1: \documentclass[11pt]{amsart}

2: \usepackage{vmargin}

3: \usepackage{harvard}

4: \usepackage{psfig}

5: \usepackage{graphics}

6: 

7: %%%% theorem-like environments %%%%

8: 

9: \newtheorem{theorem}{Theorem}[section]

10: \newtheorem*{untheorem}{Theorem}

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

12: \newtheorem*{unlemma}{Lemma}

13: \newtheorem{proposition}[theorem]{Proposition}

14: \newtheorem*{unproposition}{Proposition}

15: \newtheorem{corollary}[theorem]{Corollary}

16: \newtheorem*{uncorollary}{Corollary}

17: 

18: \theoremstyle{definition}

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

20: \newtheorem{example}[theorem]{Example}

21: \newtheorem{notation}[theorem]{Notation}

22: 

23: \theoremstyle{remark}

24: \newtheorem{remark}[theorem]{Remark}

25: \newtheorem*{unremark}{Remark}

26: \newtheorem*{claim}{Claim}

27: 

28: %%%%roman letters in enumerate%%%%

29: 

30: \makeatletter

31: \def\romenumi{%

32:   \def\theenumi{\roman{enumi}}%

33:   \def\p@enumi{\theenumi}%

34:   \def\labelenumi{(\@roman\c@enumi)}}

35: \makeatother

36: 

37: %%%%"no proof" symbol%%%%

38: 

39: \newcommand{\noproof}{\hfill\qedsymbol}

40: 

41: %%%%new operators%%%%

42: 

43: \DeclareMathOperator{\supp}{supp}

44: \DeclareMathOperator{\re}{Re}

45: \DeclareMathOperator{\im}{Im}

46: \DeclareMathOperator{\Arg}{Arg}

47: \DeclareMathOperator{\Res}{Res}

48: \DeclareMathOperator{\hess}{\mathcal{H}}

49: \DeclareMathOperator{\proj}{proj}

50: \DeclareMathOperator{\rk}{rank}

51: \DeclareMathOperator{\cone}{\bb{\kappa}}

52: \DeclareMathOperator{\dir}{\b{dir}}

53: \DeclareMathOperator{\grad}{\nabla}

54: \DeclareMathOperator{\sign}{sgn}

55: 

56: %%%%formatting%%%%%

57: 

58: \numberwithin{equation}{section}

59: 

60: \setpapersize{USletter}

61: \setmarginsrb{20mm}{20mm}{20mm}{20mm}{12pt}{10mm}{12pt}{10mm}

62: 

63: %%%% concepts whose notation may be changed %%%%%%%

64: 

65: \def\sing{\mathcal{V}}

66: \def\torus{T}

67: \def\disk{D}

68: \def\numer{{G}}

69: \def\denom{{H}}

70: \def\mult{{\psi}}

71: \def\simp{{\Delta}}

72: \def\meas{\mu}

73: \def\direc{{\bf dir}}

74: %\def\direc{{\bb{\delta}}}

75: \def\dom{{\mathcal D}}

76: \def\logdom{{\rm log} \dom}

77: \def\rate{\gamma}

78: \def\hyper{\mathcal{H}}

79: \def\nbd{{\mathcal N}}

80: \def\one{{\bf 1}}

81: \def\zero{{\bf 0}}

82: \def\GF{{\mbox{GF-sequence}}}

83: \def\cio{C^\infty_0}

84: \def\finv{\eta}

85: \def\X{\Xi}

86: 

87: %%%%% abbreviations %%%%%

88: 

89: \def\wlog{\mbox{without loss of generality}}

90: \def\b#1{\mathbf{#1}}

91: \def\bb#1{\boldsymbol{#1}}

92: \def\w#1{\widehat{#1}}

93: \def\wh{\w{\b{w}}}

94: \def\zh{\w{\b{z}}}

95: \def\thh{\w{\bf \theta}}

96: \def\zt{{\tilde{z}}}

97: \def\qt{{\tilde{q}}}

98: \def\bt{{\tilde{b}}}

99: \def\gt{{\tilde{g}}}

100: \def\ft{{\tilde{f}}}

101: \def\psit{{\tilde{\psi}}}

102: \def\R{{\mathbb{R}}}

103: \def\CC{{\mathbb{C}}}

104: \def\RP{{\mathbb{RP}}}

105: \def\N{{\mathbb{N}}}

106: \def\Z{{\mathbb{Z}}}

107: \def\A{{\mathcal A}}

108: \def\S{{\mathcal S}}

109: \def\C{{\mathcal C}}

110: \def\B{{\mathcal B}}

111: \def\ee{\varepsilon}

112: \def\dd{\delta}

113: 

114: %%%%%%%%%%%%%%END OF TOPMATTER%%%%%%%%%%%%%%%%%%%%%%

115: 

116: \begin{document}

117: 

118: \title[Asymptotics of multivariate sequences I]{Asymptotics of multivariate sequences, part I: smooth points of the singular variety}

119: \author{Robin Pemantle}

120: \address{Department of Mathematics, Ohio State University, Columbus OH 43210}

121: \email{pemantle@math.ohio-state.edu}

122: \thanks{Research supported in  part by NSF grant DMS-9803249}

123: \author{Mark C. Wilson}

124: \address{Department of Mathematical Sciences, University of Montana,

125:  Missoula, MT 59812}

126: \email{wilsonm@member.ams.org}

127: 

128: \subjclass{Primary 05A16. Secondary 32A20, 41A60.}

129: 

130: \keywords{Generating function, recurrence, linear difference equation, 

131: contour methods, central limit, oscillating integral, Cauchy integral formula.}

132: 

133: \begin{abstract}Given a multivariate generating function

134: $F(z_1 , \ldots , z_d) = \sum a_{r_1 , \ldots , r_d} z_1^{r_1}

135: \cdots z_d^{r_d}$, we determine asymptotics for the coefficients.

136: Our approach is to use Cauchy's integral formula near singular

137: points of $F$, resulting in a tractable oscillating integral. This

138: paper treats the case where the singular point of $F$ is a smooth

139: point of a surface of poles. Companion papers will treat singular

140: points of $F$ where the local geometry is more complicated, and

141: for which other methods of analysis are not known.

142: \end{abstract}

143: 

144: \maketitle

145: 

146: \section{Introduction}

147: 

148: The generating function $F(z) := \sum_{r=0}^\infty a_r z^r$ for

149: the sequence $a_0 , a_1 , a_2 , \ldots$ is one of the most useful

150: constructions in combinatorics.  If the function $F$ has a simple

151: description, it is usually not too hard to obtain $F$ as a formal

152: power series once one understands a recursive or combinatorial

153: description of the numbers $\{ a_r \}$.  One may then analyze the

154: analytic properties of $F$ in order to obtain asymptotic

155: information about the sequence $\{ a_r \}$.  While still part art

156: and part science, this latter analytic step has become quite

157: systematized. \citeasnoun{stanley;enumerative-combinatorics1} in

158: his introduction to enumerative combinatorics gives the example of

159: the function $F(z) = \exp (z + \frac{z^2}{2})$, from which he says

160: ``it is routine (for someone sufficiently versed in complex

161: variable theory) to obtain the asymptotic formula $a_r = 2^{-1/2}

162: r^{r/2} e^{-r/2 + \sqrt{r} - 1/4}$.'' Routine, in this case, means

163: a single application of the saddle point method.  When $F$ has

164: singularities in the complex plane, the analysis is often more

165: direct: the location of the singularities and the behavior of $F$

166: near these determine almost algorithmically the asymptotic

167: behavior of the sequence $\{ a_r

168: \}$.  For those not sufficiently versed in complex variable

169: theory, two useful sources are \citeasnoun{henrici;complex-analysis2} and

170: \citeasnoun{odlyzko;asymptotic-methods}.  The transfer theorems of

171: \citeasnoun{flajolet-odlyzko;singularity-analysis}

172: encapsulate much of this knowledge in a very useful way; see also

173: \citeasnoun{wilf;GFology} for an elementary introduction.

174: 

175: When the sequence $a_0 , a_1 , a_2 , \ldots$ is replaced by a

176: multidimensional array $\{ a_{r_1 , \ldots , r_d} \}$, things

177: become much more hit and miss.  Let us use boldface to denote

178: vectors in $\CC^d$ or $\N^d$, and use multi-index notation, so

179: that $a_\b{r}$ denotes the multi-index $a_{r_1 , \ldots , r_d}$ and

180: $\b{z}^\b{r}$ denotes the product $z_1^{r_1} \cdots z_d^{r_d}$ which

181: we will sometimes write in expanded form for clarity.  The

182: generating function $F : \CC^d \to \CC$ is defined analogously to

183: the one-dimensional generating function by

184: $$F(\b{z}) = \sum_{\b{r} \in \N^d} a_\b{r} \b{z}^\b{r} .$$

185: Surprisingly, techniques for extracting asymptotics of $\{ a_\b{r}

186: \}$ from the analytic properties of $F$ were, until recently,

187: almost entirely missing.  In a survey of asymptotic methods,

188: \citeasnoun{bender;asymptotic-methods} says:

189: \begin{quote}

190: Practically nothing is known about asymptotics for recursions in

191: two variables even when a generating function is available.

192: Techniques for obtaining asymptotics from bivariate generating

193: functions would be quite useful.

194: \end{quote}

195: In the intervening 25 years, some results have appeared,

196: addressing chiefly the case where the array $\{ a_\b{r} \}$ obeys a

197: central limit theorem.  Common to all of these is the following

198: method.  Treat $\{ a_\b{r} \}$ as a sequence of

199: $(d-1)$-dimensional arrays indexed by $r_d$; show that the $n^{th}$

200: $(d-1)$-dimensional generating function is roughly the $n^{th}$

201: power of a given function; use this approximation to invert the

202: characteristic function and obtain a Central Limit Theorem.  We

203: refer to these methods as $\GF$ methods.  The other body of work

204: on multivariate sequences, which we will call the diagonal method,

205: is based on algebraic extraction of the diagonal, as found in

206:  \citeasnoun{hautus-klarner;diagonal} (  

207: see also \citeasnoun{furstenberg;diagonal} and later \citeasnoun{lipshitz;diagonal} for an algebraic description of the scope of this method; variants are described in 

208: \citeasnoun{stanley;enumerative-combinatorics2} and \citeasnoun{pippenger;equicolourable}).

209: 

210: The most fundamental $\GF$ result is probably

211: \citeasnoun{bender-richmond;limit-multivariate2},

212: with extensions appearing in later work of the same

213: authors.  \citeasnoun{flajolet-sedgewick;multivariate} present a version of the

214: same idea which holds in much greater generality.

215: \citeasnoun{gao-richmond;limit-multivariate4} go beyond the central limit case,

216: using the transfer theorems of

217: \citeasnoun{flajolet-odlyzko;singularity-analysis} to handle

218: functions that are products of powers with powers of logs. Recent

219: work of Bender and Richmond

220: \cite{bender-richmond;admissible,bender-richmond;products}

221: extends the applicability of the central limit results to many

222: problems of combinatorial interest; see also

223: \cite{hwang;stirling,hwang;convergence-rates}, where

224: more precise asymptotics are given, and

225: \citeasnoun{hwang;local-limit-deviations}, which extends some results

226: to the combinatorial schemes of

227: \citeasnoun{flajolet-soria;gaussian-exponential}. This does not

228: exhaust the recent work on the problem of multivariable

229: coefficient extraction, but does circumscribe it.

230: 

231: The present paper, together with forthcoming companion papers,

232: takes aim at a large class of multivariable coefficient extraction

233: problems, for which a fair amount of information can be read off

234: in a systematic way.  An ultimate goal (not our only goal) is to

235: systematize the extraction of multivariate asymptotics

236: sufficiently that it may be automated, say in Maple.  Everything

237: we do, we do with complex contour integration.  In this regard,

238: our methods are most similar to those of

239: \citeasnoun{bertozzi-mckenna;queueing}, who, as we do, provide a general

240: framework for harnessing the multivariable theory of residues for

241: exact and series computation of coefficients. A more detailed

242: description of our method will be given in

243: Section~\ref{ss:results}, but here is an outline.

244: 

245: \begin{quote}

246:  ~~

247: 

248: (1) Use the multidimensional Cauchy integral formula to represent

249: $a_{\b{r}}$ as an integral over a $d$-dimensional torus inside

250: $\CC^d$.

251: 

252: (2) Expand the surface of integration across a point $\b{z}$ where

253: $F$ is singular, and use the residue theorem to represent

254: $a_\b{r}$ as a $(d-1)$-dimensional integral of one-variable residues.

255: The choice of $\b{z}$ determines the directions in which asymptotics

256: may be computed.

257: 

258: (3) Put this in the form of an integral $\int \exp (\lambda

259: f(\b{z})) \psi (\b{z}) \, d\b{z}$ for which the large-$\lambda$

260: asymptotics can be read off from the theory of oscillating

261: integrals.

262: \end{quote}

263: 

264: \begin{figure}

265: \centerline{\psfig{figure=taxonomy.eps,width=6in}}

266: \vspace{10pt}

267: \caption{Classification of generating functions}

268: \label{fig:taxonomy}

269: \end{figure}

270: 

271: In the rest of this introductory section, we describe the scope of

272: our methods. Figure~\ref{fig:taxonomy} depicts a classification of generating functions and illustrates the remainder of this paragraph.  If a

273: formal power series is nowhere convergent, analytic methods are

274: useless. Among those power series converging in some neighborhood

275: of the origin, there are three possibilities: a function may be

276: entire, may have singularities around which analytic continuations

277: exist, or it may be defined only on some bounded subset of

278: $\CC^d$.  Our methods are tailored to the second class.  The third

279: class, although in some sense generic, seldom arises in any

280: problem for which the generating function may be effectively

281: described.  Incomplete asymptotic information is available via

282: Darboux' method; details of this method in the univariate case are

283: given in \citeasnoun{henrici;complex-analysis2} and

284: \citeasnoun{odlyzko;asymptotic-methods}.  The first class can

285: and does arise frequently. Our methods are simply not equipped to

286: handle entire functions, and systematizing the asymptotic analysis

287: of coefficients of entire generating functions remains an

288: important open problem.

289: 

290: For the remainder of this paper, we will assume that the formal

291: power series $F$ converges in a neighborhood of the origin and may

292: be analytically continued everywhere except a set $\sing$ of complex

293: dimension $d-1$ which we call the {\em singular variety}. The

294: point $\b{z}$ in step~2 is an element of $\sing$, and the behavior of

295: $\sing$ near $\b{z}$ greatly affects the subsequent analysis in step~3.

296: This paper addresses the case where $\b{z}$ is a smooth point of

297: $\sing$ at which $F$ has a pole.  The forthcoming companion papers

298: will address cases where $\b{z}$ is a multiple point

299: or a cone point.  We do not know whether cases

300: where $\b{z}$ is a cusp of $\sing$ arise, but if so, the subsequent

301: analysis has mostly been carried out in the work of

302: \citeasnoun{arnold-varchenko;asymptotics-integrals}.

303: 

304: The chief purpose of this study is to give a solution to the

305: problem of asymptotic evaluation of coefficients that is as

306: general as possible.  An important part of this is re-derivation

307: in a general setting of results obtainable via $\GF$ or {\em ad

308: hoc} methods. We show in Section~\ref{ss:classify} how unifying

309: these results allows us to show that our method successfully finds

310: asymptotics for every function in a certain large class.  Familiar

311: examples from this class include: lattice path counting, various

312: known generating functions for polyominos and stacked balls,

313: enumeration of Catalan trees by number of components or

314: surjections by image cardinality (see

315: \citeasnoun{flajolet-sedgewick;multivariate},

316: stopping times for certain random walks

317: (see \citeasnoun{larsen-lyons;coalescing}), as well as the examples

318: given in the $\GF$ papers of

319: \citeasnoun{bender;central-local} and

320: \citeasnoun{bender-richmond;limit-multivariate2}: ordered set partitions

321: enumerated by number of blocks, permutations enumerated by rises,

322: and Tutte polynomials of recursive sequences of graphs.

323: 

324: Nevertheless, our pursuit of this problem was also motivated by

325: some specific applications which we mention briefly now and

326: discuss more thoroughly later.  These are cases where known

327: methods do not suffice to obtain complete asymptotic information.

328: There is a class of tiling enumeration problems for which an

329: explicit three variable rational generating function may be

330: obtained.  This class includes the Aztec Diamond domino tilings of

331: \citeasnoun{cohn-elkies-propp;aztec}. Asymptotics in the so-called {\em

332: region of fixation} are obtained from analysis of the smooth

333: points of $\sing$ (Theorem~\ref{th:higher d} below), while

334: asymptotics in the region of positive entropy are derived from

335: analysis of the cone point.  \citeasnoun{cohn-pemantle;fixation}  applies a

336: cone point analysis to a tiling enumeration problem for which the

337: only previous results are some pictures via simulation

338: (http://www.math.harvard.edu/\~{}cohn/picture.gif) . Another

339: motivation has been to solve the general multivariable linear

340: recursion. Depending on whether one allows forward recursion in

341: some of the variables, one obtains either rational or algebraic

342: generating functions.  The general rational function may have any

343: of the types of singularities mentioned above: smooth points,

344: nodes, cones, cusps, branchpoints, etc. Even the simple rational

345: generating function $1 / (3 - 3z - w + z^2)$ of

346: Example~\ref{ex:cubic} requires two separate analyses in order to

347: get asymptotics in all directions. We will see that

348: Theorem~\ref{th:simple} gives asymptotics in one region, while

349: Theorem~\ref{th:2d} is required for other directions.

350: 

351: Asymptotics derived near smooth pole points nearly always exhibit

352: central limit behavior.  Smooth pole points are the topic of this

353: first paper, and are exactly the case to which existing methods

354: may apply.  While one function of this paper is to lay foundations

355: for the cases in which the singularity is more complicated, there

356: are several ways in which it improves upon available analyses of

357: the smooth case.

358: 

359: First, most of the existing results assume that the singular point

360: $\b{z} \in \sing$ has positive real coordinates, and that it is

361: strictly minimal in a sense defined in the next section.  This

362: assumption often holds when the coefficients $\{ a_\b{r} \}$ are

363: nonnegative reals, though it will fail if, for example, there is

364: any periodicity.  The assumption always fails when the

365: coefficients $\{ a_\b{r} \}$ have mixed signs, as is the case for

366: example with the generating functions $(1 - zw)/(1 - 2zw +w^2)$

367: and $1 / (1 - 2zw + w^2)$ for the Chebyshev polynomials of the

368: first and second kinds \cite[page 50]{comtet;advanced}. $\GF$ methods may

369: be adapted to some of these situations. Indeed, the presentation

370: of these methods by \citeasnoun[Theorem~9.7]{flajolet-sedgewick;multivariate}

371: accomplishes this adaptation in great generality. But certainly

372: there are cases such as the rational generating function

373: $1 / (1 - z - w + \beta zw)$, where the points $\b{z}$ with given

374: moduli form a continuum and standard $\GF$ methods are not

375: sufficient.

376: 

377: Second, our methods obtain automatically a full asymptotic

378: expansion of $a_{r_1 , \ldots , r_d}$ in decreasing powers of the

379: indices $r_j$.  This is certainly not inherent in the existing

380: results, whose relatively short proofs involve inversion of the

381: characteristic function (see however

382: \citeasnoun{hwang;stirling} and \citeasnoun {hwang;central-limit-deviations} for

383: something in this direction). The expansion to $n$ terms is

384: completely effective in terms of the first $n$ partial derivatives

385: of $1/F$ at $\b{z}$, as is the error bound.

386: 

387: Third, these results explicitly cover the case where the pole at

388: $\b{z}$ has order greater than 1.  The behavior in this case is not

389: according to the central limit theorem.  The only existing work

390: addressing this case is \citeasnoun{gao-richmond;limit-multivariate4},

391: and they require

392: nonnegativity assumptions, as mentioned above.  In the case where

393: $F = G^k$ is an exact power, one could attempt first to solve the

394: problem for $G$ and then to take the $k$-fold convolution.  This

395: is much harder than the present approach, as may be seen by the

396: rather involved computation in \citeasnoun{cohn-elkies-propp;aztec}.

397: 

398: Fourth, the potential for increasing the scope to new applications

399: seems greater for contour methods than for $\GF$ methods.  The

400: contour method reduces the asymptotic problem to the problem of an

401: oscillating integral near a singularity, which can almost

402: certainly be done.  By contrast, the $\GF$ method requires first

403: an understanding of the sequence of $(d-1)$-dimensional generating

404: functions arising from the given $d$-dimensional generating

405: function, and then another result in order to transfer this

406: information to asymptotics of the coefficients $a_\b{r}$.

407: 

408: Fifth, although our results in the case of smooth pole points are

409: often similar to those obtained by $\GF$ methods, our hypotheses

410: are quite different.  In Section~\ref{ss:classify} we show how our

411: hypotheses may be universally established for functions that

412: generate nonnegative values and are meromorphic through their

413: domain of convergence.

414: 

415: Finally, we compare our method to recent results from the diagonal

416: method.  It is known \cite{lipshitz;diagonal} that the diagonal

417: sequence

418: $a_{n,n,\ldots ,n}$ of a multivariate sequence with rational

419: generating function has a generating function satisfying a linear

420: differential equation over rational functions.  Much is known

421: about how to compute this equation (see for example

422: \citeasnoun{chyzak-salvy;symbolic-integration}).

423: If one wants asymptotics on the diagonal, or in

424: any direction where the coordinate ratios are rational numbers

425: with small denominators, then these methods give results that are

426: in theory at least as good as ours.  The method, however, is

427: inherently non-uniform in the direction, so there is no hope of

428: extending it to larger sets of directions, which is what we

429: accomplish in the present work.

430: 

431: The remainder of the paper is organized as follows.  In the next

432: section we set forth notation and define the terms necessary to

433: state the main results of the paper.  The main results are stated

434: in Section~\ref{ss:results}, and examples are given.  The next

435: section contains a proof of these results, modulo the computation

436: of some oscillating integrals.  This computation is carried out in

437: Section~\ref{ss:oscillate}.  Section~\ref{ss:classify} outlines

438: some details of taxonomy and discusses universality of the method

439: of complex contour integration.  The final section states some

440: open problems.

441: 

442: \section{Notation and Preliminaries}

443: 

444: The main results of this paper give asymptotics valid under

445: certain geometric assumptions on $\sing$ and computable from some

446: quantities that are in turn effectively computable from the

447: generating function $F$. Thus in addition to setting out basic

448: notation, we need to define some terms related to the geometry of

449: $\sing$ and some quantities associated with $F$.

450: 

451: \subsection{Notation}

452: 

453: Throughout the paper, $F = \sum a_\b{r} \b{z}^\b{r}$ will denote a

454: function on $\CC^d$ analytic in a neighborhood of the origin. The

455: (open) domain of convergence of the power series will be denoted

456: $\dom$. For $\b{z} \in \CC^d$, let $\torus (\b{z})$ denote the torus

457: consisting of points $\b{w}$ with $|w_j| = |z_j|$ for $1 \leq j \leq

458: d$ and let $\disk (\b{z})$ denote the closed polydisk of points

459: $\b{w}$ with $|w_j| \leq |z_j|$ for $1 \leq j \leq d$.  Recall (see

460: \citeasnoun{hormander;intro}) that the domain $\dom$ is a union of tori

461: $\torus (\b{z})$ and is logarithmically convex, that is, the set

462: $$\logdom := \{ \b{x} \in \R^d : (e^{x_1} , \ldots , e^{x_d})

463:    \in \dom \}$$

464: is a convex subset of $\R^d$ and is an order ideal, that is, it is

465: closed under $\leq$ in the coordinatewise partial order.

466: 

467: We assume throughout that $F = \numer / \denom$, where both

468: $\numer$ and $\denom$ are analytic in a neighborhood of

469: $\disk (\b{z})$ for some point $\b{z}$.  In particular, every

470: meromorphic function satisfies this condition

471: \footnote{The greater

472: generality allows us to cover examples such as the generating

473: function for self-avoiding random walks

474: \cite{chayes-chayes;self-avoiding} or percolation paths in the

475: subcritical regime \cite{campanino-chayes-chayes;percolation}.  In

476: these cases, all the work is in showing the function is

477: meromorphic in a neighborhood of

478: $\disk (\b{z})$. Without further knowledge, the authors then

479: conclude central limit behavior.}.  The set where

480: $\denom$ vanishes will be denoted $\sing$.  Many of our examples will

481: be in dimension 2, in which case we will often use $z$ and

482: $w$ in place of $z_1$ and $z_2$, use $(z,w)$ in place of

483: $\b{z}$, and use $(r,s)$ in place of $(r_1 , r_2)$.  We sometimes

484: need to treat $\CC^d$ as $\CC^{d-1} \times \CC$ (although symmetry of

485: the coordinates is preserved most of the time).  Accordingly, when

486: the dimension is greater than~2, we use $\widehat{\b{z}}$ to denote

487: $(z_1 , \ldots , z_{d-1})$.  Partial derivatives will be denoted

488: $H_1$ for $\frac{\partial H}{\partial z_1}$ and so forth;

489: in dimension~2 we will also use $H_z$ and $H_w$.

490: 

491: As is usual for asymptotic analyses, we let $f \sim g$ denote

492: $f / g \to 1$, with the limit taken at infinity unless otherwise

493: specified.  The function $f$ is said to be {\em rapidly

494: decreasing} if $f(x) = O(x^{-N})$ for every $N$, and is said to be

495: {\em exponentially decreasing} if $f(x) = O(e^{-cx})$ for some $c

496: > 0$. We also use the symbol ``$\sim$'' to denote asymptotic

497: expansion. Thus

498: $$f \sim \sum b_n g_n$$

499: is normally taken to mean that $f - \sum_{n=0}^N b_n g_n = o(b_N

500: g_N)$, where $b_n\in \mathbb{C}$ and $\{ g_n \}$ is a fixed sequence of functions such that

501: $g_{n+1} = o(g_n)$ for each $n$. We broaden this to allow $b_n = 0$

502: when $n \neq 0$, so that the remainder term need only be $o(g_n)$

503: and not $o(b_n g_n)$.  In particular, if

504: $$f(x) \sim g(x) \cdot \sum_{n = 0}^\infty c_n x^{-n}$$

505: with $c_0 = 1$, then we say we have obtained a full asymptotic

506: expansion for $f$ in decreasing powers of $x$ with leading term

507: $g$.

508: 

509: \subsection{Geometry of $\sing$} \label{ss:geom}

510: 

511: As in the one-dimensional case, the points of $\sing$ nearest the

512: origin are the most important.  Accordingly we define a point $\b{z}

513: \in \sing$ to be {\em minimal} if $\sing \cap \disk (\b{z}) \subseteq

514: \torus (\b{z})$; we say that $\b{z}$ is {\em locally minimal}

515: if the analogous relation holds with $\sing$ replaced by a neighborhood of $\b{z}$ in $\sing$. Divide the minimal

516: points of $\sing$ into three types. Say that

517: $\b{z}$ is {\em strictly minimal}, {\em finitely minimal} or {\em

518: toral}, according to whether the cardinality of

519: $\sing \cap \disk (\b{z})$ is 1, finite, or infinite.  When infinite, the

520: intersection must be uncountable.  If $\b{z}$ is a minimal point of

521: $\sing$ then the interior of $\disk (\b{z})$ is contained in $\dom$,

522: so the assumption that $\numer$ and $\denom$ are analytic on a

523: neighborhood of $\disk (\b{z})$ is just a little stronger than what

524: is true automatically.

525: 

526: A {\em simple pole} of $F$ is a point $\b{z} \in \sing$ where $\denom$

527: vanishes to order 1.  Equivalently, the gradient $\grad \denom$

528: does not vanish.  Let $\b{z}$ be a simple pole of $F$ and assume for

529: specificity that $\denom_d$ is nonzero at $\b{z}$.   By the implicit

530: function theorem, there is a neighborhood of $\b{z}$ where

531: $\sing$ may be parametrized by $z_d = g(z_1 , \ldots , z_{d-1})$ for

532: some analytic function $g$.  We will always use $g$ to denote this

533: parametrization.

534: 

535: We will see later (in the proof of Theorem~\ref{th:classify}) that

536: under some hypotheses on $F$, minimal points of $\sing$ are always

537: found in the positive real orthant.  A relation true in complete

538: generality is the following.

539: 

540: \begin{lemma} \label{lem:dir}

541: Let $\b{z}$ be a simple pole of $F$ and suppose that $z_d \denom_d$

542: does not vanish there.  If $\b{z}$ is locally minimal then for all

543: $j < d$, the quantity $z_j \denom_j / (z_d \denom_d)$ is real and

544: nonnegative.

545: \end{lemma}

546: 

547: \begin{proof} Given $\theta$ and $j$, let $\b{z}^{(\theta)}$

548: be the result of varying $\b{z}$ by multiplying the $j^{th}$

549: coordinate by $e^{i \theta}$ and adjusting the last coordinate so

550: as to remain on $\sing$ (that is, \hfill \\ $z^{(\theta)}_d = g(z_1 ,

551: \ldots , z_{j-1} , z_j e^{i \theta} , z_{j+1} , \ldots ,

552: z_{d-1})$). Differentiating the relation $\denom (\b{z}^{(\theta)})

553: = 0$ implicitly with respect to

554: $\theta$ at 0 yields

555: \begin{equation} \label{eq:minimal}

556: i z_j \denom_j + \denom_d {d z^{(\theta)}_d \over d \theta} = 0 .

557: \end{equation}

558: By minimality of $\b{z}$, we know that the modulus of

559: $z^{(\theta)}_d$ has a minimum at $\theta = 0$, hence

560: $(d z^{(\theta)}_d / d \theta) / z_d$ is purely imaginary.

561: Plugging this into~(\ref{eq:minimal}) proves that $z_j \denom_j /

562: (z_d \denom_d)$ is real.  If $z_j \denom_j / (z_d \denom_d) =

563: -\beta < 0$ then $\sing$ has a tangent vector at $\b{z}$ in the

564: direction $- z_j e_j - \beta z_d e_d$, where $e_j$ is the

565: $j^{th}$ coordinate vector.  This contradicts minimality.

566: Hence $z_j \denom_j / (z_d \denom_d) \geq 0$.   \end{proof}

567: 

568: \begin{definition}

569: Define $\direc (\b{z})$ to be the equivalence class of (complex)

570: scalar multiples of the vector $(z_1 \denom_1 , \ldots , z_d

571: \denom_d)$, defined whenever $z_j \denom_j$ does not vanish for

572: all $j$. By the previous lemma, when $\b{z}$ is a minimal pole of

573: $F$ with nonzero coordinates, $\direc (\b{z})$ is a well defined

574: element of

575: $\RP^{d-1}$.

576: \end{definition}

577: %

578: The importance of $\direc$ is that analysis of $F$ near $\b{z}$

579: yields asymptotic information about $a_\b{r}$ with $\b{r}

580: \in \direc (\b{z})$.  The function $\direc$ appears in $\GF$

581: method literature as ${\bf m}$.  When $\b{z} \in \partial \dom$ is

582: on the boundary of the domain of convergence, $\direc (\b{z})$ is

583: the normal to the support hyperplane of the convex set $\logdom$

584: at the point $(\log |z_1| , \ldots , \log |z_d|)$.

585: 

586: We now define a few more quantities associated with $F$ and $g$.

587: Again, we will reserve the names of these functions, so as not to

588: burden the notation with subscripts and arguments.  If $\b{z}$ is a

589: simple pole of $F$ with $z_d \denom_d$ not vanishing there, define

590: a function $\psi$ on a neighborhood of $\widehat{\b{z}}$ by

591: \begin{equation} \label{eq:psi}

592: \psi (\widehat{\b{w}}) = - \lim_{w \rightarrow g(\widehat{\b{w}})}

593:    (w - g(\widehat{\b{w}})) {F (\widehat{\b{w}} , w) \over w} .

594: \end{equation}

595: Suppose now that $\widehat{\b{w}} \in \torus (\widehat{\b{z}})$ and

596: write

597: $w_j = z_j e^{i \theta_j}$.  For fixed $\b{r}$ with $r_d \neq 0$,

598: define a function $f$ on a neighborhood of $\widehat{\b{z}}$ in

599: $\torus (\widehat{\b{z}})$ by

600: \begin{equation} \label{eq:f}

601: f (\widehat{\b{w}}) = \log \left ( {g (\widehat{\b{w}}) \over

602:    g(\widehat{\b{z}})} \right ) + i \sum_{j=1}^{d-1} {r_j \over r_d}

603:    \theta_j .

604: \end{equation}

605: 

606: We will be parametrizing integrals over $\torus (\widehat{\b{z}})$

607: by ${\bf \theta}$, so we will want the above function expressed in

608: terms of $\widehat{\bf \theta}$.  We therefore compose with the

609: map $M$ taking $\widehat{\bf \theta}$ to $\widehat{\b{w}}$ defined

610: by

611: $M(\theta_1 , \ldots , \theta_{d-1}) = (z_1 e^{i \theta_1} ,

612: \ldots , z_{d-1} e^{i \theta_{d-1}})$, and define the functions

613: $\gt: = g \circ M , \ft := f \circ M , \psit := \psi \circ M$.

614: 

615: Although it is not obvious yet, $\ft$ will always vanish at

616: $\zero$  to at least two orders (Lemma~\ref{lem:stat} below), and

617: the hypothesis $Q \neq 0$ in Theorem~\ref{th:simple} is equivalent

618: to $\ft$ having nonvanishing quadratic term.  For ease of

619: reference, Table~\ref{table1} summarizes the foregoing definitions,

620: stratified by how many times the given data $\numer$ and $\denom$

621: have been manipulated.

622: 

623: %\pagebreak

624: 

625: \begin{table}

626: \label{table1}

627: \caption{Reserved notation in remainder of this article}

628: \begin{tabbing}

629: Given \= information: \\

630: \> the function $F$ in the form $\numer / \denom$ \\[3ex]

631: First level: \\

632: \> $g$ parametrizes the zero set, $\sing$ of $\denom$ \\

633: \> $\direc(\b{z})$ is the coordinatewise product $(\grad \denom)\cdot(\b{z})$

634:    in projective space \\[3ex]

635: Second level: \\

636: \> $\psi$ is the residue in $z_d$ of $F / z_d$ at points $(\widehat{\b{z}} ,

637: g(\widehat{\b{z}}))$ \\

638: \> $f$ is $\log g$, plus a term linear in $\log z_j$ and depending

639:    on $\b{r}$. \\[3ex]

640: Third level: \\

641: \> $\psit , \gt$ and $\ft$ are $\psi, g$ and $f$ expressed in

642:    terms of ${\bf \theta}$

643: \end{tabbing}

644:  ~~\\[2ex]

645: \end{table}

646: 

647: \section{Statement of results, with examples} \label{ss:results}

648: 

649: Before going on, we pause to state a prototype of our results in

650: the simplest possible setting, namely where the number of

651: variables is~2, the functions $\numer$ and $g$ are as

652: nondegenerate as possible, and only the leading term asymptotic is

653: given.  The proof is in Section~\ref{ss:proofs}.

654: 

655: \begin{theorem} \label{th:simple}

656: Let $F = \numer / \denom$ be a meromorphic function of two

657: variables, not singular at the origin.  Define

658: $$Q (z,w) := - w^2 \denom_w^2 z \denom_z - w \denom_w z^2 \denom_z^2 - w^2 z^2

659:    \left ( \denom_w^2 \denom_{zz} + \denom_z^2 \denom_{ww}

660:    - 2 \denom_z \denom_w \denom_{zw} \right ) .$$

661: Then

662: $$a_{r,s} \sim {\numer (z,w) \over \sqrt{2 \pi}} z^{-r} w^{-s}

663:    \sqrt{- w \denom_w \over s Q}$$

664: uniformly as $(z,w)$ varies over a compact set of strictly

665: minimal, simple poles of $F$ on which $Q$ and $\numer$ are

666: nonvanishing, and $(r,s) \in \direc (z,w)$.

667: \end{theorem}

668: 

669: \noindent{\em Remarks:} Usually the expression in the radical will

670: be positive real, as will the coefficients $a_{rs}$.  The result

671: is true in general, though, as long as the square root is taken to

672: be $-w \denom_w$ times the principal root of $Q / (- w

673: \denom_w^3)$. Also note that when $(r,s) \in \direc (z,w)$ then

674: the expression

675: $w \denom_w / s$ is coordinate-invariant, that is, equal to $z \denom_z / r$.

676: Thus the given expression for $a_{r,s}$ has the expected symmetry.

677: 

678: \begin{example}[Lattice paths] \label{ex:lattice} ~~ \end{example}

679: 

680: Let $a_{r,s}$ be the number of nearest-neighbor paths from the

681: origin to $(r,s)$ moving only north, east and northeast; these are

682: sometimes called {\em Delannoy numbers}~\citeasnoun[page

683: 185]{stanley;enumerative-combinatorics2}. The generating function

684: is $F(z,w) = 1 / (1 - z - w - zw)$. The zero set $\sing$ of

685: $\denom = 1 - z - w - zw$ is given by

686: $w = (1-z)/(1+z)$, and the minimal points of $\sing$ are those where

687: $w \in [0,1]$.  With the help of relations that hold when $\b{z} \in \sing$

688: we may compute as follows.

689: \begin{eqnarray*}

690: \denom_z & = & - 1 - w \\

691:    -z \denom_z & = & 1 - w \\

692:    Q & = & (1-z) (1-w) (1-zw) \\

693:    {z \denom_z \over w \denom_w} & = & {1-w \over 1-z} =

694:    {1 - w^2 \over 2w}

695: \end{eqnarray*}

696: with $\denom_w$ and $-w \denom_w$ given by reversing $z$ and $w$.

697: As $z$ varies over $[\ee , 1-\ee]$, the functions $Q$ and $\numer

698: := 1$ do not vanish.  The minimal pair $(z,w)$ that solves $(r,s)

699: \in \direc (z,w)$ is given by $z = (\sqrt{r^2 + s^2} - s)/r$ and

700: $w = (\sqrt{r^2 + s^2} - r)/s$.  Theorem~\ref{th:simple} then

701: gives

702: \begin{eqnarray*}

703: a_{rs} & \sim & \left ({\sqrt{r^2 + s^2} - s \over r} \right)^{-r}

704:    \left ( {\sqrt{r^2 + s^2} - r \over s} \right )^{-s}

705:    \sqrt{1 \over 2 \pi} \sqrt{{1-z \over s} {1 \over 1- zw}}

706:    \\[1ex]

707: & = & \left ({\sqrt{r^2 + s^2} - s \over r} \right)^{-r}

708:    \left ( {\sqrt{r^2 + s^2} - r \over s} \right )^{-s}

709:    \sqrt{1 \over 2 \pi} \sqrt{r s \over (r+s-\sqrt{r^2 + s^2})^2

710:    \sqrt{r^2 + s^2}} \, ,

711: \end{eqnarray*}

712: uniformly when $r/s$ and $s/r$ remain bounded.  In particular,

713: when $r=s=n$, this gives the following formula for the $n^{th}$

714: diagonal coefficient (which may alternatively be obtained by

715: computing the diagonal generating function $(1 - 6s+ s^2)^{-1/2}$

716: according to the method given in

717: \citeasnoun[Section~6.3]{stanley;enumerative-combinatorics2}:

718: $$(\sqrt{2} - 1)^{-2n} \sqrt{1\over 2 \pi} {2^{-1/4} \over

719:   2 - \sqrt{2}} \, .$$

720: 

721: The computations in Theorem~\ref{th:simple} in terms of the values

722: and derivatives of $\numer$ and $\denom$ are explicit.  As we

723: state more general theorems, it becomes cumbersome and in fact

724: obfuscating to give formulae for the expansion coefficients

725: directly in terms of derivatives of $\numer$ and $\denom$.  This

726: is one reason we have already introduced the functions in

727: Table~\ref{table1}. It should be emphasized, however, that while we use

728: higher level quantities in the statements of subsequent theorems,

729: each expansion coefficient can be computed from finitely many

730: derivatives of $\numer$ and $\denom$.  We begin with a relatively

731: explicit computation for the general two-variable case.

732: 

733: For $k$ at least 2, we define constants

734: \begin{eqnarray}

735: A_+ (k,l) & := & {1 \over k} \Gamma \left ({l+1 \over k} \right )

736:  \label{eq:plus} \\[2ex]

737: A (k,l) & := & {1 \over k} \Gamma \left ( {l+1 \over k} \right )

738:    \left ( 1 + e^{\sign\Arg (c_k) i \pi (l - {l+1 \over k} )} \right ) \,

739:    \mbox{ if } k \mbox{ is odd}, \label{eq:k odd} \\[2ex]

740: A (k,l) & := & {2 \over k} \Gamma \left ( {l+1 \over k} \right )

741:    \mbox{ if } k,l \mbox{ are even},\label{eq:k even} \\[2ex]

742: A (k,l) & := & 0 \mbox{ if } k \mbox{ is even and } l \mbox{ is

743:    odd.} \nonumber

744: \end{eqnarray}

745: Let

746: $$y (x) = f(x)^{1/k} = c_k^{1/k} x \left ( 1 + {f(x) - c_k x^k

747:    \over c_k x^k} \right )^{1/k} \, ,$$

748: where $c_k$ is the first nonvanishing Taylor coefficient of $f (x)

749: = \sum_{j=k}^\infty c_j x^j$ and the argument of

750: $c_k^{1/k}$ is taken between $-\pi/(2k)$ and

751: $\pi / (2k)$.  Let $\finv$ denote the inverse function to $y$

752: and let $\{ b_j \}$ be the Taylor coefficients of

753: $(\psit \circ \finv) \cdot \finv'$. Clearly each $\{ b_j \}$

754: is determined by finitely many partial derivatives of $\numer$ and

755: $\denom$, and the index $l_0$ of the first nonvanishing $b_l$ is

756: the same as the order of vanishing of $\psit$ at 0.  The

757: coefficients $b_l$ are easily computed from the coefficients

758: $\bt_j := \psit^{(j)} (0) / j!$ and $c_j := \ft^{(j)} (0) / j!$;

759: in particular, if $\ft \sim c_k x^k$ near 0 then

760: \begin{equation} \label{eq:b}

761: b_{l_0} = \bt_{l_0} c_k^{-1/k} \, .

762: \end{equation}

763: 

764: \begin{theorem} \label{th:2d}

765: Let $F = \numer / \denom = \sum a_{rs} z^r w^s$ have a strictly

766: minimal, simple pole at $(z,w)$. Let $k$ be the order of vanishing

767: of $\ft$ at 0.  Let $l_0$ be the order to which $\numer$ vanishes

768: near $(z,w)$ on $\sing$, that is, the largest $l$ such that $G(z',w')

769: = O(|z-z'|^l + |w-w'|^l)$ as $(z' , w') \to (z,w)$ in $\sing$.  Then

770: there is a full asymptotic expansion

771: \begin{equation}\label{eq:full 2d}

772: a_{r,s} \sim {1 \over 2 \pi} z^{-r} w^{-s} \sum_{l \geq l_0}

773:    \A (k,l) b_l s^{-(l+1) / k} \, ,

774: \end{equation}

775: where $\A (k,l)$ denotes $A (k,l)$ if $\im \{ c_k \} \geq 0$ and

776: $\overline{A(k,l)}$ otherwise.  The expansion is uniform as $(z,w)$

777: varies over a compact set of strictly minimal poles with $(r,s)

778: \in \direc (z,w)$ and $k$ and $l_0$ not changing.

779: \end{theorem}

780: 

781: \begin{figure}

782: \scalebox{0.6}{\input{p1-fig2.tex}}

783: \vspace{10pt}

784: \caption{$\sing$ for Example~\ref{ex:cubic}}

785: \label{fig:cuberoot}

786: \end{figure}

787: 

788: \begin{example}[Cube root asymptotics] \label{ex:cubic} 

789: 

790: Let $F(z,w) = 1 / (3 - 3z - w + z^2)$.  The set $\sing$ is the set

791: $\{ w = z^2 - 3z + 3 \}$ and $g(z) = z^2 - 3z + 3$.  The point

792: $(1,1)$ is in $\sing$, indicating that the maximal exponential growth

793: rate will be zero.  Indeed, for directions above the diagonal,

794: Theorem~\ref{th:simple} or~\ref{th:2d} may be used at the minimal

795: points $\{ (z, g(z)) : 0 < z < 1 \}$, while each direction below

796: the diagonal corresponds to a pair of complex minimal points

797: fitting the hypotheses of Corollary~\ref{cor:finitely}; the result

798: is that the coefficients decay exponentially at a rate that is

799: uniform over compact subsets of directions not containing the

800: diagonal.

801: 

802: The interesting behavior is near the diagonal.  The relevant

803: minimal point is $(1,1)$, where $z^r w^s \equiv 1$ and the decay

804: is sub-exponential.  Computing $\ft''(0)$ via

805: equation~(\ref{eq:ft''}) below gives

806: $$\ft'' (z) =  -3 {z (z^2 - 4z + 3) \over (z^2 - 3z + 3)^2} \, .$$

807: This vanishes when $z = 1$, and computing further, we find that

808: $\ft$ vanishes to order exactly 3 here, with $c_3 := \ft'''(0) /

809: 3! = i$.  Along with $\psit (0) = 1$, this then results in an

810: asymptotic expansion whose leading term is given by

811: $$a_{r,r} \sim {1 \over 2 \pi} A (3,0) i^{-1/3} (1 + e^{- i \pi /

812:    3}) r^{-1/3} = {\Gamma(2/3) \over 6 \sqrt{3} \pi} r^{-1/3}\, .$$

813: In Section~\ref{ss:open} we discuss the question of computing

814: asymptotics ``in the gaps'' so as to be able to conclude that

815: $\limsup \log a_\b{r} / \log |\b{r}| = -1/3$ or even $\limsup |\b{r}|^{1/3}

816: a_\b{r} = {\Gamma(2/3) \over 6 \sqrt{3} \pi}$.

817:  \end{example}

818:  

819: For more than two variables a result holds similar to the

820: two-variable result.

821: %

822: \begin{theorem} \label{th:higher d}

823: Let $F = \numer / \denom = \sum a_\b{r} \b{z}^\b{r}$ have a strictly

824: minimal, simple pole at $\b{z}$.  Suppose $z_d \denom_d$ does not

825: vanish.  If the Hessian of $\ft$ at $\b{z}$ is nonsingular, then

826: there is an expansion

827: $$a_\b{r} \sim \b{z}^{-\b{r}} \sum_{l \geq l_0} C_l r_d^{(1-d-l)/2}$$

828: where $l_0$ is the degree to which $\numer$ vanishes on $\sing$ near

829: the point $\b{z}$ . When

830: $\numer$ does not vanish at $\b{z}$ then $l_0 = 0$ and

831: $$C_0 = (2 \pi)^{(1-d)/2} \hess^{-1/2} { \numer (\b{z}) \over z_d H_d}$$

832: where $\hess$ is the determinant of the Hessian at $\b{z}$.

833: \end{theorem}

834: 

835: \begin{example}[Domino tilings] \label{ex:arctic} 

836: 

837: Random perfect tilings of planar regions by dominos have been a

838: subject of some interest, since the analysis by \citeasnoun{fisher;dimer}

839: of this

840: model for dimer packing uncovered an exact expression for the

841: partition function of the ensemble.  A generating function is

842: given in \citeasnoun{cohn-elkies-propp;aztec} which allowed the authors to

843: determine, after some cumbersome analysis, which parts of a

844: diamond-shaped region (a union of lattice squares approximating

845: the region $|x| + |y| \leq k$) were asymptotically deterministic

846: and which contained randomness in the limit as the edge size of

847: the diamond grew.

848: 

849: An easier analysis in the region of non-randomness is available

850: via Theorem~\ref{th:higher d} together with a slightly more

851: informative generating function than was used by

852: \citeasnoun{cohn-elkies-propp;aztec}.

853: In particular, let

854: $$F(x,y,z) = \sum_{t=0}^{\infty} \sum_{|r| + |s| \leq t}

855:    a_{r,s,t} x^r y^s z^t$$

856: be the generating function for the probability $a_{r,s,t}$ that

857: the tile covering position $(r,s)$ of a random diamond of size $t$

858: will be horizontal.  For brevity, we omit formal descriptions of

859: the diamond and its indexing.  We remark that the use of negative

860: indices (for each fixed $t$, the sum $\sum_{|r| + |s| \leq t}$ is

861: a polynomial in $x, x^{-1}, y$ and $y^{-1}$) does not require any

862: alterations in the theory (see \citeasnoun{cohn-pemantle;fixation} for

863: justification), and that the natural way to parametrize directions

864: is by the pair $(r/t , s/t)$ which varies over the diamond $|r/t|

865: + |s/t| = 1$. From \citeasnoun{cohn-elkies-propp;aztec} or from the

866: generation algorithm in \citeasnoun{gessel-ionescu-propp}, one

867: finds

868: $$F(x,y,z) = {z/2 \over 1 - (x + x^{-1} + y + y^{-1}) z/2

869:    + z^2} \, .$$

870: \citeasnoun{cohn-pemantle;fixation} show that whenever $(r,s,t)$ satisfy

871: $$t = \sqrt{r^2 + s^2 + 2 \sqrt{r^2 + 1} \sqrt{s^2 + 1} - s} \, ,$$

872: then there is a smooth minimal point $(x,y,z)$ on the pole

873: manifold of $F$ for which $(r,s,t) \in \direc (x,y,z)$, yielding

874: exponential decay in the direction $(r,s,t)$.  The set of

875: directions so parametrized turns out to be the region between the

876: diamond $|r/t| + |s/t| = 1$ and the inscribed circle $(r/t)^2 +

877: (s/t)^2 = 1/2$.  Thus they recover the description of the region

878: of non-randomness as the complement of the inscribed circle.  They

879: also obtain descriptions of the region of fixation for related

880: tiling problems in which no other analysis has been carried out.

881: \end{example}

882: 

883: The extension of all of the above results to finitely minimal

884: points is routine.

885: 

886: \begin{corollary} \label{cor:finitely}

887: Suppose $\b{z}$ is a finitely minimal point of $\sing$ with

888: $\sing \cap \torus (\b{z}) = \{ \b{z}_1 , \ldots , \b{z}_n \}$.  Then

889: $$a_\b{r} \sim \sum_{j=1}^n E_j (\b{r})$$

890: where $E_j (\b{r})$ is the asymptotic expression given by the

891: previous theorems with $\b{z} = \b{z}_j$.  In other words, if there

892: are finitely many points on $\sing \cap \torus (\b{z})$, then sum the

893: contributions as if each were strictly minimal. 

894: \noproof

895: \end{corollary}

896: 

897: \begin{example}[Chebyshev polynomials] \label{ex:Tsch} 

898: 

899: Let $F(z,w) = 1 / (1 - 2zw + w^2)$ be the generating function for

900: Chebyshev polynomials of the second kind \cite{comtet;advanced}; of

901: course asymptotics for these are well known and easy to derive by

902: other means.  To use Corollary~\ref{cor:finitely}, first find the

903: minimal points for the direction $(r,s)$, which are $(i (\beta -

904: \beta^{-1})/2 , i \beta)$ for $\beta = \pm \sqrt{s-r \over s+r}$.

905: Computing $Q = 4 a^2 (1 - a^2)$ and summing the two contributions

906: then gives

907: $$a_{rs} \sim \sqrt{2 \over \pi} (-1)^{(s-r)/2} \left ( {2r \over

908:    \sqrt{s^2 - r^2}} \right )^{-r} \left ( \sqrt{s-r \over s+r}

909:    \right )^{-s} \sqrt{s+r \over r(s-r)}$$

910: when $r+s$ is even and zero otherwise, uniformly as $r/s$ varies

911: over compact subsets of $(0,1)$.

912: 

913: \end{example}

914: 

915: \section{Proofs of main results} \label{ss:proofs}

916: 

917: Half of each theorem is easy and follows directly from Cauchy's

918: formula

919: \begin{equation} \label{eq:cauchy}

920: a_\b{r} = \left ( {1 \over 2 \pi i} \right )^d

921:    \int_T \b{w}^{-\b{r} - \one} F(\b{w}) \, d\b{w}

922: \end{equation}

923: where the multi-exponent $\b{r} - \one$ means $(r_1 - 1 , \ldots ,

924: r_d - 1)$.  Indeed, if $\b{z}$ is a minimal point of $\sing$ then

925: letting $T$ approach $\torus (\b{z})$ from the inside, we see that

926: $|\b{z}^\b{r}| a_\b{r}$ does not increase exponentially.  If, furthermore,

927:  the hyperplane through $(\log |z_1| , \ldots , \log |z_d|)$ normal

928: to $\b{r}$ is not a support hyperplane for $\logdom$, then some $\b{x}

929: \in \logdom$ has $\b{x} \cdot \b{r} > (\log |z_1| , \ldots , \log |z_d|)

930: \cdot \b{r}$, and integrating on the torus $T(e^\b{x})$ shows that

931: $|\b{z}^\b{r}| a_\b{r}$ decreases exponentially.  All the work,

932: therefore, is in showing the converse, namely that when the

933: hyperplane normal to $\b{r}$ is a support hyperplane, then

934: $\b{z}^{-\b{r}}$ does give the right exponential order for $a_\b{r}$.

935: This is done by evaluating $a_\b{r}$.

936: 

937: Theorems~\ref{th:simple}--\ref{th:higher d} all begin with the

938: reduction of an iterated Cauchy integral to an oscillating

939: integral in one fewer dimension.

940: 

941: \begin{lemma} \label{lem:X}

942: Let $\b{z}$ be a strictly minimal simple pole of $F=\numer/\denom$.

943: Assume that $z_d \denom_d \neq 0$.  For a neighborhood

944: $\widetilde{\nbd}$ of $\zero$ in $\R^{d-1}$ define a quantity

945: \begin{equation}\label{eq:X}

946: \X := (2 \pi)^{1-d} \b{z}^{-\b{r}} \int_{\widetilde{\nbd}} \exp (- r_d \ft

947:   (\widehat{\bf \theta})) \psit (\widehat{\bf \theta}) \,d\widehat{\bf \theta}.

948: \end{equation}

949: Then the quantity

950: $$|\b{z}^\b{r}| \left | a_\b{r} - \X \right |$$

951: decreases exponentially as $\widetilde{\nbd}$ remains fixed and

952: $\b{r} \to \infty$.

953: \end{lemma}

954: 

955: \begin{proof}

956: For $\ee \in (0 , |z_d|)$, let $T$ be the torus $T(\b{z})$ shrunk in

957: the last coordinate by $\ee$, that is, the set of $\b{w}$ for which

958: $|w_j| = |z_j|$, $j < d$ and $|w_d| = |z_d| - \ee$.  Write

959: Cauchy's formula as an iterated integral

960: \begin{equation} \label{eq:iter}

961: a_\b{r} = \left ( {1 \over 2 \pi i} \right )^d

962:    \int_{T(\widehat{\b{z}})} \widehat{\b{w}}^{-\widehat{\b{r}} - \one}

963:    \left [ \int_{\C_1} w_d^{-r_d} F(\b{w}) \,

964:    {d w_d \over w_d} \right ] \, d \widehat{\b{w}}  \, .

965: \end{equation}

966: Here $\C_1$ is the circle of radius $|z_d| - \ee$.  Let $K

967: \subseteq T(\widehat{\b{z}})$ be a compact set not containing

968: $\widehat{\b{z}}$.  For each fixed $\widehat{\b{w}} \in K$, the function

969: $F(\widehat{\b{w}} , \cdot)$ has radius of convergence greater than

970: $|z_d|$.   Hence the inner integral in equation~(\ref{eq:iter})

971: is $O(|z_d|+ \delta)^{-r_d}$ for some $\delta > 0$.  By continuity

972: of the radius of convergence,we may integrate over $K$ to see that

973: $$|\b{z}^\b{r}| \int_{K \times \C_1} \b{w}^{-\b{r} - \one} F(\b{w}) \, d\b{w}$$

974: decreases exponentially.  Thus if $\nbd$ is any neighborhood of

975: $\widehat{\b{z}}$ in $T(\widehat {\b{z}})$, the quantity

976: $$|\b{z}^\b{r}| \left | a_\b{r} - \left ( {1 \over 2 \pi i} \right )^d

977:    \int_\nbd \widehat{\b{w}}^{-\widehat{\b{r}} - \one}

978:    \left [ \int_{\C_1} {F(\b{w}) \over w_d^{r_d + 1}} \, dw_d

979:    \right ]  d \widehat{\b{w}} \right |$$

980: decreases exponentially.  Thus we have reduced the problem to an

981: integral over a neighborhood of $\widehat{\b{z}}$.

982: 

983: Near $\b{z}$ there is a parametrization $w_d = g(\widehat{\b{w}})$ of

984: $\sing$.  Let $\C_2$ be the circle of radius $|z_d| + \ee$.  Then

985: when $\nbd$ is sufficiently small compared to $\ee$, the image of

986: $\nbd$ under $g$ is disjoint from $\C_2$.  Fix such a neighborhood.

987: For any $\widehat{\b{w}} \in \nbd$, the function $F(\widehat{\b{w}} ,

988: \cdot)$ has a single simple pole in the annulus bounded

989: by $\C_1$ and $\C_2$, occurring at $g(\widehat{\b{w}})$.  The

990: residue in the last variable of $F$ at $g(\widehat{\b{w}})$ is equal

991: to

992: \begin{equation} \label{eq:defres}

993: R(\widehat{\b{w}}) := - \psi (\widehat{\b{w}}) g(\widehat{\b{w}})^{-r_d}

994: \end{equation}

995: where $\psi$ is defined in~(\ref{eq:psi}). Therefore, for each

996: fixed $\widehat{\b{w}} \in \nbd$,

997: $$\int_{\C_1} {F(\b{w}) \over w_d^{r_d + 1}} \, d w_d

998:    = \int_{\C_2} {F(\b{w}) \over w_d^{r_d + 1}} \, d w_d

999:    - 2 \pi i R(\widehat{\b{w}}) .$$

1000: But $|\b{z}^\b{r} \int_{\C_2} F(\b{w}) dw_d / \b{w}^{\b{r} + \one}|$ is

1001: bounded by a constant multiple of $(1 + \ee / |z_d|)^{-r_d}$ (the

1002: constant depending on the maximum of $F$ on $\C_2$) and hence

1003: $|\b{z}^\b{r}| |a_\b{r} - X|$ is exponentially decreasing, where

1004: \begin{eqnarray} \label{eq:res}

1005: X & = & (2 \pi i)^{1-d} \int_\nbd (\widehat{\b{w}})^{-\widehat{\b{r}}

1006:    - 1} g(\widehat{\b{w}})^{-r_d} \psi (\widehat{\b{w}}) \,

1007:    d \widehat{\b{w}} \\[1ex]

1008: & = & (2 \pi i)^{1-d} \b{z}^{-\b{r}} \int_\nbd {\widehat{\b{w}}^{-\widehat{\b{r}}}

1009:    \over \widehat{\b{z}}^{-\widehat{\b{r}}}} {d \widehat{\b{w}} \over

1010:    \prod_{j=1}^{d-1} w_j} \left ( {g(\widehat{\b{w}}) \over g(z_d)}

1011:    \right )^{-r_d} \psi (\widehat{\b{w}}) \nonumber

1012: \end{eqnarray}

1013: Changing variables to

1014: $w_j = z_j e^{i \theta_j}$ and $dw_j = i w_j d\theta_j$ turns the

1015: quantity $X$ into

1016: $$(2 \pi)^{1-d} \b{z}^{-\b{r}} \int_{\widetilde{\nbd}} \prod_{j=1}^{d-1}

1017:    e^{-i r_j \theta_j} \psit (\widehat{\bf \theta}) \left ( {

1018:    g(\widehat{\b{w}}) \over g(\widehat{\b{z}})} \right )^{-r_d}

1019:    \, d\widehat{\bf \theta} $$

1020: and plugging in the definitions of $f$ and $\ft$ at~(\ref{eq:f})

1021: above yields

1022: $$(2 \pi)^{1-d} \b{z}^{-\b{r}} \int_{\widetilde{\nbd}} \exp (- r_d

1023:    \ft (\widehat{\bf \theta})) \psit (\widehat{\bf \theta}) \,

1024:    d\widehat{\bf \theta}$$

1025: which is none other than $\Xi$.   \end{proof}

1026: 

1027: \begin{unremark} It is possible to compute from Cauchy's

1028: integral formula in a more coordinate-free way as follows.  There

1029: is a unique holomorphic $(d-1)$-form $\omega_F$ on $\sing$ for which

1030: $\omega \wedge d\denom = \numer \, dz_1 \wedge \cdots \wedge dz_d$.

1031: Let $\Omega$ be a ($d+1$)-manifold that is a homotopy from a small

1032: torus to a torus at infinity.  Then $M := \Omega \cap \sing$ is a

1033: $(d-1)$-manifold and $a_\b{r} = (2 \pi i)^{-d} \int_M \b{w}^{-\b{r} - \one} dF$

1034: in the sense of currents, which is none other than $\int_M

1035: \b{w}^{\b{r} - \one} \omega_F$.  See \citeasnoun{kenyon-pemantle} for a

1036: more thorough discussion of the foregoing.  The manifold $M$ is

1037: any member of a certain homology class in $\sing$ with the coordinate

1038: axes removed, and choosing $M$ to pass through the stationary

1039: phase point for the integrand replicates the selection of $\b{z}$

1040: with $\b{r} \in \direc(\b{z})$.  Although more canonical, the

1041: coordinate-free method is less suitable for explicit computation,

1042: so we do not pursue it further here.  Suffice it to point out that

1043: the conclusion of Theorem~\ref{th:higher d} may of course be

1044: written in terms more evidently symmetric, as was done in

1045: Theorem~\ref{th:simple}.

1046: \end{unremark}

1047: 

1048: Equation~(\ref{eq:X}) is easily recognized as the standard form

1049: for an oscillating integral.  The only unusual feature is that the

1050: phase is neither real nor purely imaginary.  This presents no

1051: difficulties, but it does necessitate the statement of a result in

1052: Section~\ref{ss:oscillate} that is a little different from the

1053: usual results on purely oscillating integrals, found in, for example,

1054: \citeasnoun{stein;oscillatory-integrals} or

1055: \citeasnoun{bleistein-handelsman;asymptotic-expansions-integrals}.  We first

1056: establish that $\widehat{\bf \theta} = \zero$ is a stationary

1057: phase point for the function $\ft$ when $\b{r} \in \direc (\b{z})$.

1058: %

1059: \begin{lemma} \label{lem:stat}

1060: The quantity $\ft (\zero)$ always vanishes.  If $\b{r} \in \direc

1061: (\b{z})$ then $\grad \ft (\zero) = \zero$ and the real part of $\ft$

1062: has a strict minimum at $\zero$.

1063: \end{lemma}

1064: 

1065: \begin{proof} The first statement is immediate.  To prove

1066: the second, let $j \leq d-1$ and see from the definition of $f$

1067: that

1068: $$r_d f_j (\widehat{\b{z}}) = {r_d g_j (\widehat{\b{z}}) \over

1069:    g(\widehat{\b{z}})} + {r_j \over z_j}.$$

1070: By definition of $\direc$, the ratio $r_j / (z_j \denom_j)$ is

1071: some constant $c$ independent of $j$, hence

1072: $$c^{-1} r_d f(\b{z}) = g_j (\b{z}) \denom_d (\b{z}) + \denom_j (\b{z}) .$$

1073: The right hand side of this is the derivative of $\denom (w_1 ,

1074: \ldots , w_{d-1} , g(\widehat{\b{w}}))$ with respect to $w_j$ at

1075: $\widehat{\b{z}}$.  By definition of $g$ this vanishes, and hence

1076: $f_j (\widehat{\b{z}}) = 0$.  But $\ft_j (\zero) = i z_j f_j (\b{z})$, so

1077: the gradient of $\ft$ must vanish at $\zero$.  Finally, observe

1078: that $\re \{ \ft (\widehat{\bf \theta}) \} = -\log |\gt

1079: (\widehat{\bf \theta}) / z_d|$.  By strict minimality of $\b{z}$,

1080: the modulus of $g(\widehat{\b{w}}) = \gt (

1081: \widehat{\bf \theta})$ is greater than $|z_d|$ for any

1082: $\widehat{\b{w}} \in \torus (\widehat{\b{z}})$.   \end{proof}

1083: 

1084: We now prove Theorems~\ref{th:simple},~\ref{th:2d}

1085: and~\ref{th:higher d} in reverse order.  We see from

1086: Lemma~\ref{lem:X} that proving any of these theorems amounts to

1087: evaluating the quantity $\X$ in equation~(\ref{eq:X}).  From

1088: Lemma~\ref{lem:stat} we see that $\zero$ is a stationary point for

1089: the function $\ft$ as long as $\b{r} \in \direc (\b{z})$.  The

1090: function $\ft$ is in general complex valued, but we will see in

1091: Theorem~\ref{th:nondeg} that it may be treated as if it were real

1092: valued, given the strict minimality of the zero guaranteed by

1093: Lemma~\ref{lem:stat} and the nonsingularity hypothesis. In

1094: particular the leading term of the integral in~(\ref{eq:X}) is

1095: $(2 \pi)^{(d-1)/2} \psit (\zero) r_d^{(1-d)/2}$ divided by the

1096: product of the square roots of the eigenvalues of the Hessian.

1097: Once we have identified $\psit (\zero) = \psi (\zero)$ as $\numer

1098: (\zero) / (z_d \denom_d)$, the theorem follows directly from

1099: Theorem~\ref{th:nondeg}.

1100: 

1101: Theorem~\ref{th:2d} follows from the more explicit asymptotic

1102: development given in Corollary~\ref{cor:full 2d}. Finally, to

1103: prove Theorem~\ref{th:simple}, it remains to compute the quantity

1104: $\ft'' (0)$ in terms of the partial derivatives of $\denom$.

1105: First we compute the derivatives of $g$.

1106: 

1107: \begin{lemma} \label{lem:der g}

1108: In a neighborhood of $(z,w)$, $\psi$ and the derivatives of $g$

1109: are as follows.

1110: \begin{eqnarray}

1111: g'(z) & = & -{\denom_z \over \denom_w} \label{eq:g'} \\[2ex]

1112: g''(z) & = & - {1 \over \denom_w}  \left [ \denom_{zz} - 2

1113:    {\denom_z \over \denom_w} \denom_{zw} + {\denom_z^2 \over

1114:    \denom_w^2} \denom_{ww} \right ] . \label{eq:g''} \\[2ex]

1115: \psi (z) & = & {\numer (z,w) \over - w \denom_w(z,w)} . \nonumber

1116: \end{eqnarray}

1117: \end{lemma}

1118: 

1119: \begin{proof} Differentiate the equation $\denom(z,g(z)) = 0$

1120: to get $\denom_z + g'(z) \denom_w = 0$ which is the same

1121: as~(\ref{eq:g'}).  Differentiate again to get

1122: $$\denom_{zz} + 2 g' \denom_{zw} + g'' \denom_w + (g')^2 \denom_{ww} = 0$$

1123: and use~(\ref{eq:g'}) to eliminate $g'$, giving~(\ref{eq:g''}).

1124: The formula for $\psi$ follows from the definitions of $\psi$ and

1125: of the partial derivative.

1126: \end{proof}

1127: 

1128: \noindent{\sc Proof of Theorem}~\ref{th:simple}~{\sc via direct

1129: computation:}  We know from Lemma~\ref{lem:stat} that $\ft$

1130: vanishes to order at least two at 0.  To compute $\ft''(0)$,

1131: observe first that $\ft'' - \log \gt$ is linear in $\theta$, so

1132: $\ft'' = (\log \gt)''$.  When $Z = z e^{i \theta}$, we

1133: have $(d/d\theta) = iZ (d/dZ)$, so

1134: $$\ft'' = i Z {d \over dZ} \left ( iZ {d \log g \over dZ} \right )

1135:    = - Z {d \over dZ} \left ( {Z g' \over g} \right ) \, .$$

1136: Expanding this yields

1137: \begin{equation} \label{eq:ft''}

1138: \ft'' = - Z {g' + Z g'' \over g} + {Z^2 (g')^2 \over g^2} \, .

1139: \end{equation}

1140: 

1141: By our assumption, $G$ does not vanish at $(z,w)$, so as long as

1142: $\ft'' (0) \neq 0$, we may use Theorem~\ref{th:2d} to

1143: conclude that the leading term asymptotic for $a_{r,s}$ is the

1144: $k=2, l=0$ term of~(\ref{eq:full 2d}).  The term $b_0$ there is

1145: equal to

1146: $$\psit(0) \eta' (0) = \psi (z) \sqrt{2 / \ft''(0)} =  {\numer (z,w)

1147:    \over - w \denom_w (z,w)} \sqrt{2 \over \ft'' (0)} .$$

1148: Thus from Theorem~\ref{th:2d},

1149: $$a_{r,s} \sim {A (2,0) \over 2 \pi} z^{-r} w^{-s} {\numer (z,w)

1150:    \over w \denom_w (z,w)} \sqrt{2 \over s \ft'' (0)} .$$

1151: Now evaluate this using the value $A (2,0) = \sqrt{\pi}$ and

1152: equation~(\ref{eq:ft''}) along with~(\ref{eq:g'})

1153: and~(\ref{eq:g''}) to obtain

1154: $$a_{r,s} \sim {1 \over \sqrt{2 \pi}} z^{-r} w^{-s} {\numer (z,w)

1155:    \over w \denom_w (z,w)} \sqrt{(-w \denom_w (z,w))^3 \over s Q}$$

1156: where

1157: $$Q = (- w \denom_w (z,w))^3 \ft''(0) = (- w \denom_w (z,w))^3

1158:    z {-g'(z) - z g''(z) \over g(z)} + {z^2 (g'(z))^2 \over (g(z))^2} \,

1159:    .$$

1160: With the help of Lemma~\ref{lem:der g} we see (using $g(z) = w$)

1161: that

1162: $$Q = (- w \denom_w)^3 \left [ -z {\denom_z \over -w

1163:    \denom_w} - z^2 {1 \over -w \denom_w}

1164:    \left (\denom_{zz} - 2 {\denom_z \over \denom_w} \denom_{zw}

1165:    + {\denom_z^2 \over \denom_w^2} \denom_{ww} \right )

1166:    + {z^2 \denom_z^2 \over w^2 \denom_w^2} \right ] \, ,$$

1167: evaluated at $(z,w)$, which simplifies to the expression in

1168: Theorem~\ref{th:simple}.  We see also that the nonvanishing

1169: hypotheses on $Q$ is enough to guarantee $\ft''(0) \neq 0$, which

1170: finishes the proof of Theorem~\ref{th:simple}.

1171: \noproof

1172: 

1173: \section{Some oscillating integrals} \label{ss:oscillate}

1174: 

1175: The oscillating integrals we require are integrals over a

1176: neighborhood of zero in $\R^d$ of the complex-valued integrand:

1177: $$\int_\nbd \exp (- \lambda f (\b{x})) \psi (\b{x}) \, d\b{x}$$

1178: where $f(\zero) = 0, \grad f (\zero) = \zero$ and $\re \{ f

1179: \} \geq 0$.  They are not difficult to compute, but since the

1180: standard references assume $f$ is either real or purely imaginary,

1181: we sketch the development of these results.  We mostly follow the

1182: exposition of \citeasnoun{stein;oscillatory-integrals}, adapting

1183: it to complex-valued phase functions and simplifying it to take

1184: advantage of the decay of the magnitude of the integrand in this

1185: case.

1186: 

1187: We begin with one-dimensional results.  Let $\cio$ denote the

1188: class of smooth functions with compact support.  The following

1189: proposition is a well known consequence of Watson's Lemma (see,

1190: for example, \citeasnoun[Ch.~2, Theorem~1]{wong;asymptotic-integrals}).

1191: %

1192: \begin{proposition} \label{pr:laplace}

1193: Let $\psi \in \cio (\R)$ and denote $b_j = \psi^{(j)} (0) / j!$.

1194: Then as $\lambda \to \infty$, there is an asymptotic development

1195: $$\int_0^\infty \exp (- \lambda x^k) \psi (x) \, dx \sim

1196:    \sum_{l=0}^\infty A_+ (k,l) b_l \lambda^{-(l+1)/k},$$

1197: where, as in~(\ref{eq:plus}),

1198: $$A_+ (k,l) := k^{-1} \Gamma \left ( {l+1 \over k} \right ) .$$

1199: \noproof

1200: \end{proposition}

1201: 

1202: We extend this to more general one-sided integrals by a complex

1203: change of variables. Given any analytic, complex-valued function

1204: $f$ on an interval $[0,B]$, suppose that $f(0) = 0$, that

1205: $f' \neq 0$ on $(0,B]$, and let $k \geq 1$  be the minimal so that

1206: $f^{(k)} (0) \neq 0$.  Let $\psi \in \cio$ vanishing to order $l

1207: \geq 0$ at 0.  Denote $c_j = f^{(j)} (0) / j!$ and $b_j =

1208: \psi^{(j)} (0) / j!$.  The real part of $c_k$ is necessarily

1209: nonnegative.  Define a function $y$ on $[0,B]$ by

1210: $$y(x) = f(x)^{1/k} = c_k^{1/k} x \left ( 1 + {f(x) - c_k x^k

1211:    \over c_k x^k} \right )^{1/k} \, ,$$

1212: where the argument of $c_k^{1/k}$ is between $-\pi/(2k)$ and $\pi

1213: / (2k)$.  The quantity $f(x) - c_k x^k$ is $O(x^{k+1})$ near zero,

1214: so $y$ is analytic near 0, and, in particular, is a diffeomorphism

1215: between $[0,B]$ and a contour $\gamma$ from 0 to some $B^*$.  Let

1216: $F$ invert $y$. The derivatives of

1217: $F$ at 0 are easy to compute formally and the first $j+1$ starting

1218: from the $k^{th}$ depend only on the first $j$ coefficients of $f$

1219: starting at $c_k$. Define

1220: \begin{eqnarray}

1221: \psi^* & = & (\psi \circ F) \cdot F' \; ; \nonumber \\

1222: b^*_j & = & \psi^*\,^{(j)} (0) / j! \; . \label{eq:bs}

1223: \end{eqnarray}

1224: %

1225: \begin{theorem} \label{th:one-sided}

1226: Let $f$ be analytic (complex-valued) on an interval

1227: $[0,B]$. Assume that $f(0) = 0$, that $f' \neq 0$ on $(0,B]$,

1228: and $\re \{ f \}$ has a strict minimum at 0.  Let $k \geq 2$ be

1229: minimal such that $f^{(k)} (0) \neq 0$ and $m$ be minimal so that

1230: the real part of $f^{(m)} (0)$ does not vanish. Let $\psi \in

1231: \cio$, let $l$ be minimal such that $\psi^{(l)} (0) \neq 0$,

1232: and denote $c_j := f^{(j)} (0) / j!$, $b_j := \psi^{(j)} (0) /

1233: j!$. Define $b^*_j$ as in~(\ref{eq:bs}).  Then there is an

1234: asymptotic development

1235: \begin{equation} \label{eq:series}

1236: \int_0^B \exp (- \lambda f(x)) \psi (x) \, dx \sim

1237:    \sum_{j=l}^\infty A_+ (k,j) b^*_j \lambda^{-(j+1)/k} .

1238: \end{equation}

1239: The constant in the $O(\lambda^{-(N+1)/k})$ term depends continuously (only) on the derivatives of $f$ and $\psi$ up to $(N+1) m / k - 1$.

1240: \end{theorem}

1241: 

1242: \begin{proof} Changing variables to $y = f(x)^{1/k}$,

1243: the integral becomes

1244: $$\int_\gamma \exp (- \lambda y^k) \psi^* (y) \, dy ;$$

1245: the curve $\gamma$ is the image of $[0,B]$ under $y$, so

1246: $\gamma' (0) = c_k^{1/k}$ and $\gamma$ remains in the

1247: right half plane, strictly except at 0.  For $0 < N < M$ write

1248: $\psi^*$ as $P_M + y^{M+1} R_M$, where $P_M$ is a polynomial of

1249: degree $M$ and $R_M$ is bounded; this can be done since $\psi^*$

1250: may be approximated by a degree $M$ polynomial to within

1251: $O(y^{M+1})$ at 0.

1252: 

1253: First, evaluate

1254: $$\int_\gamma \exp (- \lambda y^k) P_M (y) \, dy$$

1255: by moving the contour.  Replace $\gamma$ by two line segments, the

1256: first of which goes along the positive real axis to some distance

1257: $\ee$ and the second of which is strictly in the right half plane

1258: (we assumed $\re \{ f \} > 0$ except at 0). The integral along the

1259: second segment is exponentially small since the integrand is.

1260: Hence the combined contribution is the series~(\ref{eq:series})

1261: out to the $j=M$ term.

1262: 

1263: Next, bound

1264: $$\left | \int_\gamma \exp (- \lambda y^k) y^{M+1} R_M (y) \, dy

1265:    \right | \, .$$

1266: With $C$ representing different constants in different lines, we

1267: now observe that on $\gamma$ we have $\re \{ - y^k \} < - C

1268: |y|^m$. Thus, parametrizing $\gamma$ by arc-length, an upper bound

1269: is given by

1270: $$\int_0^\infty \exp (- \lambda C t^m) t^{M+1} C |R_M (\gamma

1271:    (t))| \, dt .$$

1272: This is easily seen to be bounded above by

1273: $C \lambda^{-(M+2)/m}$ where $C$ depends on the first $M$

1274: derivatives of $f$ and $\psi$.  Choosing $M \geq m(N+1)k - 1$ we

1275: have a remainder term that is $O(\lambda^{-(N+1)/k})$, proving the

1276: theorem. 

1277: \end{proof}

1278: 

1279: The value of a two-sided integral follows as a corollary.

1280: %

1281: \begin{corollary} \label{cor:full 2d}

1282: Assume the hypotheses of Theorem~\ref{th:one-sided}, with $f$ now

1283: defined on an interval $[-B,B]$.  Then there is an asymptotic

1284: development

1285: \begin{equation}\label{eq:two-sided}

1286: \int_{-B}^B \exp (- \lambda f(x)) \psi (x) \, dx \sim

1287:    \sum_{j=l}^\infty A (k,j) b^*_j \lambda^{-(j+1)/k} .

1288: \end{equation}

1289: with $A (k,j)$ given by~(\ref{eq:k odd}) and~(\ref{eq:k even}).

1290: The bounds on the remainder terms each depend continuously on finitely many

1291: derivatives of $f$ and $\psi$ on $[-B,B]$.

1292: \end{corollary}

1293: 

1294: \begin{proof} The two-sided integral is the sum of two

1295: one-sided integrals on intervals $[0,B]$ and $[-B,0]$.  The

1296: integral over $[-B,0]$ may be written as an integral over $[0,B]$

1297: of the function $\exp (- \lambda f(-x)) \psi (-x) \, dx$. With

1298: $b_l^*$ still denoting the coefficients resulting form the application

1299: of Theorem~\ref{th:one-sided} to the first integral, let

1300: $\check{b_l^*}$ denote the coefficients when

1301: Theorem~\ref{th:one-sided} is applied to the second integral.  In

1302: order to add the two integrals, we write $\check{b_l^*}$ in terms

1303: of $b_l^*$ by means of the following routine computation.

1304: 

1305: Let $c_k := c^k e^{i \alpha}$ with $c > 0$ and $|\alpha| \leq \pi

1306: / 2$ and define the analytic quantity $R$ so that

1307: $$y(x) = \left [ c_k x^k (1 + R(x))^k \right ]^{1/k} =

1308:    c e^{i \alpha / k} x (1 + R(x)) .$$

1309: If $k$ is odd, then then the hypothesis $\re \{ f

1310: \} \geq 0$ implies that $c_k$ is purely imaginary.  We have

1311: $$\check{y} (x) = \left [ - c_k x^k (1 + R(-x))^k \right ]^{1/k} =

1312:    c e^{- i \alpha / k} x (1 + R(-x)) = - y(-x) e^{- 2 i \alpha / k}

1313:    .$$

1314: Writing $\finv$ for the inverse function to $y$ and $\check{\eta}$

1315: for the inverse function to $\check{y}$ we then have

1316: $$\check{\finv} (x) = - \finv (-e^{2 i \alpha / k} x) \, .$$

1317: Hence, letting $\C_l [ \cdot ]$ denote the coefficient of

1318: $y^l$,

1319: \begin{eqnarray*}

1320: \check{b_l^*} & = & \C_l \left [ \psi ( - \check{\finv} (x))

1321:    \cdot \check{\finv}' (x)\right ] \\[1ex]

1322: & = & \C_l \left [ \psi ( \finv (-e^{2 i \alpha / k} x))

1323:    \cdot e^{2 i \alpha / k} \cdot \finv' (-e^{2 i \alpha / k} x) \right ]

1324: \end{eqnarray*}

1325: and thus

1326: $$\check{b_l^*} = (-1)^l e^{2 i \alpha (l+1) / k} b_l^* \, .$$

1327: When $k$ is even, the computation is similar but easier, resulting

1328: in

1329: $$\check{b_l^*} = (-1)^l b_l^* \, .$$

1330: Now observe that if $k$ is odd, hence $c_k$ is purely imaginary,

1331: then $e^{2 i \alpha (l+1) / k} = e^{\pm i \pi (l+1) / k}$

1332: according to the sign of the argument of $c_k$.  Setting $A (k,l)

1333: = (1 + (-1)^l) A_+ (k,l)$ if $k$ is even and

1334: $(1 + e^{\sign\Arg (c_k) i \pi (l+1) / k} A_+ (k,l)$

1335: if $k$ is odd, we recover the definition in~(\ref{eq:k odd})

1336: and~(\ref{eq:k even}) and prove the Corollary.

1337: \end{proof}

1338: %

1339: \begin{theorem} \label{th:nondeg}

1340: Let $f$ be a smooth complex-valued function on a neighborhood of

1341: $\zero$ in $\R^d$ such that $\re \{ f \} \geq 0$ with equality only at $\zero$.  Suppose further that $\grad f(\b{0})=0$, and that the Hessian (matrix of second partials) of $f$ has eigenvalues with positive real parts. Let $\hess$ denote the Hessian determinant at 0. Then for $\psi \in \cio$,  there is an asymptotic expansion

1342: $$\int \exp (- \lambda f (\b{x})) \psi (\b{x}) \, d\b{x} \sim

1343:    \sum_{j \geq l} C_j \lambda^{-(l+d)/2}$$

1344: where $l$ is the degree of vanishing of $\psi$ at $\zero$. If $l =

1345: 0$ then $C_0 = \psi (\zero) (2 \pi)^{d/2} \hess^{-1/2}$. The

1346: choice of square root is determined by $\hess^{-1/2} =

1347: \prod_{j=1}^d \mu_j^{-1/2}$ where $\mu_j$ are the eigenvalues of

1348: the Hessian and the principal square root is taken in each case.

1349: \end{theorem}

1350: 

1351: \begin{proof} Let $Q = \sum_{i,j=1}^d q_{i,j} z_i z_j$

1352: be the quadratic form determined by the Hessian at the origin.

1353: Denote the eigenvalues of $Q$ by $\{ \mu_j : 1 \leq j \leq d \}$

1354: and note that each $\mu_j$ has nonnegative real part.

1355: 

1356: Step 1: change coordinates to make $f$ exactly equal to the

1357: quadratic form $Q$. Indeed since $f(\b{x}) = Q(\b{x})/2 + O(|\b{x}|^3)$,

1358: and the Hessian is nondegenerate, there is a locally smooth change

1359: of variables $\{ x_j (\b{z}) : 1 \leq j \leq d \}$ such that $f(\b{z})

1360: = Q(\b{x} (\b{z}))/2$ and the Jacobian at the origin is 1.

1361: 

1362: Step 2: normalize by $\hess^{1/2}$.  For any quadratic form $Q$

1363: there is a linear change of variables $\b{y} (\b{x})$ such that

1364: $Q(\b{x}) =

1365: \sum_{j=1}^d y_j^2$.  The change of variables matrix $P$

1366: satisfies $P P^T = M(Q)$, the symmetric matrix representing $Q$.

1367: Changing variables to $\b{y}$ introduces an integrating factor of

1368: $\det P$ which is a square root of $\hess$ since $M(Q)$ is just the

1369: Hessian.  Let $\nbd'$ be the region of integration over which

1370: $\b{y}$ varies when $\b{z}$ varies over an appropriately small

1371: neighborhood of $\zero$.

1372: 

1373: Step 3: Expand $\psit$ into monomials.  The function $\psi$ has

1374: now become $\psit$, where $\psit(\zero) = \hess^{-1/2} \psi

1375: (\zero)$ and the sign of the square root will be chosen later.  We

1376: may expand $\psit$ into monomials, using the same argument as in

1377: the proof of Theorem~\ref{th:one-sided} to show the remainder term

1378: can be made

1379: $O(|\b{y}|^N)$ for any $N$.  It remains to evaluate the integral over the

1380: region of integration, $\nbd'$ of

1381: $$\int_{\nbd'} \exp (- \lambda \sum_{j=1}^d y_j^2) \psit (\b{y}) \, d\b{y}$$

1382: when $\psit$ is a monomial.

1383: 

1384: Step 4: move the region of integration to the real $d$-space. Let

1385: $\nbd''$ be the projection of $\nbd'$ onto $\R^d$ by setting the

1386: imaginary part to zero.  We claim that changing the region of

1387: integration from $\nbd'$ to $\nbd''$ alters the integral by an

1388: amount rapidly decreasing in $\lambda$.  To show this, let

1389: $\Omega$ be the region $\{ \re \{ \b{x} \} + i t \im \{ \b{x} \} :

1390: \b{x} \in \nbd' , t \in [0,1] \}$.  The boundary of $\Omega$

1391: (as a manifold) is composed of $\nbd', \nbd''$

1392: (with opposite signs) together with $S := \{ \re \{ \b{x} \} + i t

1393: \im \{ \b{x} \} : \b{x} \in \partial \nbd' , t \in [0,1] \}$.  For any

1394: $d$-form $\omega$, $\int_\Omega d\omega = \int_{\partial \Omega}

1395: \omega$.  When $\omega = \exp (- \lambda \sum_{j=1}^d \mu_j y_j^2)

1396: \b{y}^\b{r} dy_1 \wedge \cdots \wedge dy_d$ is a holomorphic $d$-form,

1397: we see that $d\omega$ vanishes (being the sum of $\partial /

1398: \partial \overline{z_j}$ terms) so that

1399: $$\int_{\nbd'} \omega = \int_{\nbd''} \omega + \int_S \omega .$$

1400: We know that $\re \{ \sum_j \mu_j y_j^2 \}$ is bounded away from 0

1401: on $\partial \nbd'$, and its minimal value on $S$ lies on

1402: $\partial \nbd'$, hence the integral over $S$ decays

1403: exponentially.

1404: 

1405: Step 5: evaluate the integral.  Factoring $\int_{\nbd''} \b{y}^\b{r}

1406: \exp (- \lambda \sum_{j=1}^d y_j^2)$ into one-dimensional

1407: integrals and plugging into Proposition~\ref{pr:laplace} yields an

1408: asymptotic expansion whose leading term (when $l=0$) is equal to

1409: $(2 \pi)^{d/2} \psi (\zero) \hess^{-1/2}$.  When $f (\b{z})$ is the function $\sum_{j=1}^d z_j^2$, then the positive square root is taken.  The

1410: choice of square root must be continuous in the analytic topology

1411: on functions having nondegenerate Hessians and having eigenvalues

1412: with positive real parts, and the only such choice is the product

1413: of the principal square roots of the eigenvalues of the Hessian.

1414: \end{proof}

1415: 

1416: \section{Classification of cases} \label{ss:classify}

1417: 

1418: For purposes of classification some natural questions are:

1419: \begin{enumerate}

1420: \romenumi

1421: \item what are all possible local geometries of minimal points of $\sing$?

1422: \item which of these can be handled by variants of the methods in this paper?

1423: \item are these sufficient to yield a good approximation to $a_\b{r}$ no

1424: matter what the direction, $\b{r} / |\b{r}|$, and no matter which

1425: generating function in the class, say, of functions meromorphic in

1426: a neighborhood of their domain of convergence?

1427: \end{enumerate}

1428: To make the last question more concrete, consider the simplest

1429: possible example, namely binomial coefficients, where $F =

1430: 1/(1-z-w)$ and $\sing$ is a complex line.  There are no singular

1431: points here, but how do we know that as $(z,w)$ varies over

1432: minimal points of $\sing$, the direction $\direc (z,w)$ will cover

1433: all of $\RP^1$?

1434: 

1435: This question will be answered by Theorem~\ref{th:classify}, but

1436: first we need to add some detail to the geometric discussion begun

1437: in Section~\ref{ss:geom}.  It will be evident that quite a few

1438: cases need to be considered, some of which require new tools and

1439: some of which require only minor modifications.  Accordingly, the

1440: results will appear in several papers, currently under

1441: preparation.  In other words, a discussion of taxonomy will

1442: necessarily refer to results not yet published, and we will

1443: indicate to the best of our knowledge which ones are expected to

1444: be routine.

1445: 

1446: Given a point $\b{z} \in \sing$, we extend the definition of

1447: $\direc (\b{z})$ to mean the set of limits of $\direc (\b{y})$ as $\b{y}

1448: \to \b{z}$ along smooth points.  When $\b{z}$ is minimal, this is just

1449: the set of normals to support hyperplanes of $\logdom$ at the

1450: point $(\log |z_1| , \ldots , \log |z_d|)$, so this is consistent

1451: with the old definition.  As we will see shortly, $\direc (\b{z})$

1452: may be a $(d-1)$-dimensional subset of $\RP^{d-1}$ when $\b{z}$ is a

1453: critical point of $\sing$.

1454: 

1455: When $\denom$ has a repeated factor, the residue computation in

1456: equation~(\ref{eq:defres}) must be replaced by one involving the

1457: derivative.  The remainder of the computation proceeds without a

1458: hitch as before.  Details are given in \citeasnoun{pemantle-wilson;toral}.

1459: For the remainder of the taxonomy, we assume $\denom$ to

1460: be square-free.  Toral smooth points may be handled by methods

1461: exactly the same as strictly minimal points.  The inner integrand

1462: in~(\ref{eq:iter}) will in this case have its maximal modulus on a

1463: set of dimension larger than zero.  A modification of the

1464: necessary oscillating integral computation that works in this case

1465: is also given in \citeasnoun{pemantle-wilson;toral}.

1466: 

1467: If $\b{z} \in \sing$ is not smooth, all the first partials vanish.  The

1468: expansion of $\denom (\b{x})$ near $\b{z}$ is then a sum of terms of

1469: degrees 2 and higher.  We call $\b{z}$ a {\em homogeneous point} of

1470: degree $k$ if this expansion contains terms

1471: $(x_j - z_j)^k$ for each $j = 1 ,\ldots , d$, and contains no

1472: terms of total degree less than $k$.

1473: 

1474: \begin{lemma} \label{lem:homogeneous}

1475: If $\b{z}$ is a locally minimal point of $\sing$ with nonzero

1476: coordinates, and $F$ is meromorphic in a neighborhood of $\b{z}$

1477: then $\b{z}$ is homogeneous.

1478: \end{lemma}

1479: 

1480: \begin{proof} Passing to $F(z_1 x_1 , \ldots , z_d x_d)$ if

1481: necessary, we may assume $\b{z} = \one$.  Setting $x_j = 1$ for all

1482: but one index $j$, we cannot obtain the zero function (by

1483: minimality), and so some term in the expansion around $\one$ is a

1484: pure power of $(x_j - 1)$, and we denote the minimal degree such

1485: term by $c_j (x_j - 1)^{k_j}$.  If $\b{z}$ is not a homogeneous

1486: point, then there is some $j$ for which some monomial has total

1487: degree lower than $k_j$.  Assume without loss of generality that

1488: $j=d$.  The function $F(x , x  , \ldots , x , y)$ then has a

1489: minimal degree pure $y-1$ term $c_0 (y-1)^k$,

1490: $k := k_d$, and some term $c' (x-1)^a (y-1)^b$ with $a+b < k$.

1491: In other words, the Newton Polygon of $F(x, \ldots , x , y)$

1492: around $(1,1)$ has a support line passing through $(0,k)$ with

1493: slope $-p/q$ in lowest terms, and $p > q$.  It is well known that

1494: we may describe the solutions $y(x)$ of the equation

1495: $$F(1+x , \ldots , 1+x , 1 + y) = 0$$

1496: as follows.  Write

1497: $$H := (y-1)^k (c_0 + c_1 (y-1)^{-p} (x-1)^q + c_2(y-1)^{-2p} (x-1)^{2q}

1498:   + \cdots + c_s (y-1)^{-sp} (x-1)^{sq})$$

1499: for the polynomial collecting all the terms on this support line.

1500: Then for each $q^{th}$ root of unity, $\omega$, and each root

1501: $\lambda$ of $\sum c_{s-j} \lambda^j = 0$, there is a solution

1502: $y = \lambda^{1/p} x^{q/p} (\omega + o(1))$ as $x \to 0$.

1503: A proof may be found in \citeasnoun{brieskorn-knorrer}.

1504: 

1505: Varying $x$ over the set $|\pi - \arg (x)| \leq \pi / 4$, we see

1506: that the solutions $y(x)$ must sometimes be in this set as well.

1507: For those $x$, the points $(1+x , \ldots , 1+x , 1+y)$ will be in

1508: $\sing \cap \disk (\one) \setminus \torus (\one)$, violating

1509: minimality of $\one$.  By contradiction, we have shown that no

1510: monomial in the expansion around $\one$ has lower total degree

1511: than any pure power term, hence $\one$ is minimal.    \end{proof}

1512: 

1513: Continuing the taxonomy, suppose that $\b{z}$ is a homogeneous point

1514: of $\sing$ of degree $k \geq 2$.  We say that $\b{z}$ is a {\em

1515: multiple point} if $\sing$ is locally the union of $k$ analytic

1516: surfaces.  Algebraically, this means that the leading (order $k$)

1517: terms in the expansion of $\denom$ near $\b{z}$ factors into linear

1518: pieces.  If the homogeneous point $\b{z}$ is not a multiple point,

1519: we say it is a {\em cone point}.  When $d=2$ there are no cone

1520: points, since any homogeneous polynomial in 2 variables factors

1521: completely over $\CC$.

1522: 

1523: Our understanding of cone points is not yet complete, but an

1524: analysis involving cone points is underway in \citeasnoun{cohn-pemantle;fixation}.

1525: For multiple points, most of the story is given in

1526: \citeasnoun{pemantle-wilson;multiple}. In particular, the following theorem

1527: is proved there.

1528: %

1529: \begin{theorem}[\citeasnoun{pemantle-wilson;multiple}]

1530: \label{th:PW2000a}

1531: Let $\b{z}$ be an isolated, minimal, multiple point of $\sing$ with

1532: multiplicity $k$.  Let $S \subseteq \RP^{d-1}$ be the set of

1533: outward normals to support hyperplanes to $\logdom$ at the point

1534: $(\log |z_1| , \ldots , \log |z_d|)$.  Then there is an integer $p \geq 0$

1535: and a polynomial function $\phi : S \to \R$ such that the

1536: asymptotic expansion

1537: \begin{equation}\label{eq:PW2000a}

1538: a_\b{r} \sim \b{z}^{-\b{r}} \phi (\b{r}) \sum_j C_j (r_d)^{k-p/2-j/2}

1539: \end{equation}

1540: holds uniformly as $\b{r}$ varies over compact subsets of the

1541: interior of $S$.

1542: \noproof

1543: \end{theorem}

1544: 

1545: 

1546: The extension to toral multiple points is given in

1547: \citeasnoun{pemantle-wilson;toral}. If a multiple point is not

1548: isolated or toral, then the degree of multiplicity, $k$, must be

1549: less than the dimension,

1550: $d$.  This cannot happen of course when $d=2$, but does happen

1551: when $d \geq 3$.  The method for handling this case, toral or

1552: otherwise, is given in \citeasnoun{pemantle-wilson;toral}.  That paper

1553: will also contain some subcases of the isolated multiple point

1554: case, namely when the sheets of $\sing$ intersect non-transversely.

1555: 

1556: Having more or less completed the taxonomy, we now discuss when we

1557: can guarantee that our methods yield asymptotics in all

1558: directions.

1559: %

1560: \begin{theorem} \label{th:classify}

1561: Let $F = \numer / \denom = \sum a_{r,s} z^r w^s$ be the quotient

1562: of analytic functions $\numer , \denom : \CC^2 \to \CC$.  Suppose

1563: that the coefficients $a_{r,s}$ are all nonnegative, and that

1564: $F(z,0)$ and $F(0,w)$ are not entire.  Then for every direction

1565: $\alpha \in \RP^1$ there is a minimal $\b{z} \in \sing$ with

1566: $\alpha \in \direc (\b{z})$.

1567: \end{theorem}

1568: 

1569: \begin{proof} Let $(x,y)$ be any point on the boundary of

1570: $\logdom$.  For $u < e^x$ and $v < e^y$ the power series for $F$

1571: is convergent at $(u,v)$.  As $u \uparrow e^x$ and $v \uparrow

1572: e^y$ therefore, $F(u,v)$ is finite and increasing.  On the other

1573: hand, the power series for $F$ is not absolutely convergent on

1574: $\torus (e^x , e^y)$, since we know $F$ to have some singularity

1575: on this torus.  Hence $F(u,v) \uparrow \infty$ as $(u,v) \uparrow

1576: (e^x,e^y)$.  Since $F$ is meromorphic, it must have a pole at

1577: $(e^x , e^y)$, hence $(e^x , e^y) \in \sing$ and is a minimal point of $\sing$.

1578: As $(x,y)$ varies over the boundary of $\logdom$, we let $\gamma

1579: \subseteq \sing$ denote the curve traced out by this minimal point.

1580: 

1581: Pick any $\alpha \in \RP^1$.  The convex set $\logdom$ has

1582: horizontal and vertical support hyperplanes (by non-entirety of

1583: $F(z,0)$ and $F(0,w)$), and therefore has a support hyperplane

1584: normal to $\alpha$; let $(x,y)$ be a point of intersection of this

1585: support plane with $\logdom$. We have just seen that $\b{z} (\alpha)

1586: := (e^x , e^y)$ is a minimal point of $\sing$.  If $\b{z}$ is a smooth

1587: point of $\sing$ then $\alpha \in \direc (\b{z})$: either $\b{z}$ is

1588: finitely minimal, in which case Theorem~\ref{th:2d} applies, or it

1589: is toral, in which case the toral version of this theorem from

1590: \citeasnoun{pemantle-wilson;toral} applies.

1591: 

1592: Assume now that $\b{z}$ is not a smooth point.  By

1593: Lemma~\ref{lem:homogeneous}, $\b{z}$ is a homogeneous point, and

1594: since $d=2$, $\b{z}$ is a multiple point.  Theorem~\ref{th:PW2000a}

1595: then shows that $\alpha \in \direc (\b{z})$ in this case as well.

1596: This finishes the proof. 

1597: \end{proof}

1598: 

1599: \section{Further details and open questions} \label{ss:open}

1600: 

1601: The theorems in this and subsequent papers give estimates that are

1602: uniform away from the boundary of the domain in which they are

1603: valid.  In order for all of these to be patched together so as to

1604: give estimates valid now matter how $\b{r} \to \infty$, one must

1605: determine the bandwidth around the boundary for which the boundary

1606: estimates on either side hold.  For instance, suppose $(z,w)$ is a

1607: multiple point of degree 2 and that $\direc (z,w)$ is the set of

1608: slopes between $1/2$ and 2.  It appears that the asymptotic

1609: estimate in \citeasnoun{pemantle-wilson;multiple} holding near the line $\{

1610: s = 2r \}$ can be written so it is valid out to $s = 2r + c

1611: \sqrt{r}$.  If the estimate for the region $s/r > 2 + \ee$ can be

1612: widened so it holds to $s = 2r + c \sqrt{r}$ and a description

1613: given that is valid in the regime $(s-2r)/\sqrt{r} \to c$, then

1614: the estimates will patch together completely.

1615: 

1616: Another natural question is the universality of the method when

1617: the coefficients have mixed signs.  We conjecture that

1618: Theorem~\ref{th:classify} still holds, in the sense that for every

1619: direction there is point $\b{z} \in \sing$ for which integration near

1620: $\b{z}$ yields correct asymptotics.  What we know is that $\b{z}$ may

1621: no longer be minimal.  For example, if $\numer = 1$ and

1622: $$\denom = (1 - (2/3) w - (1/3) z) (1 + (1/3) w - (2/3) z )$$

1623: then the point $(3/2 , 3/4)$ is not minimal but yields asymptotics

1624: in the diagonal direction; one sees this by integrating along a

1625: deformed torus rather than along $\torus (3/2 , 3/4)$.  In fact we

1626: conjecture that such a deformation always exists, but the topology

1627: seems not transparent enough to yield an easy proof.

1628: 

1629: The class of algebraic functions is in some ways almost as nice as

1630: the set of rational functions, and nicer than the meromorphic

1631: functions.  For one thing, an algebraic function is determined by

1632: a finite amount of data, and may thus easily be input into a

1633: symbolic math package.

1634: \citeasnoun{gao-richmond;limit-multivariate4} give an analysis of

1635: algebraic and logarithmic singularities, but sometimes the

1636: relevant singularities for algebraic functions are poles.  For

1637: example, in Larsen and Lyons' analysis

1638: \cite{larsen-lyons;coalescing} of merge times for coalescing

1639: particles, they find an algebraic function of the form

1640: $$F(z,w) = {\chi (z,w) \over w - 1 - \sqrt{1-z}}$$

1641: with $\chi$ analytic.  The branch of the square root is chosen so

1642: that at the origin the denominator is 2, not 0.  There is a

1643: branchline at $z=1$, but for all directions in $\RP^1$, there is a

1644: smooth pole on the curve $w = 1 + \sqrt{1-z}$ yielding asymptotics

1645: in the desired direction.  It is natural to ask when this will

1646: happen, and how one can tell effectively.  Some questions of

1647: effectiveness are addressed in \citeasnoun{pemantle-wilson;multiple} and

1648: \citeasnoun{pemantle-wilson;toral},

1649: but there is probably substantial room for improvements on an

1650: algorithmic level.

1651: 

1652: \noindent{\bf Acknowledgements:} We thank Steve

1653: Wainger for his generous help, including pointing out

1654: equation~(\ref{eq:X}).  Thanks also to Andreas Seeger for his help

1655: with the material in Section~\ref{ss:oscillate}.  We are indebted

1656: to Manuel Lladser for many helpful comments on an earlier draft.

1657: 

1658: \bibliographystyle{agsm}

1659: \bibliography{asymp}

1660: \end{document}

1661: