {\bf X}^{(01)}(\sigma|_A) \\

1129: & - {\bf X}^{(01)}(\sigma|_C) + ad + {\bf

1130: X}^{(01)}(\sigma|_{\sigma^{-1}(\bar{J})}) - {\bf

1131: X}^{(01)}(\sigma|_B) - {\bf X}^{(01)}(\sigma|_D)

1132: \end{align*}

1133: By {\bf [2S]}, for sufficiently large $n$, each term ${\bf

1134: X}^{(01)}(\sigma|_S)$ can be approximated by $\binom{|S|}{2}/2$ to

1135: within $\epsilon n^2/2$, and therefore by $|S|^2/4$ to within $3

1136: \epsilon n^2/4$, since

1137: $$

1138: \left | \binom{|S|}{2} - \frac{|S|^2}{2} \right | = \frac{|S|}{2}

1139: < \frac{\epsilon}{2} n^2.

1140: $$

1141: Therefore, rewriting and multiplying by $4$, we have that

1142: $$

1143: \left | n^2 - (a+b)^2 - (c+d)^2 - (a+c)^2 - (a+d)^2 + a^2 + b^2 +

1144: c^2 + d^2 - 4ad \right |

1145: $$

1146: is bounded above by $27 \epsilon n^2$.  Since $n = a+b+c+d$, we

1147: may simplify down to

1148: \begin{equation} \label{eq1}

1149: | bc - ad | < \frac{27 \epsilon}{2} n^2

1150: \end{equation}

1151: Let $\delta n = \left | I \cap \sigma^{-1}(J) \right | -

1152: |I||J|/n$.  Then, by (\ref{tst}),

1153: \begin{align*}

1154: | bc - ad | &= \left | \left ( \frac{|I||\bar{J}|}{n} - \delta n \right ) \left ( \frac{|\bar{I}||J|}{n} - \delta n \right ) - \left ( \frac{|I||J|}{n} + \delta n \right ) \left ( \frac{|\bar{I}||\bar{J}|}{n} + \delta n \right ) \right | \\

1155: &= |\delta| (|I||J| + |I||\bar{J}| + |\bar{I}||J| + |\bar{I}||\bar{J}|) \\

1156: &= \frac{D_J(\sigma(I))}{n} \cdot (|I|+|\bar{I}|)(|J|+|\bar{J}|)

1157: \geq \frac{27 \epsilon}{2} n^2

1158: \end{align*}

1159: contradicting (\ref{eq1}).

1160: \end{proof}

1161: 

1162: We present two simple applications of these results.  The following observation has some relevance to the investigations

1163: of [\ref{AF1}] and [\ref{B1}].

1164: 

1165: \begin{prop} If a permutation $\sigma \in S_n$ excludes $\tau \in

1166: S_m$ (in the sense that ${\bf X}^\tau(\sigma) = 0$), then

1167: $$

1168: D(\sigma) \geq \frac{n\binom{n}{m}}{4 \, e^{2m} \,m!\, n^m}.

1169: $$

1170: \end{prop}

1171: \begin{proof} Let $\epsilon = D(\sigma)/n$.  We show that, if

1172: $$

1173: \epsilon \leq \frac{\binom{n}{m}}{4 \, e^{2m} \, m! \, n^m}

1174: $$

1175: then there is at least one copy of every element $\tau$ of $S_m$

1176: in $\sigma$.  According to the implication from {\bf [UB]} to {\bf

1177: [mS]}, if $D(\sigma) < \epsilon n$, then

1178: $$

1179: \left | \textbf{X}^\tau(\sigma) - \frac{1}{m!} \binom{n}{m} \right

1180: | < 4 e^{2m} \epsilon n^m \leq \frac{1}{m!} \binom{n}{m},

1181: $$

1182: so that $\textbf{X}^\tau(\sigma) > 0$.

1183: \end{proof}

1184: 

1185: \begin{cor} There is a constant $c > 0$ so that, if $n \geq 2m$ and

1186: $\sigma \in S_n$ excludes $\tau \in S_m$, then

1187: $$

1188: \frac{D(\sigma)}{n} > \left(\frac{c}{m^2} \right)^m.

1189: $$

1190: \end{cor}

1191: 

1192: We can also use Theorem \ref{qrp} to calculate the discrepancy of a random permutation.

1193: 

1194: \begin{theorem} If a permutation $\sigma$ is chosen randomly and uniformly from $S_n$, then $D(\sigma) = O(\sqrt{n \log n})$ almost surely.

1195: \end{theorem}

1196: \begin{proof} We use $c_i$, $i \in \mathbb{N}$ to denote absolute constants throughout.  Let the random variable $\xi_n$ be the number of inversions in a randomly and uniformly chosen element of $S_n$.  Define $\eta_n$ to be the normalized random variable given by

1197: $$

1198: \eta_n = \frac{\xi_n - \mbox{\textbf{E}}(\xi_n)}{(\mbox{Var}\, \xi_n)^{1/2}} = \frac{\xi_n - {\frac{1}{2} \binom{n}{2}}}{(\mbox{Var}\, \xi_n)^{1/2}},

1199: $$

1200: and let $u_n$ denote the distribution function with an atom of mass $1/n$ at each of $0, \ldots , n-1$.  It is well known ([\ref{St1}],[\ref{St2}]) that the generating function $g_n(q)$ for the number of permutations with a given number of inversions is the $q$-factorial $[n]!$, and that its coefficients are symmetric and unimodal.  Therefore, $g_n = g_{n-1} \cdot (q^{n-1} + \ldots + 1)$, so that $\xi_n = \xi_{n-1} \ast u_n$ (the convolution product).  The unimodality of the coefficients of $g_{n-1}$ implies the concavity of the cumulative distribution function $F_n(x)$ of $\xi_n$ on the interval $[\lceil \binom{n}{2}/2 \rceil,\infty]$.  Thus, if $x \geq \binom{n-1}{2}/2+n-1$, then $F_n(x) = (F_{n-1} \ast u_n)(x) \leq F_{n-1}(x)$.  (The convolution is a finite sum, so the implicit change of order of summation is legitimate.)  Similarly, $F_n(x) = (F_{n-1} \ast u_n)(x) \geq F_{n-1}(x)$ if $x \leq \binom{n-1}{2}/2$.  We have, then,

1201: \begin{equation} \label{monotone}

1202: \mbox{\textbf{Pr}}( \left|\xi_n - \mbox{\textbf{E}}(\xi_n)\right| > \lambda) \geq \mbox{\textbf{Pr}}( \left|\xi_{m} - \mbox{\textbf{E}}(\xi_m)\right| > \lambda )

1203: \end{equation}

1204: whenever $\lambda \geq (n-1)/2$ and $m \leq n$.

1205: 

1206: 

1207: It is a theorem of Sachkov [\ref{Sa1}] that the cumulative distribution function of $\eta_n$ converges to $\Phi(0,1)$ (the c.d.f. of the standard normal distribution), and that $\sigma_n^2 = \mbox{Var}\, \xi_n = n^3/36 + O(n^2)$.  In particular, the moment generating function $M(t,n)=\mbox{\textbf{E}}(e^{t \eta_n})$ of $\eta_n$ is given by

1208: $$

1209: \log M(t,n) = \frac{t^2}{2} + \sum_{k=2}^\infty \frac{B_{2k} \, t^{2k}}{2k \, \sigma_n^{2k} \, (2k)!} \sum_{j=1}^n (j^{2k}-1)

1210: $$

1211: where $B_{2k}$ is the $2k^{\mbox{th}}$ Bernoulli number.  Let $f_n(t) = \log M(t,n) - t^2/2$.  Then, using the approximation $|B_{2k}| < 4 (2k)!/(2 \pi)^{2k}$, we have

1212: \begin{equation} \label{momest}

1213: |f_n(t)| \leq c_1 \sum_{k=2}^\infty \frac{t^{2k}}{(2 \pi)^{2k} \, 2k \, n^{3k}} \cdot n^{2k+1} \leq c_2 n \sum_{k=2}^\infty \left ( \frac{t}{2 \pi n^{1/2}} \right )^{2k} \leq \frac{c_3 t^4}{4 \pi^2 n - t^2},

1214: \end{equation}

1215: so long as $t < n^{1/2}$.

1216: 

1217: By Theorem \ref{qrp}, there exists an $\alpha>0$ so that

1218: $$

1219: \mbox{\textbf{Pr}}(D(\sigma) > \lambda \sqrt{n \log n}) \leq \mbox{\textbf{Pr}}(\max_{I,J} \left|{\bf X}^{(10)}(\sigma|_{I \cap \sigma^{-1}(J)}) - \frac{1}{2} \binom{|I \cap \sigma^{-1}(J)|}{2}\right| > c_4 \lambda n^{3/2} \sqrt{\log n}).

1220: $$

1221: By (\ref{monotone}), we may write

1222: \begin{align*}

1223: \mbox{\textbf{Pr}}(D(\sigma) > \lambda \sqrt{n \log n}) & \leq \sum_{I,J} \mbox{\textbf{Pr}}(\left|\xi_{|\sigma(I) \cap J|} - \mbox{\textbf{E}} \xi_{|\sigma(I) \cap J|}\right| > c_4 \lambda n^{3/2} \sqrt{\log n}) \\

1224: & \leq n^4 \mbox{\textbf{Pr}}(\left|\xi_n - \mbox{\textbf{E}} \xi_n \right| > c_4 \lambda n^{3/2} \sqrt{\log n})

1225: \end{align*}

1226: so long as $n$ is sufficiently large.  Furthermore, by Markov's inequality and the estimate on $\sigma_n$, for fixed $\lambda > 0$,

1227: \begin{align*}

1228: \mbox{\textbf{Pr}}(\left|\xi_n - \mbox{\textbf{E}} \xi_n \right| > c_4 \lambda n^{3/2} \sqrt{\log n}) & \leq \mbox{\textbf{Pr}}(\left|\eta_n\right| > c_5 \lambda \sqrt{\log n}) \\

1229: &= 2 \mbox{\textbf{Pr}}(e^{\eta_n} > e^{c_5 \lambda \sqrt{\log n}} ) \\

1230: &\leq 2 \mbox{\textbf{E}}(e^{t \eta_n}) e^{-t c_5 \lambda \sqrt{\log n}}.

1231: \end{align*}

1232: Setting $t = c_5 \lambda \sqrt{\log n}$ and applying the bound (\ref{momest}), we have

1233: $$

1234: \mbox{\textbf{Pr}}(\left|\xi_n - \mbox{\textbf{E}} \xi_n \right| > c_4 \lambda n^{3/2} \sqrt{\log n}) \leq e^{- c_6 \lambda^2 \log n + c_7 \log^2 n/n} \leq c_8 e^{- c_6 \lambda^2 \log n}.

1235: $$

1236: Therefore,

1237: $$

1238: \mbox{\textbf{Pr}}(D(\sigma) > \lambda \sqrt{n \log n}) \leq c_8 n^{4 - c_6 \lambda^2},

1239: $$

1240: which tends to zero if we choose $\lambda > 2 c_6^{-1/2}$.

1241: \end{proof}

1242: 

1243: 

1244: \bigskip

1245: \section{Constructions}

1246: \setcounter{theorem}{0}

1247: 

1248: In this section, we present a construction for a large class of permutations which are highly quasirandom.  We will assume throughout that $\sigma \in S_n$ and $\tau \in S_m$, unless indicated otherwise.

1249: 

1250: \begin{definition} For permutations $\sigma \in S_n$ and $\tau \in S_m$, considered as actions on $\mathbb{Z}_n$ and $\mathbb{Z}_m$, respectively, define $\sigma \otimes \tau \in S_{nm}$ by $(\sigma \otimes \tau)(x) = \tau(\lfloor \frac{x}{n} \rfloor) + m \sigma(x \! \! \mod n)$.  We will also denote the $k^{th}$ product of $\sigma$ with itself as $\sigma^{(k)}$.

1251: \end{definition}

1252: 

1253: A special case of this product appears in [\ref{D2}], where the authors define a sequence of permutations lacking ``monotone $3$-term arithmetic progressions'' by taking iterated products of the elements of $S_2$.

1254: 

1255: Note that $\sigma \otimes \tau$ has the property that $(\sigma

1256: \otimes \tau)([0,n-1])$ is the set of all elements of

1257: $\mathbb{Z}_{nm}$ congruent to $0$ mod $m$ (i.e., $m \cdot

1258: [0,n-1]$), a set which necessarily lacks the ``weak translation''

1259: property of quasirandom sets.  Thus, a sequence $\{

1260: \sigma_1,\sigma_1 \otimes \sigma_2,\sigma_1 \otimes \sigma_2

1261: \otimes \sigma_3,\ldots \}$ sends intervals to sets which are not

1262: quasirandom in the sense of [\ref{CG2}].  Nonetheless, we will

1263: prove shortly that it does satisfy ${\bf UB}$.  First, we

1264: offer a justification for the lack of parentheses in the

1265: expression for this sequence.

1266: 

1267: \begin{prop} $\otimes$ is associative.

1268: \end{prop}

1269: \begin{proof}

1270: Suppose $\sigma_i \in S_{n_i}$ for $i=1,2,3$.  Then, applying the

1271: definition of $\otimes$ twice,

1272: \begin{align*}

1273: \left [(\sigma_1 \otimes \sigma_2) \otimes \sigma_3 \right] (x) & = \sigma_3 \left( \left \lfloor \frac{x}{n_1n_2} \right \rfloor \right ) + n_3 \sigma_2 \left ( \left \lfloor \frac{x \mbox{ mod } n_1n_2}{n_1} \right \rfloor \right ) \\

1274: & \, \, \, \, \, + n_2n_3 \sigma_1(x \mbox{ mod } n_1)

1275: \end{align*}

1276: and

1277: \begin{align*}

1278: \left [\sigma_1 \otimes (\sigma_2 \otimes \sigma_3) \right ](x) & = \sigma_3 \left( \left \lfloor \frac{\left \lfloor x/n_1 \right \rfloor}{n_2} \right \rfloor \right ) + n_3 \sigma_2 \left ( \left \lfloor \frac{x}{n_1} \right \rfloor \mbox{ mod } n_2 \right ) \\

1279: & \, \, \, \, \, + n_2n_3 \sigma_1(x \mbox{ mod } n_1)

1280: \end{align*}

1281: Note that every element $x$ of $\mathbb{Z}_{n_1n_2n_3}$ can be

1282: represented uniquely as $x = an_1n_2 + bn_1 + c$, with $0 \leq a <

1283: n_3$, $0 \leq b < n_2$, and $0 \leq c < n_1$.  Using this

1284: notation, we find that

1285: \begin{align*}

1286: \left [(\sigma_1 \otimes \sigma_2) \otimes \sigma_3 \right] (x) & = \sigma_3(a) + n_3 \sigma_2 \left ( \left \lfloor \frac{bn_1 + c}{n_1} \right \rfloor \right ) + n_2n_3 \sigma_1(c) \\

1287: & = \sigma_3(a) + n_3 \sigma_2(b) + n_2n_3 \sigma_1(c)

1288: \end{align*}

1289: and

1290: \begin{align*}

1291: \left [\sigma_1 \otimes (\sigma_2 \otimes \sigma_3) \right ](x) = \,\, & \sigma_3 \left( \left \lfloor \frac{an_2 + b}{n_2} \right \rfloor \right ) + n_3 \sigma_2 \left ( (a n_2 + b) \mbox{ mod } n_2 \right ) \\

1292: & \,\, + n_2n_3 \sigma_1(c) \\

1293: = & \,\, \sigma_3(a) + n_3 \sigma_2(b) + n_2n_3 \sigma_1(c).

1294: \end{align*}

1295: \end{proof}

1296: 

1297: Define $d(\sigma)$ by

1298: $$

1299: d(\sigma) = \max_{I,J} D_J(\sigma(I))

1300: $$

1301: where $I$ is allowed to vary over all possible intervals, but $J$

1302: is restricted to initial intervals.  We denote the analogue for

1303: final intervals by $d^\prime$.  Then we have the following result:

1304: 

1305: \begin{prop} $d(\sigma \otimes \tau) \leq m - 1 + d(\sigma)$.

1306: \end{prop}

1307: \begin{proof}

1308: Let the interval $I_k=[kn,(k+1)n-1] \subset \mathbb{Z}_{nm}$.

1309: Then, any initial interval $S$ of $\mathbb{Z}_{nm}$ can, for some

1310: $l<m$, be written

1311: $$

1312: S = \bigcup_{k=0}^l I_k \cup S_0

1313: $$

1314: where $S_0$ is an initial segment of $I_{l+1}$.  For any interval

1315: $J \subset \mathbb{Z}_{nm}$, then, we may write

1316: $$

1317: D_J((\sigma \otimes \tau)(S)) \leq \sum_{k=0}^l D_J((\sigma

1318: \otimes \tau)(I_k)) + D_J((\sigma \otimes \tau)(S_0))

1319: $$

1320: by Lemma \ref{subadd}.  First, we estimate $D_J(\sigma(I_k))$.

1321: \begin{align*}

1322: D_J((\sigma \otimes \tau)(I_k)) &= \left | \, | \, (\sigma \otimes \tau)(I_k) \cap J | - \frac{|(\sigma \otimes \tau)(I_k)||J|}{nm} \right | \\

1323: &= \left | \, | \, (m[0,n-1]+k) \cap J | - \frac{n|J|}{nm} \right | \\

1324: &\leq \left | \, \frac{|J| + m - 1}{m} - \frac{|J|}{m} \right | =

1325: \frac{m-1}{m}

1326: \end{align*}

1327: Let $J_0 \subset \mathbb{Z}_n$ denote the set $\left \{ \lfloor

1328: \frac{x}{m} \rfloor | \, x \in J \right \}$, and let $S_1 \subset

1329: \mathbb{Z}_n$ be the set $S_0$ reduced mod $n$.  Then,

1330: \begin{align*}

1331: D_J((\sigma \otimes \tau)(S_0)) &= \left | \, | \, (\sigma \otimes \tau)(S_0) \cap J | - \frac{|(\sigma \otimes \tau)(S_0)||J|}{nm} \right | \\

1332: &= \left | \, | \, \sigma(S_1) \cap J_0 | - \frac{|S_1||J|}{nm} \right | \\

1333: &= \left | \, | \, \sigma(S_1) \cap J_0 | - \frac{|S_1|m|J_0|}{nm} + \frac{|S_1|}{nm} (m|J_0|-|J|) \right | \\

1334: & \leq D_{J_0}(\sigma(S_1)) + \left ( \, \frac{n}{nm} (m-1) \right

1335: ) \leq d(\sigma) + \frac{m-1}{m}

1336: \end{align*}

1337: Thus,

1338: \begin{align*}

1339: d(\sigma \otimes \tau) & \leq (m-1) \frac{m-1}{m} + \frac{m-1}{m} + d(\sigma) \\

1340: & = m - 1 + d(\sigma).

1341: \end{align*}

1342: \end{proof}

1343: 

1344: An identical result holds for $d^\prime$, by symmetry.  We use

1345: this in the next proposition, which allows us to bound

1346: discrepancies recursively.

1347: 

1348: \begin{prop} \label{inh} $D(\sigma \otimes \tau) \leq m-1 + d(\sigma) + d^\prime(\sigma)$.

1349: \end{prop}

1350: \begin{proof}

1351: Note that every interval $I$ of $\mathbb{Z}_{nm}$ is of the form

1352: $$

1353: S = \bigcup_{k \in [l,L]} I_k \cup S_0 \cup S^\prime_0

1354: $$

1355: where $[l,L]$ is an interval of $\mathbb{Z}_m$ of length no more

1356: than $m-2$, $S_0$ is an initial segment of $I_{L+1}$, and

1357: $S^\prime_0$ is a final segment of $I_{l-1}$.  Applying Lemma

1358: \ref{subadd},

1359: $$

1360: D_J((\sigma \otimes \tau)(S)) \leq \sum_{k=l}^L D_J((\sigma

1361: \otimes \tau)(I_k)) + D_J((\sigma \otimes \tau)(S_0)) +

1362: D_J((\sigma \otimes \tau)(S^\prime_0))

1363: $$

1364: By the arguments presented in the proof of the previous

1365: proposition,

1366: \begin{eqnarray*}

1367: D_J((\sigma \otimes \tau)(I_k)) & \leq & \frac{m-1}{m} \\

1368: D_J((\sigma \otimes \tau)(S_0)) & \leq & d(\sigma) + \frac{m-1}{m} \\

1369: D_J((\sigma \otimes \tau)(S^\prime_0)) & \leq & d^\prime(\sigma) +

1370: \frac{m-1}{m}

1371: \end{eqnarray*}

1372: Thus,

1373: $$

1374: D(\sigma \otimes \tau) \leq m - 1 + d(\sigma) + d^\prime(\sigma)

1375: $$

1376: \end{proof}

1377: 

1378: If we apply these results to a product of permutations,

1379: \begin{cor} \label{prodperm} If, for $i \leq k$, $\sigma_i \in S_{n_i}$, where $n_i > 1$, then

1380: $$

1381: D(\bigotimes_{i=1}^{k} \sigma_i ) \leq n_k + 2 \sum_{i=1}^{k-1}

1382: n_i - 2k + 1.

1383: $$

1384: \end{cor}

1385: \begin{proof}

1386: By Proposition \ref{inh}, $d(\bigotimes_{i=1}^{m} \sigma_i ) \leq

1387: n_m - 1 + d(\bigotimes_{i=1}^{m-1} \sigma_i )$.  Inductively,

1388: then, we find

1389: $$

1390: d(\bigotimes_{i=1}^{m} \sigma_i ) \leq \sum_{i=2}^{m} (n_i - 1) +

1391: d(\sigma_1).

1392: $$

1393: Since $d(\sigma_1) \leq n_1 - 1$, we may write

1394: $d(\bigotimes_{i=1}^{m} \sigma_i ) \leq \sum_{i=1}^{m} n_i - m$

1395: and, similarly, $d^\prime(\bigotimes_{i=1}^{m} \sigma_i ) \leq

1396: \sum_{i=1}^{m} n_i - m$.  Thus,

1397: \begin{align*}

1398: D(\bigotimes_{i=1}^{k} \sigma_i ) & \leq n_k - 1 + d(\bigotimes_{i=1}^{k-1} \sigma_i ) + d^\prime(\bigotimes_{i=1}^{k-1} \sigma_i ) \\

1399: & \leq n_k + 2 \sum_{i=1}^{k-1} n_i - 2k + 1.

1400: \end{align*}

1401: \end{proof}

1402: 

1403: Corollary \ref{prodperm} provides us with a large family of very

1404: strongly quasirandom permutations.  To see this, let

1405: $\{\sigma_i\}_{i=1}^\infty$ be a sequence of permutations with

1406: $\sigma_i \in S_{n_i}$.  Then, letting $\lambda_k = \sum_{i=1}^{k}

1407: n_i / \prod_{i=1}^{k} n_i$ and applying the corollary,

1408: $$

1409: \frac{D(\bigotimes_{i=1}^{k} \sigma_i)}{\prod_{i=1}^{k} n_i} < 2

1410: \frac{\sum_{i=1}^{k} n_i}{\prod_{i=1}^{k} n_i} = 2 \lambda_k

1411: $$

1412: But,

1413: \begin{align*}

1414: \lambda_k &= \frac{1}{n_k} \cdot \frac{\sum_{i=1}^{k-1} n_i}{\prod_{i=1}^{k-1} n_i} + \frac{n_k}{\prod_{i=1}^{k} n_i} \\

1415: &= \left(\frac{1}{n_k} + \frac{1}{\sum_{i=1}^{k-1} n_i} \right )

1416: \lambda_{k-1}

1417: \end{align*}

1418: Thus, the ratio of the discrepancy to the size of the product

1419: permutations tends to zero quickly.  In particular, $\sigma^{(k)}$

1420: is very strongly quasirandom, since $\sigma^{(k)} \in S_{n^k}$.

1421: That is, if $\sigma^{(k)} \in S_N$, then $D(\sigma^{(k)}) < 2kn =

1422: O(\log N)$.  Immediately one wonders whether permutations exist

1423: with discrepancies which grow slower than $\log N$.  A theorem of Schmidt [\ref{S1}] answers this question in the negative, implying that the $D(\sigma^{(k)})$ are, in a sense, ``maximally'' quasirandom.

1424: 

1425: \begin{theorem} [Schmidt] \label{schmidt} Let $\{x_i\}_{i=0}^{N-1} \subset [0,1)$, and define

1426: $$

1427: D(m) = \sup_{\alpha \in [0,1)} \left | |\{x_i\}_{i=0}^{m-1} \cap

1428: [0,\alpha) | - m \alpha \right |.

1429: $$

1430: Then there exists an integer $n \leq N$ so that $D(n) > \log N /

1431: 100$.

1432: \end{theorem}

1433: 

1434: We may immediately conclude that discrepancies grow at least as

1435: fast as $\log N$.

1436: 

1437: \begin{cor} \label{disclowerbound} For any $\sigma \in S_N$, $D(\sigma) > \log N / 100 - 1$.

1438: \end{cor}

1439: \begin{proof}

1440: Take $x_i = \sigma(i)/N$ in Theorem \ref{schmidt}.  Then there

1441: exists an $\alpha \in [0,1)$ and an $n \leq N$ so that

1442: $$

1443: \left | \left |\frac{\sigma([0,n-1])}{N} \cap [0,\alpha) \right |

1444: - n \alpha \right | > \frac{\log N}{100}.

1445: $$

1446: Defining $k=\lfloor \alpha N \rfloor$, we have

1447: $$

1448: \left | |\sigma([0,n-1]) \cap [0,k] | - \frac{n(k+1)}{N} \right |

1449: + \left | \frac{n(k+1)}{N} - n \alpha \right | > \frac{\log

1450: N}{100}.

1451: $$

1452: Therefore, if we let $I$ and $J$ vary over all intervals in

1453: $\mathbb{Z}_N$,

1454: $$

1455: \max_{I,J} \left | |\sigma(I) \cap J | - \frac{|I||J|}{n} \right |

1456: > \frac{\log N}{100} - 1

1457: $$

1458: so that $D(\sigma) > \log N / 100 - 1$.

1459: \end{proof}

1460: 

1461: One might expect that the algebraic properties of quasirandom

1462: permutations, such as the number of cycles, should be

1463: approximately that of random permutations (in this case, $\log

1464: n$).  However, we have the following counterexample.  Let $i_n$ be

1465: the identity permutation on $\mathbb{Z}_n$.  Then $i^{(k)}_n$ is

1466: always an involution in $S_{n^k}$ -- and the sequence $\{i^{(k)}_n/n^k\}_{i=0}^{n-1} \subset [0,1)$ is an initial segment of the van der Corput sequence.  In fact, under this interpretation, Corollary \ref{prodperm} can be considered a generalization of the classical theorem that the discrepancy of the van der Corput sequence is $O(\log N)$.  (See, for example, [\ref{F1}] for a modern version of this result.)

1467: 

1468: \begin{prop} $i^{(k)}_n(x)$ is the element of $\mathbb{Z}_{n^k}$ whose base $n$ expansion is the reverse of the base $n$ expansion of $x$.

1469: \end{prop}

1470: \begin{proof}

1471: The proof is by induction on $k$.  The case $k=1$ is obvious.

1472: Suppose it were true of $i^{(k)}_n$.  Let $x \in

1473: \mathbb{Z}_{n^{k+1}}$ have the base $n$ expansion $x_{k} x_{k-1}

1474: \cdots x_1 x_0$.  Then

1475: \begin{align*}

1476: i^{(k+1)}_n(x) & = (i^{(k)}_n \otimes i_n)(x) \\

1477: & = i_n(\lfloor \frac{x}{n^k} \rfloor) + n i^{(k)}_n(x \! \! \! \! \mod n^k) \\

1478: & = (\lfloor \frac{x}{n^k} \rfloor) + n i^{(k)}_n(x_{k-1} x_{k-2} \cdots x_1 x_0) \\

1479: & = x_k + n (x_0 x_1 \cdots x_{k-2} x_{k-1}) \\

1480: & = x_0 x_1 \cdots x_{k-2} x_{k-1} x_k.

1481: \end{align*}

1482: \end{proof}

1483: 

1484: \section{{\sc Conclusion}}

1485: \setcounter{theorem}{0}

1486: 

1487: The original motivation for this paper was a (still unanswered)

1488: question of R. L. Graham [\ref{G2}].  For a sequence of

1489: permutations $\sigma_j \in S_{n_j}$, let {\bf P(k)} be the

1490: property of \textit{asymptotic $k$-symmetry}: for each $\tau \in

1491: S_k$,

1492: $$

1493: \left | X^{\tau}(\sigma_j) - \frac{\binom{n_j}{k}}{k!} \right | =

1494: o(n_j^k).

1495: $$

1496: Note that this property is weaker than property \textbf{[kS]} of

1497: Theorem \ref{qrp}, which we will call \textit{strong} asymptotic

1498: $k$-symmetry.  Theorem \ref{qrp} says that strong asymptotic

1499: $k$-symmetry implies strong asymptotic $(k+1)$-symmetry for any $k

1500: \geq 2$.  Graham asks whether there exists an analogous $N$ so

1501: that, for all $k > N$, $\mbox{\bf P(k)} \Rightarrow \mbox{\bf

1502: P(k+1)}$?  At first it might seem like one is asking for too much.

1503: However, precisely this type of phenomenon occurs for graphs

1504: ([\ref{CGW1}]).  It turns out that, if we let {\bf G(k)} be the

1505: property that all graphs on $k$ vertices occur as subgraphs at

1506: approximately the same rate, then

1507: $$

1508: \mbox{\bf G(1)} \Leftarrow \mbox{\bf G(2)} \Leftarrow \mbox{\bf

1509: G(3)} \Leftarrow \mbox{\bf G(4)} \Leftrightarrow \mbox{\bf G(5)}

1510: \Leftrightarrow \mbox{\bf G(6)} \Leftrightarrow \cdots

1511: $$

1512: In particular, {\bf G(4)} implies quasirandomness, which in turn

1513: implies {\bf G(k)} for all $k$.

1514: 

1515: The fact that $\mbox{\bf P(1)} \not \Rightarrow \mbox{\bf P(2)}$

1516: is trivial.  To show that $\mbox{\bf P(2)} \not \Rightarrow

1517: \mbox{\bf P(3)}$, let $\sigma_n \in S_{2n}$ be the permutation

1518: which sends $x$ to $x+n$.  Then $X^{01}(\sigma_n) = 2n(2n-1)$, and

1519: $X^{10}(\sigma_n) = 4n^2$, so that $\left | X^{01}(\sigma_n) -

1520: X^{10}(\sigma_n) \right | = o((2n)^2)$.  However, the pattern

1521: $(021)$ {\it never} appears in $\sigma_n$.  We have been unable to

1522: date to provide an analogous result for any $P(k)$ with $k

1523: > 2$.

1524: 

1525: A second, very natural question is that of the existence of

1526: \textit{perfect} $m$-symmetry: the property of having all

1527: subsequence statistics \textit{precisely} equal to their expected

1528: values.  That is, for $\sigma \in S_n$,

1529: $$

1530: X^{\tau}(\sigma) = \frac{\binom{n}{k}}{k!}

1531: $$

1532: for all $\tau \in S_m$.  For this to occur, the number of

1533: permutations of length $m$ must evenly divide $\binom{n}{m}$.  Let

1534: \textbf{D($m$)} be the property of an integer $N$ that

1535: $$

1536: m! \left | \binom{n}{m} \right. \!\! .

1537: $$

1538: It is easy to see that a permutation $\sigma \in S_n$ with perfect

1539: $m$-symmetry must have perfect $m^\prime$-symmetry for any

1540: $m^\prime \leq m$, so $n$ must satisfy \textbf{D($m^\prime$)} for

1541: all such $m^\prime$.  Let $h(m)$ be the least $n$ for which this

1542: occurs.  A quick calculation reveals that $h(2)=4$, $h(3)=9$,

1543: $h(4)=64$, and $h(5)=128$.  In fact, there is a perfect

1544: $2$-symmetric permutation on $4$ symbols: 3012.  A computer search

1545: revealed that there are exactly two $3$-symmetric permutations on

1546: $9$ symbols: 650147832 and its reverse, 238741056.  No

1547: $m$-symmetric permutation is known for $m>3$, and the question of

1548: whether such permutations exist remains open.  We conjecture that

1549: an $m$-symmetric permutation on sufficiently many symbols exists

1550: for all $m$, and believe it likely that one exists on $h(m)$

1551: symbols.

1552: 

1553: Finally, the selection of intervals as the sets which measure ``proximity''

1554: in the definition of quasirandomness was a natural but somewhat arbitrary choice.

1555: It would be worth investigating the properties of ``$(\mathcal{A},\mathcal{B})$-quasirandom''

1556: permutations for families $\mathcal{A}$, $\mathcal{B} \subset 2^{\mathbb{Z}_n}$, i.e., permutations $\sigma$ such that $\max_{A,B} D_B(\sigma(A)) = o(n)$ for $A \in \mathcal{A}$ and $B \in \mathcal{B}$.

1557: 

1558: \section*{{\sc Acknowledgements}}

1559: 

1560: The author wishes to thank Fan Chung Graham and Ron Graham for

1561: their tremendous help in formulating and attacking the problems discussed above.  He also thanks Chris Dillard, Robert Ellis, and Lei Wu for helpful discussions during the development of this work.

1562: 

1563: \section*{{\sc References}}

1564: \begin{enumerate}

1565: \item \label{Ah1} {\sc L. V. Ahlfors}, ``Complex Analysis,'' McGraw-Hill Book Co., New York, 1978.

1566: \item \label{A1} {\sc N. Alon}, Generating Pseudo-Random Permutations and Maximum Flow Algorithms, {\it Inform. Process. Lett.} {\bf 35} (1990), 201--204.

1567: \item \label{AF1} {\sc N. Alon and E. Friedgut}, On the number of permutations avoiding a given pattern, {\it J. Comb. Theory Ser. A} {\bf 89} (2000), 133--140.

1568: \item \label{AS1} {\sc N. Alon and J. H. Spencer}, ``The probabilistic method,'' Wiley-Interscience Series in Discrete Mathematics and Optimization. Wiley-Interscience [John Wiley \& Sons], New York, 2000. 

1569: \item \label{B1} {\sc M. B\'{o}na}, The solution of a conjecture of Stanley and Wilf for all layered patterns, {\it J. Comb. Theory Ser. A} {\bf 85} (1999), 96--104.

1570: \item \label{Ch1} {\sc B. Chazelle}, ``The Discrepancy Method,'' Cambridge University Press, Cambridge, 2000.

1571: \item \label{CG1} {\sc F. R. K. Chung and R. L. Graham}, Quasi-random set systems, {\it J. Amer. Math. Soc.} {\bf 4} (1991), 151--196.

1572: \item \label{CG2} {\sc F. R. K. Chung and R. L. Graham}, Quasi-random subsets of $Z\sb n$, {\it J. Combin. Theory Ser. A} {\bf 61} (1992), 64--86.

1573: \item \label{CGW1} {\sc F. R. K. Chung, R. L. Graham, and R. M. Wilson}, Quasi-random graphs, {\it Combinatorica} {\bf 9} (1989), 345--362.

1574: \item \label{D1} {\sc P. J. Davis}, ``Circulant Matrices,'' Wiley, New York, 1979.

1575: \item \label{D2} {\sc J. A. Davis, R. C. Entringer, R. L. Graham, and G. J. Simmons}, On permutations containing no long arithmetic progressions, {\it Acta Arithmetica}, {\bf 34} (1977/78), 81--90.

1576: \item \label{ES1} {\sc P. Erd\H{o}s and G. Szekeres}, A combinatorial problem in geometry, {\it Compocito Math.} {\bf 2} (1935), 464--470.

1577: \item \label{F1} {\sc H. Faure}, Discr\'{e}pance quadratique de suites infinies en dimension un. Th\'{e}orie des nombres (Quebec, PQ, 1987), 207--212, de Gruyter, Berlin, 1989. 

1578: \item \label{G1} {\sc W. T. Gowers}, A new proof of Szemer\'edi's Theorem, {\it Geometric and Functional Analysis} {\bf 11} (2001), 465--588.

1579: \item \label{G2} {\sc R. L. Graham}, personal communication.

1580: \item \label{L1} {\sc J. C. Lagarias}, Well-spaced labelling of points in rectangular grids, {\it SIAM J. Discrete Math.} {\bf 13} (2000), 521--534.

1581: \item \label{N1} {\sc H. Niederreiter}, Quasi-Monte Carlo methods and pseudo-random numbers. {\it Bull. Amer. Math. Soc.} {\bf 84} (1978), 957--1041. 

1582: \item \label{MMW1} {\sc B. D. McKay, J. Morse, and H. S. Wilf}, The distributions of the entries of Young tableaux, to appear.

1583: \item \label{Sa1} {\sc V. N. Sachkov}, ``Probabilistic Methods in Combinatorial Analysis,'' Encyclopedia of Mathematics and its Applications {\bf 56}, Cambridge University Press, Cambridge, 1997.

1584: \item \label{SS1} {\sc F. W. Schmidt and R. Simion}, Restricted permutations, {\it European J. Combin.} {\bf 6} (1985), 383--406.

1585: \item \label{S1} {\sc W. M. Schmidt}, Irregularities of distribution VII, {\it Acta Arith.} {\bf 21} (1972), 45–-50.

1586: \item \label{St1} {\sc R. P. Stanley}, Log-concave and unimodal sequences in algebra, combinatorics, and geometry. Graph theory and its applications: East and West (Jinan, 1986), 500--535, Ann. New York Acad. Sci., 576, New York Acad. Sci., New York, 1989.

1587: \item \label{St2} {\sc R. P. Stanley}, Enumerative combinatorics, Vol. 1, {\it Cambridge Studies in Advanced Mathematics} {\bf 49}, Cambridge University Press, Cambridge, 1997.

1588: \end{enumerate}

1589: 

1590: \ifx\ispreprint\undefined

1591: \pagebreak

1592: \def\baselinestretch{1}

1593: \begin{center}

1594: \begin{figure}[ht]

1595: \hspace{.7in} \input{figa.pstex_t} \caption{Diagram of

1596: implications for Theorem \ref{balance}.}

1597: \end{figure}

1598: \end{center}

1599: \else

1600: \fi

1601: 

1602: \ifx\ispreprint\undefined

1603: \pagebreak

1604: \def\baselinestretch{1}

1605: \begin{center}

1606: \begin{figure}[ht]

1607: \hspace{1.25in} \input{figb.pstex_t} \caption{Diagram of

1608: implications for Theorem \ref{qrp}.}

1609: \end{figure}

1610: \end{center}

1611: \else

1612: \fi

1613: 

1614: \end{document}

1615: