1: \def\versiondate{23 Jan.\ 2003}

2: \input math.macros

3: %\input Ref.macros

4: \input lrlEPSfig.macros

5: 

6: %\proofmodetrue

7: %\leftsectionheadtrue

8: \checkdefinedreferencetrue

9: %\continuousnumberingtrue

10: \continuousfigurenumberingtrue

11: \theoremcountingtrue

12: \sectionnumberstrue

13: %\figuresectionnumberstrue

14: \forwardreferencetrue

15: %\lefteqnumberstrue

16: \citationgenerationtrue

17: \nobracketcittrue

18: \tocgenerationtrue

19: \hyperstrue

20: \initialeqmacro

21: 

22: \bibsty{myapalike}

23: %\input\jobname.lbl 

24: 

25: \def\T{{\Bbb T}}

26: \def\Td{{\T^d}}

27: \def\dbar{{\overline{d}}}

28: \def\Pfp{{(\P^f)^+}}

29: \def\Pfm{{(\P^f)^-}}

30: \def\calA{{\cal A}}

31: \def\calT{{\cal T}}

32: \def\F{{\cal F}}

33: \def\U{{\cal U}}

34: \def\C{{\Bbb C}}

35: \def\hE{\ell^2(E)}

36: \def\Hs{{\ell^2(E)}}

37: \def\hD{\ell^2({\Z}^d)}

38: \def\id{{\rm id}}  %% The identity operator.

39: \def\ev#1{{\cal #1}}

40: \def\A{{\ev A}}

41: \def\ip#1{(\changecomma #1)}

42: \def\bigip#1{\bigl(\bigchangecomma #1\bigr)}

43: \def\biggip#1{\biggl(\bigchangecomma #1\biggr)}

44: \def\leftip#1{\left(\leftchangecomma #1\right)}  

45: \def\changecomma#1,{#1,\,}

46: \def\bigchangecomma#1,{#1,\;}

47: \def\leftchangecomma#1,{#1,\ }

48: \def\Var{{\rm Var}}

49: \def\bfz{{\bf 0}}

50: \def\bfo{{\bf 1}}

51: \def\GM{{\ss GM}}  %% geometric mean

52: \def\mult#1{M_{\times #1}} %% multiplication of the variable

53: \def\strle{\preccurlyeq_{\rm s}}   %% strongly dominated by

54: %\def\frac#1#2{{#1 \over #2}}  % to work with Mathematica TeXForm

55: %\def\imag{i}  % for TeXForm

56: \def\fav{\widehat f(0)}  % the average of $f$

57: \def\past{{\ss Past}}

58: \def\cat{{\rm G}}  % Catalan's constant

59: \def\others{{\cal Z}}  % the $\sigma$-field generated by non-zero sites

60: \def\HM{{\ss HM}}  % harmonic mean

61: \def\fullle{\preccurlyeq_{\rm f}}   %% fully dominated by

62: \def\Zd{{\Z^d}}

63: \def\ee{{\bf e}}  % exponential

64: \def\frac#1#2{{#1 \over #2}}

65: \def\nonfrac#1#2{{#1 / #2}}

66: \def\hl{{\ss HL}}  % Helson-Lowdenslager space

67: \def\supp{{\rm supp}\,}

68: \def\Cov{{\rm Cov}}

69: \def\ord{\Pi}  % order

70: 

71: \def\BLPSusf{Benjamini, Lyons, Peres, and Schramm (2000)%

72: \def\BLPSusf{BLPS (2000)}}

73: 

74: 

75: 

76: \def\firstheader{\eightpoint\ss\underbar{\raise2pt\line 

77:     {

78:     To appear in {\it Duke Math. J.}

79:     \hfil Version of \versiondate}}}

80: 

81: \beginniceheadline

82: 

83: %\ifproofmode \relax \else\head{} {Version of \versiondate}\fi 

84: \vglue20pt

85: 

86: \title{Stationary Determinantal Processes:}

87: \title{Phase Multiplicity, Bernoullicity,}

88: \title{Entropy, and Domination}

89: 

90: \author{Russell Lyons and Jeffrey E. Steif}

91: 

92: 

93: 

94: 

95: 

96: \abstract{%

97: We study a class of stationary processes indexed by $\Z^d$ that

98: are defined via minors of $d$-dimensional (multilevel) Toeplitz matrices.

99: We obtain necessary and sufficient conditions for

100: phase multiplicity (the existence of a phase transition)

101: analogous to that which occurs in 

102: statistical mechanics.

103: Phase uniqueness is equivalent to the presence of a strong

104: $K$ property, a particular strengthening of the usual $K$ 

105: (Kolmogorov) property. 

106: We show that all of these processes are

107: Bernoulli shifts (isomorphic to i.i.d.\ processes in the sense of ergodic

108: theory).

109: We obtain estimates of their entropies and

110: we relate these processes via stochastic domination to product measures.

111: }

112: 

113: 

114: \bottomII{Primary 

115: 82B26, %phase transitions (general)

116: 28D05, %measure-preserving transformations

117: 60G10. %Stationary processes

118: Secondary

119: 82B20, %lattice systems (Ising, dimer, Potts, etc.) and systems on graphs

120: 37A05, %measure-preserving transformations

121: 37A25, %ergodicity, mixing, rates of mixing

122: 37A60, %dynamical systems in statistical mechanics

123: 60G25, %Prediction theory

124: 60G60, %Random fields

125: 60B15. %Probability measures on groups, Fourier transforms, factorization

126: }

127: {Random field, Kolmogorov mixing, Toeplitz determinants, negative

128: association, fermionic lattice model, stochastic domination, entropy,

129: Szeg\H{o} infimum, Fourier coefficients, insertion tolerance, deletion

130: tolerance, geometric mean.}

131: {Research partially supported by 

132: NSF grants DMS-9802663, DMS-0103897 (Lyons), and

133: DMS-0103841 (Steif), the Swedish Natural Science Research Council (Steif)

134: and the Erwin Schr\"odinger Institute for Mathematical

135: Physics, Vienna (Lyons and Steif).}

136: 

137: 

138: \articletoc

139: 

140: 

141: 

142: 

143: \bsection {Introduction}{s.intro}

144: 

145: Determinantal probability measures and point processes arise in numerous

146: settings, such as

147: mathematical physics (where they are called fermionic point processes),

148: random matrix theory, representation theory, and certain other areas

149: of probability theory. 

150: See \ref b.Sosh/ for a survey and \ref b.L:det/ for additional developments

151: in the discrete case.

152: We present here a detailed analysis of the discrete

153: stationary case.

154: After this paper was first written, we learned of independent but slightly

155: prior work of \ref b.ShiTak:II/, announced in \ref b.ShiTak:ann/.

156: We discuss the (small) overlap between their work and ours in the appropriate

157: places below.

158: See also \ref b.ShiTak:I/ and \ref b.ShiYoo:glauber/ for related

159: contemporaneous work.

160: 

161: Stationary determinantal processes are interesting from several viewpoints.

162: First, they have interesting relations with the 

163: theory of Toeplitz determinants.

164: As in that theory, the {\bf geometric mean} of a nonnegative function $f$,

165: defined as 

166: $$

167: \GM(f) := \exp \int \log f

168: \,,

169: $$

170: will play an important role in some of our results.

171: (In fact, the arithmetic mean and the harmonic mean of $f$ will also

172: characterize certain properties of our processes.)

173: Second, such processes arise in certain combinatorial models, such as uniform

174: spanning trees and dimer models.

175: Third, these systems have a rich infinite-dimensional parameter space,

176: consisting, in the case of a $\Z^d$ action, of all measurable functions $f$

177: from the $d$-dimensional torus to $[0, 1]$.

178: We illustrate the variety of possible behaviors with examples throughout

179: the paper.

180: Fourth, they have the unusual property of negative association.

181: Though unusual, negative association occurs in various places and a fair

182: amount is known about it

183: (see \ref b.Joag/, \ref b.Pemantle:negass/, 

184: \ref b.Newman:asympt/, \ref b.ShaoSu:LIL/, \ref b.Shao:comparison/,

185: \ref b.ZhangWen:weak/, \ref b.Zhang:Strassen/, and the references therein,

186: for example). We offer a whole new class of examples of negatively

187: associated stationary processes.

188: As such, determinantal processes provide an easy way to construct examples of

189: many kinds of behavior that might otherwise be difficult to construct, such as

190: negatively associated (stationary) processes with slow decay of correlations,

191: or with the even sites independent of the odd sites (in one dimension, say),

192: or with the property of being finitely dependent.

193: Fifth, all our processes are Bernoulli shifts, i.e., isomorphic to i.i.d.\

194: processes. This may be surprising in view of the fact that only measurability,

195: rather than smoothness, of the parameter $f$ is required.

196: Sixth, in one dimension,

197: some of the processes are strong $K$, while others are not.

198: Namely, strong $K$ is equivalent to $f(\bfo-f)$ having a positive

199: geometric mean.

200: Similarly, we characterize exactly, in all dimensions, which $f$ give strong

201: {\it full\/} $K$ systems. It turns out that not only the rate at which $f$

202: approaches 0 or 1 matters, but also where. For example, in two dimensions,

203: if $f$ is real analytic, then the system is strong full $K$ iff the

204: (possibly empty) sets $f^{-1}(0)$ and $f^{-1}(1)$ belong to nontrivial

205: algebraic varieties. The strong full $K$ property is 

206: analogous to phase uniqueness

207: in statistical physics, as

208: we explain in \ref s.sk/.

209: 

210: We now state our results somewhat more precisely and present several

211: examples.

212: Let $f : \Td \to [0, 1]$ be a Lebesgue-measurable function on the

213: $d$-dimensional torus $\Td := \R^d/\Z^d$.

214: Define a $\Z^d$-invariant probability measure $\P^f$ on the Borel sets of $ \{

215: 0, 1 \}^{\Z^d}$ by defining the probability of the cylinder sets 

216: \begineqalno

217: \P^f[\eta(e_1)=1,\dots, \eta(e_k)=1] 

218: &:= 

219: \P^f[\{ \eta \in \{ 0, 1 \}^{\Z^d} \st \eta(e_1)=1,\dots, \eta(e_k)=1\}]

220: \cr&:= 

221: \det [\widehat{f}(e_j-e_i)]_{1\le i, j \le k}

222: \cr

223: \endeqalno

224: for all $e_1,\ldots,e_k\in \Z^d$, where $\widehat f$ denotes the Fourier

225: coefficients of $f$.

226: We shall prove in \ref s.back/ that this does indeed define a probability

227: measure.

228: Note that when $d=1$ and when $e_1,\ldots,e_k$

229: are chosen to be $k$ consecutive integers, the

230: right-hand side above is the usual $k \times k$ Toeplitz determinant of $f$

231: denoted $D_{k-1}(f)$.

232: In particular, we have $\P^f\big[\eta(e) = 1\big] = \widehat f(\bfz)

233: = \int_{\Td} f$ for every $e \in \Z^d$.

234: 

235: \procl x.Bernoulli

236: As a simple example, if $f \equiv p$, then $\P^f$ is i.i.d.\ Bernoulli($p$)

237: measure.

238: \endprocl

239: 

240: 

241: \procl x.ustH

242: For a more interesting example, consider

243: $$

244: f(x, y) := {\sin^2 \pi x \over \sin^2 \pi x + \sin^2 \pi y}

245: \,.

246: \label e.ustH

247: $$

248: A portion of a sample of a configuration from $\P^f$ is shown in \ref

249: f.window-ustH/, where a square with upper left corner $(i, j)$ is colored

250: black iff $\eta(i, j) = 1$.

251: These correspond to the horizontal edges of a uniform spanning tree in the

252: square lattice. That is, if $T$ is a spanning tree of $\Z^2$, then $\eta(i,

253: j)$ is the indicator that the edge from $(i, j)$ to $(i+1, j)$ belongs to

254: $T$. The portion of the spanning tree from which \ref f.window-ustH/ was

255: constructed is shown in \ref f.window-ust/.

256: When $T$ is chosen ``uniformly" (see \ref b.Pemantle:ust/, \ref

257: b.Lyons:bird/, or \ref b.BLPSust/ for definitions and information on

258: this), then $\eta$ has the law $\P^f$. This follows from the Transfer

259: Current Theorem of \ref b.BurPem/ and the representation of the Green

260: function as an integral; see \ref b.Lyons:book/ for more details.

261: Similarly, the edges of the uniform spanning forest (it is a tree only for $d

262: \le 4$, as shown by \ref b.Pemantle:ust/) parallel to the $x_1$-axis in $d$

263: dimensions correspond to the

264: function 

265: $$

266: f(x_1, x_2, \ldots, x_d) := {\sin^2 \pi x_1 \over \sum_{j=1}^d \sin^2 \pi x_j}

267: \,.

268: \label e.ustH-d

269: $$

270: We remark that

271: the uniform spanning tree is the so-called random-cluster model when 

272: one takes the limit $q \downarrow 0$,

273: then $p \downarrow 0$, and finally the thermodynamic limit, the latter

274: shown to exist by \ref b.Pemantle:ust/.

275: \endprocl

276: 

277: \twoefiginslabel window-ustH {A sample from $\P^f$ of \ref e.ustH/.} x3 

278: window-ust {A uniform spanning tree.} y3

279: 

280: \procl x.ustHx

281: Let 

282: $$

283: g(x) :=

284: {\sin \pi x \over \sqrt{1+\sin^2 \pi x}}

285: \,.

286: $$

287: Then the edges of the uniform spanning tree in the plane that

288: lie on the $x$-axis have the law $\P^g$, as shown in \ref x.ustH_x-dom/ below.

289: \endprocl

290: 

291: 

292: \procl x.USTzz

293: Let 

294: $$

295: f(x) := {1 \over 2} + {|\sin 2\pi x| - 1 \over 2 \cos 2\pi x}

296: \,.

297: $$

298: An elementary calculation shows that 

299: $$

300: \widehat f(k) = \cases{1/2  &if $k=0$,\cr

301:                        0    &if $k \ne 0$ is even,\cr

302: \displaystyle

303: (-1)^{(k-1)/2}\left(-{1 \over 2} + {2 \over \pi} 

304:       \sum_{j=0}^{(k-1)/2} {(-1)^j \over 2j+1}\right) - {1 \over \pi k}

305:                             &if $k$ is odd.\cr}

306: $$

307: We shall show in \ref x.USTzz-dom/ that the measure $\P^f$ arises as follows.

308: Given a spanning tree $T$ of the square lattice, let $\eta(n)$ be the

309: indicator that $e_n \in T$, where $e_n$ is the edge 

310: $$

311: e_n := \cases{[(n/2, n/2), (n/2 + 1, n/2)]  &if $n$ is even,\cr

312:               [((n+1)/2, (n-1)/2), ((n+1)/2, (n+1)/2)]  &if $n$ is odd.\cr}

313: $$

314: The collection of edges $\{e_n \st n \in \Z \}$ is a zig-zag path in the

315: plane. The law of $\eta$ is $\P^f$ when $T$ is chosen as a uniform spanning

316: tree.

317: (Although it is not hard to see that the law of $\eta$ is $\Z$-invariant,

318: by using planar duality, and although the law of $\eta$ must be a

319: determinantal probability measure because the law of $T$ is,

320: it is not apparent {\it a priori\/} that the law of $\eta$ has the

321: form $\P^f$ for some $f$.)

322: \endprocl

323: 

324: \procl x.lozenge

325: Fix a horizontal edge of the hexagonal lattice (also known as the

326: honeycomb lattice) and index all its vertical translates by $\Z$.

327: If one considers the standard measure of maximal entropy on perfect

328: matchings of the hexagonal lattice, also called the dimer model

329: and equivalent to lozenge tilings of the plane, and

330: looks only at the edges indexed as above by $\Z$, then the law is $\P^f$

331: for $f := \I{[1/3, 2/3]}$, as shown by \ref b.Kenyon:local/.

332: \endprocl

333: 

334: \procl x.domino It is interesting that the function

335: $f := \I{[0, 1/2]}$ for $d = 1$ also arises from a combinatorial model.

336: In this case,

337: $$

338: \widehat f(n) = \cases{1/2 &if $n=0$,\cr

339:                   0 &if $n \ne 0$ is even,\cr

340:                   1/(\pi i n) &if $n$ is odd.\cr}

341: $$

342: The measure $\P^f$ is the zig-zag process of \ref b.Joh:noninter/

343: derived from uniform domino tilings in the plane.

344: For the definition of ``uniform" in this case, see \ref b.BurPem/.

345: A picture of a portion of such a tiling is shown in \ref f.window-tiling/.

346: Consider the squares on a diagonal from upper left to lower right.

347: The domino covering any such square also covers a second square.

348: If this second square is above the diagonal, then we color the original

349: square black, as shown in \ref f.window-diagprocess/.

350: \ref b.Joh:noninter/ showed that the law of this process is $\P^f$ when the

351: diagonal squares are indexed by $\Z$ in the natural way.

352: More generally, the processes $\P^f$ for $f$ the indicator of any arc of $\T$

353: are used by \ref b.BOO:plancherel/, Theorem 3, to describe the typical shape

354: of Young diagrams.

355: \endprocl

356: 

357: \twoefiginslabel window-tiling {A uniform domino tiling.} x3 

358: window-diagprocess {A sample from $\P^f$ of \ref x.domino/.} y3

359: 

360: 

361: \procl x.renewal Let $0 < a < 1$ and $d = 1$.

362: If $f(x) := (1 - a)^2/|e^{2\pi i x} - a|^2$, then $\P^f$ is a renewal

363: process (\ref B.Sosh/).

364: The number of 0s between successive 1s has the same distribution as

365: the number of tails until 2 heads appear for a coin that has probability $a$

366: of coming up tails.

367: More explicitly, for $n\ge 1$,

368: $$

369: \P^f[\eta(1) = \cdots = \eta(n-1) = 0, \eta(n) = 1 \mid \eta(0) = 1]

370: =

371: n (1-a)^2 a^{n-1}

372: \,.

373: \label e.renewal-distribution

374: $$

375: Since 

376: $$

377: f(x) = {1-a \over 1+a} \left( {a e^{2\pi i x} \over 1 - a e^{2\pi i x}} + {1

378: \over 1 - a e^{-2\pi i x}} \right)

379: \,,

380: $$

381: expansion in a geometric series shows that

382: $$

383: \widehat f(k) = {1-a \over 1+a} a^{|k|}

384: \,.

385: $$

386: We prove that $\P^f$ is indeed this explicit renewal process after we prove

387: \ref p.regenerate/, in which we extend this example to other regenerative

388: processes.

389: \endprocl

390: 

391: 

392: \procl x.perconf

393: If $0 < p < 1$ and $f : \Td \to [0, 1]$ is measurable, then a fair sample of

394: $\P^{pf}$ can be obtained from a fair sample of $\P^f$ simply by independently

395: changing each 1 to a 0 with probability $1-p$.

396: \endprocl

397: 

398: 

399: For general systems, note that the covariance of

400: $\eta(\bfz)$ and $\eta(k)$ for $k \in \Z^d$ is $-|\widehat f(k)|^2$. This is

401: summable in $k$ since $f \in L^2(\Td)$, but that is essentially the most one

402: can say for its rate of decay.

403: That is, given any $\Seq{a_k} \in \ell^1(\Z^d)$, there is some

404: (even continuous) $f : \Td \to [0, 1]$ and some constant $c > 0$ such that

405: $|\widehat f(k)|^2 \ge c |a_k|$ for all $k \in \Z^d$, as shown by \ref

406: b.LKK/.

407: Observe also that as is the case for Gaussian processes, the processes

408: studied here have the property that if the random variables are uncorrelated, then

409: they are mutually independent.

410: 

411: It is shown (in a much more general context) by \ref b.L:det/ that these

412: measures, as well as these measures conditioned on the values of $\eta$

413: restricted to any finite subset of $\Z^d$,

414: have the following negative association property:

415: If $A$ and $B$ are increasing events that are measurable with respect to the

416: values of $\eta$ on disjoint subsets of $\Z^d$, then $A$ and $B$ are

417: negatively correlated.

418: 

419: In the order presented in this paper, our principal results are the following.

420: \beginbullets

421: 

422: For all $f$, the process $\P^f$ is a Bernoulli shift. % (\ref t.Bern/).

423: This was shown by \ref b.ShiTak:II/ for those $f$ such that $\sum_{n \ge 1}

424: n |\widehat f(n)|^2 <\infty$ by showing that those $\P^f$ are weak Bernoulli.

425: 

426: For all $f$, the process $\P^f$ stochastically dominates product measure

427: $\P^{\GM(f)}$ and is stochastically dominated by product measure

428: $\P^{1-\GM(\bfo-f)}$, and these bounds are optimal.

429: This is rather unexpected for the process of \ref x.ustH/ related to the

430: uniform spanning tree; explicit calculations are given below in \ref

431: x.ustH-dom/ for this particular process.

432: We give similar optimal bounds for full domination (uniform insertion and

433: deletion tolerance) in terms of harmonic means.

434: 

435: We present methods to estimate the entropy of $\P^f$.

436: For example, we show that for the function $g$ in \ref x.ustHx/,

437: the entropy of the system $\P^g$ lies in the interval $[0.69005, 0.69013]$;

438: see \ref x.ustH-x-ent/.

439: 

440: The process $\P^f$ is strong full $K$ iff there is a nonzero

441: trigonometric polynomial $T$ such that 

442: ${|T|^2 \over f (\bfo-f)} \in L^1(\Td)$.

443: This is equivalent to phase uniqueness in the sense that

444: no conditioning at infinity can change the measure.

445: 

446: In one dimension, $\P^f$ is strong $K$ iff $f(\bfo-f)$ has a positive

447: geometric mean.

448: This is equivalent to phase uniqueness when conditioning on one

449: side only. Higher-dimensional versions of this will also be obtained.

450: \endbullets

451: 

452: We shall give full definitions as they become needed. 

453: Some general background on determinantal probability measures is presented

454: in \ref s.back/, where we also exhibit a key representation of certain

455: conditional probabilities as Szeg\H{o} infima.

456: The property of being a Bernoulli shift 

457: is proved in \ref s.bern/, while the auxiliary result

458: that all $\P^f$ have full support (except in two degenerate cases) is shown in

459: \ref s.fs/.

460: Properties concerning stochastic domination are proved in \ref

461: s.dom/. These are used to estimate entropy in \ref s.ent/.

462: More sophisticated methods

463: of estimating entropy are also developed and illustrated in \ref s.ent/.

464: The main result about phase multiplicity is proved in \ref s.sk/, while the

465: one-sided case for $d=1$ and its higher-dimensional generalizations are

466: treated in \ref s.1side/.

467: Finally, we end with some open questions in \ref s.open/.

468: 

469: The above definitions can be generalized to any countable abelian group with

470: the discrete topology. 

471: For example, if we use $\Z \times \Z_2$, 

472: we obtain nontrivial joinings of the above systems with themselves. That

473: is, suppose $f : \T \to [0, 1]$ and $h : \T \to [0, 1]$ are such that $f \pm

474: h : \T \to [0, 1]$. Then the function $f_h : \T \times \{ -1, 1 \}

475: \to [0, 1]$ defined by $(x, \epsilon) \mapsto f(x) + \epsilon h(x)$

476: gives a system that, when restricted to each copy of $\Z$, is just $\P^f$, but

477: has correlations between the two copies that are given by $h$

478: (the case $h = \bfz$ gives the independent joining). 

479: We can obtain slow decay of correlations between the two copies and

480: negative associations in the joining itself;

481: this is perhaps something that is not easy to construct directly.

482: 

483: 

484: \bsection {Background}{s.back}

485: 

486: We first quickly review the probability measures studied in 

487: \ref b.L:det/;

488: see that paper for complete details.

489: 

490: Let $E$ be a finite or countable set and consider the complex

491: Hilbert space $\hE$. 

492: Given any closed subspace $H \subseteq \ell^2(E)$,

493: let $P_H$ denote the orthogonal projection onto $H$.

494: There is a unique probability measure $\P^H$ on $2^E := \{ 0, 1 \}^E$ defined

495: by 

496: $$

497: \P^H[\eta(e_1) = 1, \ldots, \eta(e_k) = 1]

498: =

499: \det [(P_H e_i, e_j)]_{1\le i,j\le k}

500: \label e.(3)

501: $$

502: for all $k \ge 1$ and any set of distinct $e_1,\ldots,e_k\in E$; see, e.g.,

503: \ref b.L:det/ or \ref b.DVJ/, Exercises 5.4.7--5.4.8.

504: On the right-hand side, we are identifying each $e \in E$ with the element

505: of $\hE$ that is 1 in coordinate $e$ and 0 elsewhere. 

506: In case $H$ is finite-dimensional, then $\P^H$ is concentrated on subsets

507: of $E$ of cardinality equal to the dimension of $H$.

508: 

509: 

510: More generally,

511: let $Q$ be a positive contraction, meaning that $Q$ is a self-adjoint

512: operator on $\Hs$ such that for all $u \in \Hs$, we have $0 \le (Q u, u) \le

513: (u, u)$.

514: There is a unique probability

515: measure $\P^Q$ such that 

516: $$

517: \P^Q[\eta(e_1) = 1, \ldots, \eta(e_k) = 1]

518: = \det [(Q {e_i}, {e_j})]_{i, j \le

519: k}

520: \label e.Qcorrel

521: $$

522: for all $k \ge 1$ and distinct $e_1, \ldots, e_k \in E$.

523: When $Q$ is the orthogonal projection onto a closed

524: subspace $H$, then $\P^Q = \P^H$.

525: In fact, properties of $\P^Q$ can be deduced from the special case of

526: orthogonal projections. Since this will be useful for our analysis, we

527: review this reduction procedure.

528: 

529: Note first that uniqueness follows from the fact that \ref e.Qcorrel/

530: determines all finite-dimen\-sion\-al marginals via the inclusion-exclusion

531: theorem. Indeed, we have the following formula for any disjoint pair of

532: finite sets $A, B \subseteq E$ (see, e.g., \ref b.L:det/):

533: $$

534: \P^Q[\eta \restrict A \equiv 1,\, \eta \restrict B \equiv 0]

535: =

536: \det \left[\big(\I B(e)e + (-1)^{\I B(e)} Q e, e'\big)\right]_{e, e' \in A

537: \cup B}

538: \,.

539: \label e.incexc

540: $$

541: To show existence, let $P_H$ be any orthogonal projection that is

542: a dilation of $Q$, i.e., $H$ is

543: a closed subspace of $\ell^2(E')$ for some $E' \supseteq E$ and for all $u

544: \in \Hs$, we have $Qu = P_{\Hs} P_H u$, where we regard $\ell^2(E') = \Hs

545: \oplus \ell^2(E' \setminus E)$. (In this case, $Q$ is also called the

546: compression of $P_H$ to $\Hs$.)

547: The existence of a dilation is standard and is easily

548: constructed; see, e.g., \ref b.L:det/.

549: Having chosen a dilation, we simply define $\P^Q$ as the law of $\eta$

550: restricted to $E$ when $\eta$ has the law $\P^H$.

551: Then \ref e.Qcorrel/ is a special case of \ref e.(3)/.

552: 

553: A probability measure $\P$ on $2^E$ is said to have {\bf negative

554: associations} if for all pairs

555: $A$ and $B$ of increasing events that are measurable with respect to the

556: values of $\eta$ on disjoint subsets of $E$, we have that $A$ and $B$ are

557: negatively correlated with respect to $\P$.

558: The following {\bf conditional negative association} (CNA)

559: property is proved in \ref b.L:det/ and a consequence of this

560: (see \ref p.JNRD/) will be used frequently here.

561: 

562: \procl t.negass

563: If $Q$ is any positive contraction on $\ell^2(E)$, $A$ is a

564: finite subset of $E$, and $\eta_0 \in 2^A$,

565: then $\P^Q[\;\cdot \mid \eta \restrict A = \eta_0]$ has

566: negative associations.

567: \endprocl

568: 

569: We now assume that $E=\Z^d$. Then the group structure of $\Z^d$ allows

570: $\Z^d$ to act naturally on $\ell^2(\Z^d)$ and on $2^{\Z^d}$.

571: The proof of the following lemma is straightforward and 

572: therefore skipped.

573: 

574: \procl l.pm

575: If $Q$ is a $\Z^d$-invariant positive contraction on

576: $\ell^2(\Z^d)$, then $\P^Q$ is also $\Z^d$-invariant.

577: \endprocl

578: 

579: As is well known, there exists a complex Hilbert-space isomorphism

580: between $L^2(\T^d, \lambda_d)$ and $\ell^2(\Z^d)$ where

581: $\T^d$ is the $d$-dimensional torus $\R^d/\Z^d$ and

582: $\lambda_d$ is unit Lebesgue measure on $\T^d$. 

583: This isomorphism is given by the Fourier transform

584: $f\mapsto \widehat f$, where for $f \in L^2(\T^d, \lambda_d)$, we have

585: $\widehat{f}(k)=\int_{\T^d} f(x) e^{-2\pi i k\cdot x} \, d\lambda_d(x)$ 

586: for $k\in\Z^d$.

587: If $\ee_k$ denotes the function $x \mapsto e^{2\pi i k \cdot x}$, then the

588: isomorphism takes the set $ \{ \ee_k \st k \in \Zd \}$ to the standard basis

589: for $\ell^2(\Zd)$.

590: From now on, we shall abbreviate $L^2(\T^d, \lambda_d)$ by $L^2(\T^d)$.

591: 

592: The following is well known.

593: 

594: \procl t.rudin 

595: \beginitems

596: \itemrm{(i)} Let $A\subseteq\Td$ be measurable and consider the operator 

597: $T_A:L^2(\T^d)\to L^2(\T^d)$ given by

598: $$

599: T_A(g)=g \I A\,.

600: $$

601: Then these projections (as $A$ varies over the measurable subsets of $\Td$) 

602: correspond (via the Fourier

603: isomorphism) to the $\Z^d$-invariant projections on $\ell^2(\Z^d)$.

604: \itemrm{(ii)} More generally, let $f:\Td\to [0,1]$ be measurable and

605: consider the operator $M_f:L^2(\T^d)\to L^2(\T^d)$ given by

606: $$

607: M_f(g)=fg\,.

608: $$

609: Then these positive contractions (as $f$ varies

610: over the measurable functions from $\T^d$ to $[0,1]$) correspond (via the

611: Fourier

612: isomorphism) to the $\Z^d$-invariant positive contractions on $\ell^2(\Z^d)$.

613: More specifically, $M_f$ corresponds to convolution with $\widehat{f}$.

614: \enditems

615: \endprocl

616: 

617: As in the above theorem, an $f:\Td\to [0,1]$ yields a

618: $\Z^d$-invariant positive contraction $Q_f$ on

619: $\ell^2(\Z^d)$, which in turn yields a translation-invariant

620: probability measure $\P^{Q_f}$ on $2^{\Z^d}$ that we denote more simply

621: by $\P^f$.

622: 

623: \procl l.fourier

624: Given $f:\Td\to [0,1]$ measurable and $e_1,\ldots,e_k\in \Z^d$,

625: $$

626: \P^f[\eta(e_1)=1,\dots, \eta(e_k)=1] = 

627: \det [\widehat{f}(e_j-e_i)]_{1\le i, j \le k}

628: \,.

629: $$

630: \endprocl

631: 

632: \proof 

633: By definition, the left-hand side is

634: $\det [(Q_f e_i,e_j)]_{1\le i, j \le k}$. 

635: By \ref t.rudin/(ii), 

636: $$

637: \ip{Q_f e_i,e_j}= \bigip{M_f \ee_{e_i}, \ee_{e_j}}

638: = \widehat{f}(e_j-e_i)\,.

639: \Qed

640: $$

641: 

642: 

643: \procl r.toeplitz

644: \ref l.fourier/ says that for $d=1$,

645: the probability of having 1s on some finite collection of elements of

646: $\Z$ is a particular minor of the Toeplitz matrix associated to $f$.

647: \endprocl

648: 

649: Equation \ref e.incexc/ shows a symmetry of $\P^f$ and

650: $\P^{\bfo-f}$, namely, if $\eta$ has the distribution $\P^f$, then $\bfo - \eta$

651: has the distribution $\P^{\bfo-f}$.

652: \ref b.ShiTak:II/ prove the existence of $\P^f$ by a different method and

653: also note this symmetry.

654: 

655: Although the last lemma gives us a formula for $\P^f$ directly in terms of

656: $f$ without reference to any projections, it is still useful 

657: to know a specific projection of which $M_f$ is a compression. 

658: Let 

659: $f:\Td\to [0,1]$ be measurable. 

660: Identifying $\T^{d+1}$ with $\Td \times \CO{0, 1}$,

661: we let $A_f\subseteq \T^{d+1}$ be the set

662: $\{(x,y)\in \T^{d+1} \st y\le f(x)\}$. 

663: Consider the projection $T_{A_f}$ of $L^2(\T^{d+1})$ given 

664: in \ref t.rudin/(i).

665: We view $L^2(\T^d)$ as a subspace of $L^2(\T^{d+1})$ by identifying

666: $g\in L^2(\T^d)$ with $g\otimes \bfo \in L^2(\T^{d+1})$, where

667: $(g\otimes \bfo) (x,y):= g(x)$ for $x\in \Td$, $y\in \T$.

668: The orthogonal projection $P$ of $L^2(\T^{d+1})$ onto $L^2(\T^d)$ is then

669: given by $g \mapsto (x \mapsto \int_\T g(x,y) \, dy)$. A very simple

670: calculation, 

671: left to the reader, shows that $M_f$ is a compression of $T_{A_f}$;

672: i.e., 

673: $$

674: M_f=P T_{A_f}

675: \label e.compression

676: $$ 

677: on $L^2(\T^d)$ viewed as a subspace of $L^2(\T^{d+1})$. For later use, let 

678: $$

679: H_f:=\{g \in L^2(\T^{d+1}) \st g = 0 \hbox{ a.e.\ on } (A_f)^c \}

680: $$

681: be the image of $T_{A_f}$.

682: 

683: We next remind the reader of the notion of stochastic domination between 

684: two probability measures

685: on $2^E$. First, if $\eta,\delta$ are elements of $2^E$, we write 

686: $\eta \preccurlyeq\delta$ if $\eta(e) \le \delta(e)$ for all $e\in E$. A 

687: subset $A$ of

688: $2^E$ is called {\bf increasing} if $\eta\in A$ and $\eta

689: \preccurlyeq\delta$ imply that $\delta\in A$.

690: If $\nu$ and $\mu$ are two probability measures on $2^E$, we write $\nu

691: \preccurlyeq \mu$ if $\nu(A) \le \mu(A)$ for all increasing sets $A$.

692: A theorem of \ref b.Strassen/ says that this is equivalent to the 

693: existence of a probability

694: measure $m$ on $2^E \times 2^E$ that has $\nu$ and $\mu$ as its first and

695: second marginals (i.e., $m$ is a {\bf coupling} of $\nu$ and $\mu$)

696: and such that $m$ is concentrated on the set

697: $\{(\eta,\delta):\eta \preccurlyeq\delta\}$ (i.e., $m$ is {\bf monotone}).

698: 

699: Throughout the paper, we shall use 

700: the following consequence of conditional negative association (\ref

701: t.negass/), sometimes called

702: {\bf joint negative regression dependence}

703: (see \ref b.Pemantle:negass/).

704: 

705: \procl p.JNRD

706: Assume that $\{X_i\}_{i\in I}$ has conditional negative association,

707: $I$ is the disjoint union of $I_1$ and $I_2$, and 

708: $a,b\in \{0,1\}^{I_1}$ with $a_i\le b_i$ for each $i\in I_1$.

709: Then 

710: $$

711: [\{X_i\}_{i\in I_2}\mid X_i=b_i, i\in I_1]

712: \preccurlyeq [\{X_i\}_{i\in I_2}\mid X_i=a_i, i\in I_1]\,,

713: $$

714: where $[ Y \mid A]$ stands for the law of $Y$ conditional on $A$.

715: \endprocl

716: 

717: \procl l.mon

718: Let $f_1,f_2:\Td\to [0,1]$ with $f_1\le f_2$ a.e. Then

719: $\P^{f_1} \preccurlyeq  \P^{f_2}$.

720: \endprocl 

721: 

722: \proof

723: This follows from a more general result (see \ref b.L:det/) that 

724: says that if $Q_1$ and $Q_2$ are two commuting positive contractions 

725: on $\ell^2(E)$ such that $Q_1\le Q_2$ in the sense that $Q_2-Q_1$ 

726: is positive, then $\P^{Q_1} \preccurlyeq  \P^{Q_2}$. 

727: However, here is a more concrete proof in our case. Since

728: $f_1\le f_2$ a.e., it follows that $H_{f_1}\subseteq H_{f_2}$, which implies

729: by Theorem 6.2 in \ref b.L:det/ that the projection measures 

730: $\P^{H_{f_1}}$ and $\P^{H_{f_2}}$ on $\Z^{d+1}$ satisfy 

731: $\P^{H_{f_1}}  \preccurlyeq  \P^{H_{f_2}}$ and hence their restrictions to $\Z^d$ satisfy

732: the same relationship; i.e., $\P^{f_1} \preccurlyeq\P^{f_2}$.

733: \Qed

734: 

735: 

736: We close this section with a key

737: representation of certain conditional probabilities and an application.

738: The minimum in \ref e.closest/ below is often referred to as a Szeg\H{o}

739: infimum.

740: For an infinite set $B \subseteq \Z^d \setminus \{ \bfz \}$, write

741: $$

742: \P^f[\eta(\bfz) = 1 \mid \eta \restrict B \equiv 1]

743: :=

744: \lim_{n \to\infty}

745: \P^f[\eta(\bfz) = 1 \mid \eta \restrict B_n \equiv 1]

746: \,,

747: $$

748: where $B_n$ is any increasing sequence of finite subsets of $B$ whose union is

749: $B$. This is a decreasing limit by virtue of \ref p.JNRD/.

750: Let $\ip{\cdot, \cdot}_f$ denote the usual inner product in the complex

751: Hilbert space $L^2(f)$.

752: For any set $B \subset \Z^d$, write $[B]_f$ for the closure in $L^2(f)$

753: of the linear span of the complex exponentials $ \{ \ee_k \st k

754: \in B \} $.

755: 

756: \procl t.closest

757: Let $f : \Td \to [0, 1]$ be measurable and $B\subset\Z^d$ with

758: $\bfz \notin B$.

759: Then 

760: $$

761: \P^f[\eta(\bfz) = 1 \mid \eta \restrict B \equiv 1]

762: =

763: \min \left\{ \int |\bfo - u|^2 f \,d\lambda_d \st u \in [B]_f \right\}

764: \,.

765: \label e.closest

766: $$

767: \endprocl

768: 

769: 

770: \proof

771: It suffices to prove the theorem for $B$ finite since the infinite case then

772: follows by a simple limiting argument.

773: So assume that $B$ is finite.

774: Note that $\widehat f(k-j) = \ip{\ee_j, \ee_k}_f$, so that 

775: $$

776: \P^f[\eta \restrict B \equiv 1] = 

777: \det [ \ip{\ee_j, \ee_k}_f]_{j, k \in B}

778: \,,

779: $$

780: and similarly for $\P^f[\eta \restrict (B \cup \{ \bfz \}) \equiv 1]$.

781: Thus, the left-hand side of \ref e.closest/ is a quotient of

782: determinants. The fact that such a quotient has the form of the right-hand

783: side is sometimes called Gram's formula. We include the proof

784: for the convenience of the reader.

785: Since $\ee_\bfz = \bfo$, it follows by row operations on the matrix

786: $[\ip{\ee_j, \ee_k}]_{j, k \in B \cup \{ \bfz \}}$ that

787: $$

788: \P^f[\eta(\bfz) = 1 \mid \eta \restrict B \equiv 1]

789: =

790: \|P_{[B]_f}^\perp \bfo\|_f^2

791: \,,

792: \label e.byproj

793: $$

794: where $P_{[B]_f}^\perp$ denotes orthogonal projection onto the orthogonal

795: complement of $[B]_f$ in $L^2(f)$.

796: Since this is the squared distance from $\bfo$ to $[B]_f$,

797: the equation \ref e.closest/ now follows.% from the next lemma.

798: \Qed

799: 

800: An extension of the above reasoning, given in \ref b.L:det/, provides

801: the entire conditional probability measure:

802: 

803: \procl t.condonB

804: Let $f : \Td \to [0, 1]$ be measurable and $B\subset\Z^d$.

805: Then the law of $\eta \restrict (\Zd \setminus B)$ conditioned on

806: $\eta \restrict B \equiv 1$ is the determinantal probability measure

807: corresponding to the positive contraction on $\ell^2(\Zd \setminus B)$

808: whose $(j, k)$-matrix entry is 

809: $$

810: \bigip{P_{[B]_f}^\perp \ee_j, P_{[B]_f}^\perp \ee_k}_f

811: \,.

812: $$

813: for $j, k \notin B$.

814: \endprocl

815: 

816: 

817: 

818: The Szeg\H{o} infimum that appears in \ref t.closest/ involves

819: trigonometric approximation, a classical area that

820: has strong connections to the topics of prediction and interpolation

821: for wide-sense stationary processes. We briefly discuss these topics

822: now. In later sections, we describe more explicit

823: connections to our results.

824: Recall that a mean-0 wide-sense stationary process is a

825: (not necessarily stationary) process

826: $\Seq{Y_n}_{n\in \Zd}$ for which all the variables have finite

827: variance and mean 0, and such that for each $k \in \Zd$, the covariance

828: $\Cov(Y_{n+k}, Y_n) = \E[Y_{n+k} \overline{Y_n}]$ does not depend on $n$.

829: There is then a positive measure $G$ (called the spectral measure) 

830: on $\T^d$ satisfying

831: $$

832: \widehat{G}(k)=\Cov(Y_{n+k}, Y_n)

833: $$

834: for $n,k\in \Zd$.

835: It turns out that for a one-dimensional wide-sense stationary process,

836: if $G$ is absolutely continuous with density $g$, then

837: $\GM(g)=0$ iff 

838: perfect linear prediction is possible, which means that $Y_0$ is in the closed

839: linear span of $\{Y_n\}_{n\le -1}$ in $L^2(\Omega)$,

840: where $\Omega$ is the underlying probability

841: space. This was proved in various versions by

842: Szeg\H{o}, Kolmogorov and Kre{\u\i}n.

843: Since the covariance with respect to $\P^f$ of 

844: $\eta(\bfz)$ and $\eta(k)$ for $k \in \Z^d$ is also

845: given by a Fourier coefficient (namely $-|\widehat f(k)|^2$ for $k \ne

846: \bfz$)

847: and since, as we shall see, the geometric mean of $f$ will play an 

848: important role in classifying the behavior of $\P^f$ as well,

849: one might wonder about the relationship between our

850: determinantal processes and wide-sense stationary processes.

851: It is not hard to show that $\P^f$ 

852: (viewed as a wide-sense stationary process)

853: has a spectral measure $G$ that is absolutely continuous with

854: density $g$ given by the formula

855: $$

856: g:= \widehat{f}(\bfz) \bfo -f*\widetilde{f}

857: \,,

858: $$

859: where $*$ denotes convolution and

860: $\widetilde{f}(t):=f(1-t)$.

861: Other than the trivial cases $f = \bfz$ and $f = \bfo$,

862: it is easy to check that $g$ is bounded away from 0 

863: and so, in particular, its geometric mean is always

864: strictly positive.

865: This suggests that our results are perhaps not so connected to

866: prediction and interpolation. However, it turns out that

867: some of the questions that

868: we deal with here (such as phase multiplicity and domination)

869: concerning $\P^f$ do have interpretations in terms of prediction and

870: interpolation for wide-sense stationary processes whose spectral

871: density is $f$ (which does not include $\P^f$).

872: More specifics will be given in the relevant sections.

873: 

874: Special attention is often devoted to stationary Gaussian processes, one

875: reason being that their distribution is determined entirely by their

876: spectral measure. It is known that

877: a stationary Gaussian process with no deterministic component

878: is a multistep Markov chain iff its spectral density is the reciprocal of a

879: trigonometric polynomial; see \ref b.Doob:elem/. 

880: The analogous property for determinantal processes is regeneration:

881: 

882: \procl p.regenerate

883: If $d=1$, then $f$ is the reciprocal of a trigonometric polynomial of

884: degree at most $n$ iff $\P^f$ is a regenerative process that regenerates

885: after $n$ successive $1$s appear.

886: \endprocl

887: 

888: This last property means that for any $k$, given that

889: $\eta \restrict [k+1, k+n] \equiv 1$, the future, $\eta \restrict

890: \CO{k+n+1,\infty}$, is conditionally independent of the past, $\eta

891: \restrict \OC{-\infty, k}$.

892: 

893: \proof

894: Note first that because of \ref t.condonB/, this

895: regenerative property holds for $\P^f$ iff for all $j \ge 0$ and all $C

896: \subset \OC{-\infty, -n-1}$, we have $P_{[B]_f}^\perp \ee_j = P_{[B \cup

897: C]_f}^\perp \ee_j$, where $B := [-n, -1]$.

898: (Here, we are relying on the fact that $\|P_{[B]_f}^\perp \ee_j\|_f > \|P_{[B

899: \cup C]_f}^\perp \ee_j\|_f$ if the vectors are not equal.)

900: This is the same as $P_{[B]_f}^\perp \ee_j \perp [C]_f$ for all $j \ge 0$, or,

901: in other words, there exists some $u_j \in [B]_f$ with $\ee_j - u_j \perp [B

902: \cup C]_f$. As this would have to hold for all $C$

903: (and $u_j$ is independent of $C$), 

904: it is the same as the existence of

905: some $u_j \in [B]_f$ such that $T_j := (\ee_j - u_j) f$ is analytic, i.e.,

906: $\widehat {T_j}(k) = 0$ for all $k < 0$. 

907: Now this implies that $\overline{T_0} = (\bfo - \overline{u_0}) f$, whence 

908: $T_0 (\bfo - \overline{u_0}) = \overline{T_0} (\bfo - u_0)$.

909: %%(Note that we may divide by analytic and conjugate analytic functions since

910: %%they vanish only on a set of measure 0; see \ref b.Rudin:RCA/, Theorem 17.18,

911: %%p.~345.)

912: The left side of this last 

913: equation is an analytic function, while the right side

914: is the conjugate of an analytic function (i.e., its Fourier coefficients

915: vanish on $\Z^+$).

916: Therefore, both equal some constant, $c$.

917: Hence 

918: $$

919: f = {T_0 \over \bfo - u_0} = {c \over (\bfo - u_0) (\bfo - \overline{u_0})}

920: = {c \over |\bfo - u_0|^2}

921: \,,

922: $$

923: which is indeed the reciprocal of a trigonometric polynomial of degree at most

924: $n$.

925: 

926: Conversely, if $1/f$ is the reciprocal of a trigonometric polynomial of degree

927: at most $n$, then since $f \ge 0$, the theorem of Fej\'er and Riesz

928: (see \ref b.GrenanderSzego/, p.~20)

929: allows us to write $f = c/|\bfo - u_0|^2$ for some conjugate-analytic

930: polynomial $u_0 \in [B]_f$ such that the analytic extension of

931: $\bfo - \overline{u_0}$ to the unit disc has no zeroes.

932: We may rewrite this as $(\bfo - u_0) f = T_0$ for $T_0 := c/(\bfo -

933: \overline{u_0})$. 

934: Since $T_0$ has an extension to the unit disc as the reciprocal of an

935: analytic polynomial with no zeroes, it follows that $T_0$ is also analytic.

936: Multiplying both sides of this equation by $\ee_1$ and rewriting $\ee_1 u_0 f =

937: c_1 (\bfo - u'_1) f = c_1 (u_0 - u'_1) f + c_1 T_0$ for some constant $c_1$ and

938: some $u'_1 \in [B]_f$, we see that $(\ee_1 - u_1) f = T_1$ for $u_1 := c_1 (u_0

939: - u'_1) \in [B]_f$ and $T_1 := \ee_1 T_0 + c_1 T_0$, an analytic function.

940: Similarly, we may establish by induction that for each $j \ge 0$,

941: there is some $u_j \in [B]_f$ such that $T_j := (\ee_j - u_j) f$ is 

942: analytic.

943: This proves the equivalence desired.

944: %(One can also see that

945: %$T_{k+1} = \ee_1 T_k + c_{k+1} T_0$ for some constant $c_{k+1}$.)

946: %\msnote{There is a more explicit form available, so I don't think this adds

947: %much.}

948: \Qed

949: 

950: We now use this proof to establish the explicit probabilistic form of the

951: renewal process in \ref x.renewal/.

952: In this case, $B = \{-1\}$ and one easily verifies that $u_j = a^{j+1}

953: \ee_{-1}$, as is standard in the theory of linear prediction. 

954: Therefore, 

955: $$

956: \P^f[\eta(j) = 1 \mid \eta(-1) = 1]

957: =

958: \|P_{[B]_f}^\perp \ee_j\|_f^2

959: =

960: \|\ee_j - u_j\|_f^2

961: =

962: \|\ee_j\|_f^2 - \|u_j\|_f^2

963: =

964: {1-a \over 1+a} (1 - a^{2j+2})

965: \,.

966: $$

967: It now suffices to verify that this is also true for the explicit renewal

968: process described in \ref x.renewal/.

969: First, it is well known from basic renewal theory that for a renewal 

970: process, the probabilities

971: $\P^f[\eta(j) = 1 \mid \eta(-1) = 1]$ determine the distribution of the 

972: number of 0s between two 1s. Hence to verify the above statement,

973: one simply needs to check that these latter probabilities are related to

974: the interrenewal distribution via the appropriate convolution-type equation.

975: In this case, it comes down to verifying that for all $j\ge 0$,

976: $$

977: {1-a \over 1+a}(1-a^{2j+2})

978: =

979: (j+1)(1-a)^2a^j

980: +{1-a \over 1+a}\sum_{k=1}^j k (1-a)^2a^{k-1}(1-a^{2(j-k)+2})\,.

981: $$

982: This identity is easy to check.

983: 

984: 

985: 

986: \bsection {The Bernoulli Shift Property}{s.bern}

987: 

988: In this section,

989: we prove that the stationary determinantal processes studied here

990: are Bernoulli shifts. We 

991: assume the reader is familiar with the

992: basic notions of a Bernoulli shift (see, e.g., \ref b.Ornstein:book/).

993: 

994: \procl t.Bern

995: Let $f : \T^d \to [0, 1]$ be measurable.

996: Then $\P^f$ is a Bernoulli shift; i.e., it is isomorphic 

997: (in the sense of ergodic theory) to an i.i.d.\ process.

998: \endprocl

999: 

1000: Before beginning the proof, we present a few preliminaries.

1001: We first recall the definition of the $\dbar$-metric.

1002: 

1003: \procl d.overd

1004: If $\mu$ and $\nu$ are 

1005: $\Z^d$-invariant probability measures

1006: on $2^{\Z^d}$, then

1007: $$

1008: \dbar(\mu,\nu) := \inf_m m \Big[ \big \{ (\eta,\delta) \in 2^{\Z^d} \times

1009: 2^{\Z^d} \st \eta(\bfz)\ne \delta(\bfz) \big \}\Big]\, ,

1010: $$

1011: where the infimum is taken over all couplings $m$ of $\mu$ and 

1012: $\nu$ that are $\Z^d$-invariant.

1013: \endprocl 

1014: 

1015: The following is a slight generalization of a well-known result

1016: (see, for example, page 75 in \ref b.Liggett/). The 

1017: proof is an immediate consequence of the existence of a monotone coupling

1018: and is left to the reader.

1019: 

1020: 

1021: \procl l.weak

1022: Suppose that $\sigma_1$, $\sigma_2$, $\nu$ and $\mu$ are

1023: $\Z^d$-invariant probability measures on $2^{\Z^d}$ such that $\nu

1024: \preccurlyeq \sigma_i \preccurlyeq  \mu$ for both $i=1, 2$. 

1025: Then 

1026: $$

1027: \dbar(\sigma_1, \sigma_2) \le \mu[\eta(\bfz) = 1] - \nu[\eta(\bfz)= 1]

1028: \,.

1029: $$

1030: \endprocl 

1031: 

1032: \procl p.L1givesdbar

1033: Let $f, g : \T^d \to [0, 1]$ be measurable.

1034: Then 

1035: $$

1036: \dbar(\P^f, \P^g) \le \int_{\T^d} |f-g| \,d\lambda_d

1037: \,.

1038: $$

1039: \endprocl 

1040: 

1041: \proof

1042: Because of \ref l.mon/, we may apply \ref l.weak/ to $\sigma_1 := \P^f$,

1043: $\sigma_2 := \P^g$, $\nu := \P^{f \wedge g}$, and $\mu := \P^{f \vee g}$.

1044: We obtain that 

1045: $$

1046: \dbar(\P^f, \P^g)

1047: \le

1048: \int_{\T^d} f \vee g \,d\lambda_d - \int_{\T^d} f \wedge g \,d\lambda_d

1049: =

1050: \int_{\T^d} |f-g| \,d\lambda_d

1051: \,.

1052: \Qed

1053: $$

1054: 

1055: \proofof t.Bern

1056: The first step is to approximate $f$ by 

1057: trigonometric polynomials. Let $K_r$ be the $r$th Fej\'er kernel for $\T$, 

1058: $$

1059: K_r := \sum_{|j| \le r} \Big(1-{|j|\over r+1}\Big) \ee_j 

1060: \,,

1061: $$

1062: and define $K^d_r(x_1, \ldots, x_d) := \prod_{i=1}^d K_r(x_i)$.

1063: It is well known 

1064: that $K^d_r$ is a positive summability kernel, so that

1065: if we define $g_r$ by

1066: $$

1067: g_r:= f * K^d_r\,,

1068: $$

1069: then $0\le g_r \le 1$ and $\lim_{r\to\infty} g_r =f$ a.e.\ and in $L^1(\Td)$. 

1070: 

1071: Next, since each $K^d_r$ is a trigonometric polynomial, so is each $g_r$.

1072: {}From this and \ref l.fourier/, it is easy to see that there is a constant 

1073: $C$ such that $\P^{g_r}(A\cap B) = \P^{g_r}(A)\P^{g_r}(B)$ 

1074: if $A$ is of the form $\eta\equiv 1 \hbox{ on } S_1$ and 

1075: $B$ is of the form $\eta\equiv 1 \hbox{ on } S_2$ 

1076: with $S_1$ and $S_2$ being finite sets 

1077: and having distance at least $C$ between them.

1078: {}From this, a standard argument in probability (a $\pi$-$\lambda$

1079: argument) shows that if $A$ is any event depending only on

1080: locations $S_1$ and if $B$ is any event depending only on locations $S_2$

1081: with $S_1$ and $S_2$ possibly infinite sets

1082: having distance at least $C$ between them, then

1083: $\P^{g_r}(A\cap B) = \P^{g_r}(A)\P^{g_r}(B)$. A process 

1084: with this property is called a {\bf finitely dependent} process 

1085: and this property implies it is a Bernoulli shift 

1086: (e.g., the so-called very weak Bernoulli condition 

1087: is immediately verified; see \ref b.Ornstein:book/ for this definition for

1088: $d=1$ and, e.g., \ref b.Steif/ for general $d$).

1089: 

1090: Since the processes that are Bernoulli shifts 

1091: are closed in the $\dbar$ metric (see

1092: \ref b.Ornstein:book/) and \ref p.L1givesdbar/ tells us that

1093: $\lim_{r\to\infty} \dbar(\P^{g_r},\P^f) = 0$, we 

1094: conclude that $\P^f$ is a Bernoulli shift.

1095: \Qed

1096: 

1097: \procl r.WB

1098: An important property of 1-dimensional 

1099: stationary processes is the weak Bernoulli

1100: (WB) property. This is also referred to as ``$\beta$-mixing" and ``absolute

1101: regularity" in the literature. Despite its name, it is known that WB is

1102: strictly stronger than Bernoullicity. It is easy to check that ``aperiodic''

1103: regenerative processes and finitely dependent processes are WB. Hence,

1104: by our earlier results, if $f$ is a trigonometric polynomial or the inverse

1105: of a trigonometric polynomial (or if $\bfo-f$ is the inverse

1106: of a trigonometric polynomial), then $\P^f$ is WB. This is subsumed by the

1107: independent

1108: work of \ref b.ShiTak:II/, who showed that $\P^f$ is WB whenever

1109: $\sum_{n \ge 1} n |\widehat f(n)|^2 <\infty$.

1110: The precise class of $f$ for which $\P^f$ is WB is not known.

1111: We also note that it follows from 

1112: \ref b.Smor/ that if

1113: $f$ is a trigonometric polynomial, then $\P^f$ is finitarily isomorphic

1114: to an i.i.d.\ process.

1115: \endprocl

1116: 

1117: \bsection {Support of the Measures}{s.fs}

1118: 

1119: In this section, we show that all of the probability measures $\P^f$ have

1120: full support except in two degenerate cases. We begin with the following

1121: lemma.

1122: 

1123: \procl l.lindep

1124: Let $H$ be a closed subspace of $\hE$ and let $A$ and $B$ be finite disjoint

1125: subsets of $E$. Then

1126: $$

1127: \P^H[\eta(e)=1 \hbox{ for } e\in A, \,\, \eta(e)=0 \hbox{ for } e\in B] >0

1128: $$

1129: if and only if 

1130: $\{P_H(e)\}_{e\in A}\cup\{P_{H^\perp}(e)\}_{e\in B}$ is linearly independent.

1131: \endprocl

1132: 

1133: \proof

1134: In the special case that $A$ or $B$ is empty, the result follows from

1135: \ref b.L:det/.

1136: In general,

1137: $\{P_H(e)\}_{e\in A}\cup\{P_{H^\perp}(e)\}_{e\in B}$ is linearly independent

1138: if and only if

1139: $\{P_H(e)\}_{e\in A}$ and $\{P_{H^\perp}(e)\}_{e\in B}$ are each linearly

1140: independent sets. Also,

1141: $\P^H[\eta \restrict A \equiv 1,\, \eta \restrict B \equiv 0] >0$

1142: if and only if

1143: $\P^H[\eta \restrict A \equiv 1] >0$ and

1144: $\P^H[\eta \restrict B \equiv 0] >0$ since 

1145: $$

1146: \P^H[\eta \restrict A \equiv 1, \,\, \eta \restrict B \equiv 0] \ge  

1147: \P^H[\eta \restrict A \equiv 1] \P^H[\eta \restrict B \equiv 0] 

1148: $$ 

1149: by the negative association property, \ref t.negass/.

1150: Thus, the general case follows from the special case.

1151: \Qed

1152: 

1153: \procl t.support

1154: $\P^f$ has full support

1155: for any function $f:\Td\to [0,1]$ other than $\bfz$ or $\bfo$.

1156: \endprocl

1157: 

1158: \proof

1159: Since marginals of probability measures with full support clearly

1160: have full support, it suffices to prove the result for

1161: $\P^H$ when $H$ is a closed $\Z^d$-invariant

1162: subspace of $\hD$ other than

1163: $\hD$ or 0. If we translate over to $L^2(\T^d)$, we see, by \ref l.lindep/,

1164: that it suffices to show that

1165: if $A\subseteq \Td$ with $0< \lambda_d(A) < 1$

1166: and $n_1,\ldots,n_k, m_1,\ldots,m_\ell$ are

1167: all distinct elements of $\Z^d$, then

1168: $$

1169: \{\ee_{n_j} \I A\}_{1\le j \le k} \cup

1170: \{\ee_{m_r} \I {A^c}\}_{1\le r \le \ell} 

1171: $$

1172: is linearly independent.

1173: Suppose that

1174: $c_1,\ldots,c_k, d_1,\ldots,d_\ell$ are complex numbers such that

1175: $$

1176: \sum_{j=1}^k c_j \ee_{n_j} \I A +

1177: \sum_{r=1}^\ell d_r \ee_{m_r}  \I {A^c} = \bfz 

1178: \,.

1179: $$

1180: From this it follows that

1181: $\sum_{j=1}^k c_j \ee_{n_j}=0$ a.e.\ on $A$. 

1182: Since $\lambda_d(A) >0$,

1183: we have that $\sum_{j=1}^k c_j \ee_{n_j}$ is 0 

1184: on a set of positive measure.

1185: It is well known that this implies that $c_1,\ldots,c_k$ vanish (the proof

1186: uses induction on $d$ and Fubini's theorem). Similarly, $d_1,\ldots,d_\ell$

1187: vanish.

1188: \Qed

1189: 

1190: 

1191: \bsection {Domination Properties}{s.dom}

1192: 

1193: In this section, we study the question of

1194: which product measures are stochastically dominated by $\P^f$ and

1195: which product measures stochastically dominate $\P^f$. 

1196: For simplicity, we give our first results for $d=1$, and only afterwards

1197: describe how these results extend to higher

1198: dimensions. We also treat a different 

1199: notion of ``full domination" at the end of the

1200: section.

1201: Recall that for $f : \Td \to [0, 1]$, we define 

1202: $$

1203: \GM(f) := \exp \int_\Td \log f \,d\lambda_d

1204: \,.

1205: $$

1206: 

1207: We introduce an auxiliary stronger

1208: notion of domination than $\preccurlyeq$, but only in the case where

1209: one of the measures is a product measure. Let $\mu_p$ denote product

1210: measure with density $p$. 

1211: 

1212: \procl d.sdom

1213: A stationary process $\{\eta_n\}_{n\in\Z}$ with distribution $\nu$

1214: {\bf strongly dominates} $\mu_p$, written 

1215: $\mu_p\strle\nu$,

1216: if for any $n$ and any $a_1,\ldots,a_n \in \{0,1\}$,

1217: $$

1218: \P[\eta_0=1 \mid \eta_i=a_i, i=1,\ldots, n]\ge p

1219: \,.

1220: $$

1221: Similarly, we define $\nu\strle\mu_p$ if the above inequality holds

1222: when $\ge$ is replaced by $\le$.

1223: \endprocl

1224: 

1225: The following lemma is easy and well known; it is sometimes referred to as 

1226: Holley's lemma.

1227: 

1228: \procl l.holley

1229: If $\mu_p\strle \nu$, then $\mu_p\preccurlyeq\nu$.

1230: \endprocl 

1231: 

1232: The converse is not true, as we shall see later (\ref r.support/).

1233: In light of \ref l.mon/, we have that

1234: if $p \le f \le q$, then 

1235: $$

1236: \mu_p \preccurlyeq \P^f \preccurlyeq \mu_q

1237: \,.

1238: $$

1239: The optimal improvement of these stochastic bounds is as follows.

1240: 

1241: \procl t.dom

1242: For any $f : \T \to [0, 1]$, we have

1243: $\mu_p\preccurlyeq\P^f$ iff $p\le \GM(f)$ iff $\mu_p\strle\P^f$. Similarly,

1244: $\P^f\preccurlyeq\mu_q$ iff $q\ge 1-\GM(\bfo-f)$ iff $\P^f\strle\mu_q$.

1245: In addition, for any stationary process $\mu$ that 

1246: has conditional negative association, we have

1247: $\mu_p\preccurlyeq\mu$ iff $\mu_p\strle\mu$.

1248: \endprocl 

1249: 

1250: \proof

1251: Let $d_n := D_{n-1}(f)$ be the probability of having $n$ 1s in a row.

1252: According to Szeg\H{o}'s theorem (\ref b.GrenanderSzego/, pp.~44, 66),

1253: $d_{n+1}/d_n$ is decreasing in $n$

1254: and

1255: $$

1256: \lim_{n\to\infty}d_{n+1}/d_n= \GM(f)

1257: = \lim_{n \to\infty} d_n^{1/n}

1258: \,.

1259: $$

1260: In particular, $d_{n+1}/d_n\ge \GM(f)$ for all $n$.

1261: 

1262: \ref p.JNRD/ implies that

1263: for any fixed $n$,

1264: $$

1265: \P^f[\eta_0=1 \mid \eta_i=a_i, i=1,\ldots, n]

1266: $$

1267: is minimized among all $a_1,\ldots,a_n \in \{0,1\}$

1268: when $a_1 = a_2 = \cdots = a_n = 1$.

1269: In this case, the value is $d_{n+1}/d_n$.

1270: Since this is at least $\GM(f)$, 

1271: we deduce that if $p\le \GM(f)$, then

1272: $\mu_p\strle \P^f$ and hence that $\mu_p\preccurlyeq\P^f$. 

1273: 

1274: Conversely, if $\mu_p \preccurlyeq \P^f$, then certainly

1275: $p^n\le d_n$ for all $n$. 

1276: Hence $p \le \GM(f)$. The second to last statement can be proved in 

1277: the same way or can be concluded by symmetry. 

1278: Finally, the last statement can be proved in a similar fashion.

1279: \Qed

1280: 

1281: % The proof also shows the equivalence with $d_n \ge p^n$ for all $n$.

1282: 

1283: \procl r.examples

1284: The above theorem gives us two interesting examples. 

1285: First, if we take $f:=\I A$ where $A$ has Lebesgue measure $1-\epsilon$, then

1286: $\dbar(\P^f, \delta_{\bfo}) \le \epsilon$ by 

1287: \ref p.L1givesdbar/, but nonetheless, $\P^f$ does not dominate

1288: any nontrivial product measure since $\GM(f) = 0$.

1289: Second, if we take

1290: $f:={\bf 1}_{[0,1/2]}+ .4\cdot {\bf 1}_{[1/2,1]}$, 

1291: then $f < 1/2$ on a set of positive

1292: measure, but nonetheless $\P^f$ dominates $\P^{1/2}=\mu_{1/2}$

1293: since $\GM(f) = (.4)^{1/2} > 1/2$.

1294: \endprocl

1295: 

1296: 

1297: \procl x.ustH_x-dom

1298: Let $f$ be as in \ref x.ustH/, so that $\P^f$ is the law of the horizontal

1299: edges of the uniform spanning tree in the plane.

1300: In order to examine a 1-dimensional process, let us consider only

1301: the edges lying on the $x$-axis.

1302: If we let 

1303: $$

1304: g(x) := \int_\T f(x, y) \,d\lambda_1(y)

1305: \,,

1306: $$

1307: then the edges lying on the $x$-axis have the law $\P^g$

1308: since $\widehat g(k) = \widehat f(k, 0)$ for all $k \in \Z$.

1309: Since an antiderivative of $1/(1 + a\sin^2 \pi y)$ is 

1310: $$

1311: {\arctan \left( \sqrt{1+a}\tan \pi y\right) \over \pi \sqrt{1+a}}

1312: \,,

1313: $$

1314: we have 

1315: $$

1316: g(x) =

1317: {\sin \pi x \over \sqrt{1+\sin^2 \pi x}}

1318: \,,

1319: $$

1320: as given in \ref x.ustHx/.

1321: We claim that $\P^g$ strongly dominates $\mu_p$ for $p := \sqrt 2 - 1$, and

1322: this is optimal.

1323: In order to show this, we calculate $\GM(g)$.

1324: Write $g_1(x) := \sin^2 \pi x$. 

1325: Then $(\GM(g))^2 = \GM(g_1)/\GM(\bfo + g_1)$.

1326: Let $G_1$ be the harmonic extension of $\log g_1$ from the circle

1327: to the unit disc.

1328: Then $\int \log g_1 \,d\lambda_1 = G_1(0)$ by the mean value property of

1329: harmonic functions.

1330: Since $g_1 = |(1-\ee_1)/2|^2$, we see that $G_1$ is the

1331: real part of the analytic function $z \mapsto 2 \log [(1-z)/2]$

1332: in the unit disc, from which we conclude that $G_1(0) = \log (1/4)$.

1333: Therefore, $\GM(g_1) = 1/4$. 

1334: Similarly, $\bfo+g_1 = |[\sqrt 2 + 1 - (\sqrt 2 - 1)\ee_1]/2|^2$, whence

1335: $\GM(\bfo+g_1) = [(\sqrt 2 + 1)/2]^2$.

1336: Therefore $\GM(g) = \sqrt 2 - 1 = 0.4142^+$, as desired.

1337: It turns out that $q := 1 - \GM(\bfo - g) = 1 - 2(\sqrt 2 -1)e^{-2\cat/\pi} =

1338: 0.5376^+$, where 

1339: $$

1340: \cat := \sum_{k=0}^\infty {(-1)^k \over (2k+1)^2} = 0.9160^-

1341: \label e.defcat

1342: $$

1343: is {\bf Catalan's constant}.

1344: Thus, $\P^g \strle \mu_q$.

1345: It is interesting how close $p$ and $q$ are.

1346: \endprocl

1347: 

1348: We now introduce some new mixing conditions.

1349: Given a positive integer $r$, we may restrict $\P^f$ to $2^{r\Z}$.

1350: If we identify $r\Z$ with $\Z$, then we obtain the process $\P^{f_r}$, where 

1351: $$

1352: f_r(t) := {1 \over r} \sum_{x \in r^{-1} t} f(x)

1353: \,,

1354: $$

1355: where $r^{-1} t := \{ x \in \T \st r x = t \}$.

1356: The reason that this restriction is equal to $\P^{f_r}$ is that

1357: for all $k \in \Z$, we have $\widehat{f_r}(k) = \widehat f(rk)$, as

1358: is easy to check.

1359: Because of this relation, we have that $\int |f_r(t) - \widehat

1360: f(0)|^2\,d\lambda_1(t) \to 0$ as $r \to\infty$, so that by \ref l.weak/, it

1361: follows that $\dbar(\P^{f_r}, \P^{\widehat f(0)}) \to 0$ as $r \to\infty$.

1362: (It is not hard to show that a similar property holds for any Kolmogorov

1363: automorphism (see \ref d.K/), while the example in \ref r.support/ shows

1364: that this property can occur in other cases as well.)

1365: In fact, we often have a stronger convergence for our determinantal processes,

1366: as we show next.

1367: 

1368: \procl t.dom-mix

1369: Let $f : \T \to [0, 1]$ be measurable.

1370: If $\GM(f) > 0$ or $f$ is bounded away from 0 on an interval of positive length,

1371: then there exist constants $p_r \to \widehat f(0)$ such that $\P^{f_r}

1372: \succcurlyeq \P^{p_r}$ for all $r$.

1373: Therefore, if $\GM\big(f(\bfo - f)\big) > 0$ or if $f$ is continuous and not

1374: equal to $\bfz$ nor $\bfo$, then there exist constants $p_r, q_r \to \widehat

1375: f(0)$ such that $\P^{p_r} \preccurlyeq \P^{f_r} \preccurlyeq \P^{q_r}$ for all

1376: $r$.

1377: \endprocl

1378: 

1379: Of course, in view of \ref t.dom/, we use $p_r := \GM(f_r)$ and $q_r := 1 -

1380: \GM(\bfo - f_r)$.

1381: Thus, the theorem is an immediate consequence of the following lemma.

1382: 

1383: \procl l.dom-mix

1384: Let $f : \T \to [0, 1]$ be measurable.

1385: If $\GM(f) > 0$ or $f$ is bounded away from 0 on an interval of positive length,

1386: then $\GM(f_r) \to \widehat f(0)$ as $r \to\infty$.

1387: \endprocl

1388: 

1389: \proof

1390: We have seen that $f_r \to \fav \cdot \bfo$ in $L^2$, whence also in measure.

1391: Therefore $\log f_r \to (\log \fav) \cdot \bfo$ in measure.

1392: Thus, it remains to show that $\{ \log f_r \}$ is uniformly integrable (at

1393: least, for all large $r$).

1394: Suppose first that $\GM(f) > 0$.

1395: 

1396: Given $h \in L^1(\T)$, write $h^{(r)}(t) := h(r t)$, so that

1397: $\widehat{h^{(r)}}(k)$ is 0 when $k$ is not a multiple of $r$ and is

1398: $\widehat h(k/r)$ when $k$ is a multiple of $r$.

1399: For any $g, h \in L^2(\T)$, we have

1400: \begineqalno

1401: \int_\T g_r(t) \overline{h(t)} \,d\lambda_1(t)

1402: &=

1403: \sum_{k \in \Z} \widehat{g_r}(k) \overline{\widehat h(k)}

1404: =

1405: \sum_{k \in \Z} \widehat{g}(r k) \overline{\widehat h(k)}

1406: =

1407: \sum_{k \in \Z} \widehat{g}(k) \overline{\widehat{h^{(r)}}(k)}

1408: \cr&=

1409: \int_\T g(t) \overline{h^{(r)}(t)} \,d\lambda_1(t)

1410: =

1411: \int_\T g(t) \overline{h(r t)} \,d\lambda_1(t)

1412: \,.

1413: \label e.turn

1414: \cr

1415: \endeqalno

1416: Therefore, for any set $A \subseteq \T$, we have 

1417: $$

1418: \int_{A} \log (f_r) \,d\lambda_1

1419: \ge

1420: \int_{A} (\log f)_r \,d\lambda_1

1421: =

1422: \int_{r^{-1} A} \log f\,d\lambda_1

1423: \,;

1424: $$

1425: the inequality follows by concavity of log, while the equality follows from

1426: \ref e.turn/ applied to $g := \log f$ and $h := \I A$.

1427: Given any $\epsilon \in \big(0, \fav\big)$, let $A^r_\epsilon := \{ t \st

1428: f_r(t) < \epsilon \}$.

1429: Note that for any measurable $A \subset \T$, we have

1430: $\lambda_1(r^{-1} A) = \lambda_1(A)$.

1431: Since $f_r \to \fav \cdot \bfo$ in measure, it follows that

1432: $$

1433: \lim_{r \to\infty} \lambda_1(r^{-1} A^r_\epsilon) =

1434: \lim_{r \to\infty} \lambda_1(A^r_\epsilon)

1435: = 0

1436: \,.

1437: $$

1438: Since $\log f$ is integrable, it follows that 

1439: $$

1440: \lim_{r \to\infty} \int_{r^{-1} A^r_\epsilon} \log f \,d\lambda_1 =

1441: 0

1442: \,.

1443: $$

1444: This establishes the uniform integrability in the first case.

1445: 

1446: In the second case where $f \ge c > 0$ on an interval of length

1447: $\epsilon > 0$, we have that $f_r \ge c \epsilon$ on all of $\T$ for

1448: all $r > 1/\epsilon$.

1449: It is then obvious that $\{\log f_r\}$ is uniformly integrable for all 

1450: $r > 1/\epsilon$.

1451: \Qed

1452: 

1453: \procl r.nasty

1454: There are functions $f$ with $f > 0$ a.e., yet $\GM(f_r) = 0$ for all $r$.

1455: For example, enumerate the rationals $\{ x_j \st j \ge 1\}$ in $(0, 1)$ and

1456: choose $\epsilon_j > 0$ such that $x_j + \epsilon_j < 1$ and $\sum_j

1457: \epsilon_j < 1$.

1458: Define $A_n := [0, 1] \setminus \bigcup_{j > n} [x_j, x_j + \epsilon_j]$ and

1459: $f_0(x) := e^{-1/x}$.

1460: Then one can show that 

1461: $$

1462: f(x) := f_0(x) \I{A_0}(x) +

1463: \sum_{n=1}^\infty f_0(x-x_n) \I{[x_n, x_n+\epsilon_n] \cap A_n}(x) 

1464: $$

1465: is such an example.

1466: \endprocl

1467: 

1468: 

1469: 

1470: 

1471: We turn next to the extension of the prior results to higher dimensions.

1472: This is basically straightforward, but has interesting applications, as we

1473: shall see.

1474: First, we recall the usual notion of the {\bf past $\sigma$-field} $\past(k)$

1475: at a point $k \in \Z^d$.

1476: We use lexicographic ordering on $\Z^d$, i.e., write $(k_1, k_2,

1477: \ldots, k_d) \prec (l_1, l_2, \ldots, l_d)$ if $k_i < l_i$ when $i$ is the

1478: smallest index such that $k_i \ne l_i$.

1479: Then define $\past(k)$ to be the $\sigma$-field generated by $\eta(j)$ for all

1480: $j \prec k$.

1481: More generally, the past $\sigma$-field could be defined with respect to 

1482: any {\bf ordering} of $\Z^d$, which means the selection of a set $\ord \subset

1483: \Z^d$ that has the properties $\ord \cup (-\ord) = \Zd \setminus \{ \bfz \}$, 

1484: $\ord \cap (-\ord) = \emptyset$, and $\ord+\ord \subset \ord$. The associated ordering 

1485: is that where $k \prec l$ iff $ l-k \in \ord$.

1486: $\past_\ord(u)$ will denote the past of $u$ with respect to the ordering

1487: $\ord$,

1488: i.e., $\{v \st v \prec u\}$.

1489: Thus, $\past_\ord(\bfz)$ is just $-\ord$. 

1490: (For a characterization of

1491: all orders, see \ref b.Teh/, \ref b.Zaiceva/, or \ref b.Trevisan/.)

1492: As before, let $\mu_p$ denote product measure with density $p$. 

1493: 

1494: 

1495: \procl d.sdom-d

1496: Given an ordering $\ord$, 

1497: a stationary process $\{\eta_n\}_{n\in\Z^d}$ with distribution $\nu$

1498: {\bf strongly dominates} $\mu_p$, written 

1499: $\mu_p\strle\nu$,

1500: if 

1501: $$

1502: \P[\eta_\bfz=1 \mid \past_\ord(\bfz)]\ge p \quad\nu\hbox{-a.s.}

1503: $$

1504: Similarly, we define $\nu\strle\mu_p$ if the above inequality holds

1505: when $\ge$ is replaced by $\le$. (Note that the ordering $\ord$ here is 

1506: suppressed in the notation.)

1507: \endprocl

1508: 

1509: Although we have phrased it differently, in one dimension, this is equivalent

1510: to \ref d.sdom/ when $\ord=\{-1,-2,\ldots\}$.

1511: Again, we have a version of Holley's lemma:

1512: 

1513: \procl l.holley-d

1514: Given any ordering $\ord$, 

1515: if $\mu_p\strle \nu$, then $\mu_p\preccurlyeq\nu$.

1516: \endprocl 

1517: 

1518: (Note that to produce a monotone coupling for a general ordering with

1519: respect to a set $\ord$, it is enough to do so for the measures

1520: restricted to any finite $B \subset \Zd$. Given such a $B$, order $B$ by the

1521: restriction of $\prec$ to $B$ and couple the measures by adding sites from

1522: $B$ in this order.)

1523: 

1524: We now prove

1525: 

1526: \procl t.dom-d

1527: Fix any ordering $\ord$.

1528: For any measurable $f : \Td \to [0, 1]$, we have

1529: $\mu_p\preccurlyeq\P^f$ iff $p\le \GM(f)$ iff $\mu_p\strle\P^f$. Similarly,

1530: $\P^f\preccurlyeq\mu_q$ iff $q\ge 1-\GM(\bfo-f)$ iff $\P^f\strle\mu_q$.

1531: In addition, for any stationary process $\mu$ that

1532: has conditional negative association, we have

1533: $\mu_p\preccurlyeq\mu$ iff $\mu_p\strle\mu$.

1534: \endprocl 

1535: 

1536: \proof

1537: \ref p.JNRD/ and \ref t.closest/ imply that 

1538: \begineqalno

1539: \essinf \P^f\Big[\eta(\bfz) = 1 \mid \past_\ord(\bfz)\Big] 

1540: &=

1541: \inf \Big\{ \P^f[\eta(\bfz) = 1 \mid \eta \restrict A \equiv 1] \st

1542: A \subset -\ord \hbox{ is finite} \Big\}

1543: \cr&=

1544: \P^f[\eta(\bfz) = 1 \mid \eta \restrict (-\ord) \equiv 1]

1545: \cr&=

1546: \min \left\{ \int |\bfo - T|^2 f \,d\lambda_d \st T \in [-\ord]_f \right\}

1547: \,.

1548: \endeqalno

1549: By \ref b.HelLow:I/, the latter quantity equals $\GM(f)$.

1550: Therefore, $p\le \GM(f)$ iff $\mu_p\strle\P^f$. 

1551: On the other hand, if $\mu_p\preccurlyeq\P^f$, then let $A(r)$ be the

1552: finite subset of $-\ord$ consisting of all points within some large radius $r$

1553: about $\bfz$.

1554: Let $a_1 \prec a_2 \prec \cdots \prec a_n$ be the elements of $A(r)$ in

1555: order, where $n := |A(r)|$, and write $A_j := A(r) \cap (-\ord+a_j)$.

1556: Then 

1557: $$

1558: p^n \le \P^f[\eta \restrict A(r) \equiv 1]

1559: =

1560: \prod_{j=1}^n \P^f[\eta(a_j) = 1 \mid \eta \restrict A_j \equiv 1]

1561: \,.

1562: $$

1563: Most terms in this product are quite close to $\GM(f)$, while all lie in

1564: $[\GM(f), 1]$.

1565: It follows that

1566: $$

1567: \lim_{r \to\infty} \P^f[\eta \restrict A(r) \equiv 1]^{1/n}

1568: =

1569: \GM(f)

1570: \,,

1571: $$

1572: so that $p \le \GM(f)$.

1573: Finally, the remaining statements can be proved in a similar fashion.

1574: \Qed

1575: 

1576: \procl r.past

1577: It follows from \ref t.dom-d/ that the choice of ordering $\ord$ does not

1578: determine whether $\mu_p \strle \P^f$.

1579: This is not true for general stationary processes, even in one dimension. 

1580: We are grateful to Olle H\"aggstr\"om for the following simple example.

1581: Let $\mu$ be the distribution on $\{0, 1 \}^2$ given by 

1582: $$

1583: \mu = (1/7) (\delta_{(0,0)} +\delta_{(0,1)}) + (2/7)\delta_{(1,0)} + (3/7)

1584: \delta_{(1,1)}

1585: \,.

1586: $$

1587: Let $\eta \in \{ 0, 1 \}^\Z$ be such that with probability 1/2, $(\eta_{2n},

1588: \eta_{2n+1})$ are chosen independently each with distribution $\mu$ and

1589: otherwise $(\eta_{2n-1}, \eta_{2n})$ are chosen independently each with

1590: distribution $\mu$.

1591: Let $\nu$ be the law of $\eta$.

1592: If $\ord := \Z^-$, then $\mu_p \strle \nu$ iff $p \le 1/2$, while

1593: if $\ord := \Z^+$, then $\mu_p \strle \nu$ iff $p \le 4/7$.

1594: \endprocl

1595: 

1596: \procl x.ustH-dom

1597: Let $f$ be as in \ref x.ustH/, so that $\P^f$ is the law of the horizontal

1598: edges of a uniform spanning tree in the square lattice $\Z^2$.

1599: By \ref b.Kas:dim/ or \ref b.Montroll/, we have

1600: $$

1601: \int_{\T^2} \log 4(\sin^2 \pi x + \sin^2 \pi y) \,d\lambda_2(x, y)

1602: =

1603: {4\cat \over \pi}

1604: =

1605: 1.1662^+

1606: \,,

1607: \label e.catalan

1608: $$

1609: where $\cat$ is again Catalan's constant \ref e.defcat/.

1610: As we have shown in \ref x.ustH_x-dom/, $\GM(\sin^2 \pi x)=1/4$,

1611: whence by \ref e.catalan/,

1612: $\GM(f) = e^{-4\cat/\pi} =\GM(\bfo - f)$, 

1613: where we are using the observation that $1-f(x, y) = f(y,x)$. 

1614: (This last identity has a combinatorial reason arising from planar dual

1615: trees.) 

1616: Therefore $\mu_p \preccurlyeq \P^f\preccurlyeq \mu_{1-p}$ for $p :=

1617: e^{-4\cat/\pi} = 0.3115^+$ and this $p$ is optimal (on each side).

1618: This result in itself is rather surprising and it would be fascinating to see

1619: an explicit monotone coupling.

1620: \endprocl

1621: 

1622: For $f:\Td\to [0,1]$ measurable and

1623: $r = (r_1, r_2, \ldots, r_d) \in \Z^d$ with all $r_j > 0$, define

1624: $$

1625: f_r(t) := {1 \over \prod_{j=1}^d r_j} \sum_{x \in r^{-1} t} f(x)

1626: \,,

1627: $$

1628: where $r^{-1} (t_1, t_2, \ldots, t_d) := \{ (x_1, x_2, \ldots, x_d) \in \Td

1629: \st \all j\ r_j x_j = t_j \}$.

1630: Write $r \to\infty$ to mean that $\min r_j \to\infty$.

1631: The following is a straightforward extension of \ref t.dom-mix/.

1632: We leave its proof to the reader.

1633: 

1634: \procl t.dom-mix-d

1635: Let $f : \Td \to [0, 1]$ be measurable.

1636: If $\GM(f) > 0$ or $f$ is positive on a non-empty open set,

1637: then there exist constants $p_r \to \widehat f(\bfz)$ as $r \to\infty$

1638: such that $\P^{f_r}

1639: \succcurlyeq \P^{p_r}$ for all $r$.

1640: Therefore, if $\GM\big(f(\bfo - f)\big) > 0$ or if $f$ is continuous and not

1641: equal to $\bfz$ nor $\bfo$, then there exist constants $p_r, q_r \to \widehat

1642: f(\bfz)$ such that $\P^{p_r} \preccurlyeq \P^{f_r} \preccurlyeq \P^{q_r}$ for

1643: all $r$.

1644: \endprocl

1645: 

1646: Consider now a domination property even stronger than our previously defined

1647: strong domination. 

1648: Let $\others$ be the $\sigma$-field generated by $\eta(k)$ for $k \ne \bfz$.

1649: 

1650: \procl d.fulldom

1651: A stationary process $\{\eta_n\}_{n\in\Z^d}$ with distribution $\nu$

1652: {\bf fully dominates} $\mu_p$, written 

1653: $\mu_p\fullle\nu$,

1654: if 

1655: $$

1656: \P[\eta_\bfz=1 \mid \others]\ge p \quad\nu\hbox{-a.s.}

1657: $$

1658: In this situation, we also say that $\nu$ is {\bf uniformly insertion tolerant

1659: at level} $p$.

1660: Similarly, we define $\nu\fullle\mu_p$ if the above inequality holds

1661: when $\ge$ is replaced by $\le$, and

1662: say that $\nu$ is {\bf uniformly deletion tolerant at level} $1-p$.

1663: We say that $\nu$ is {\bf uniformly insertion tolerant} if

1664: $\mu_p\fullle\nu$ for some $p > 0$ and

1665: that $\nu$ is {\bf uniformly deletion tolerant}

1666: if $\nu\fullle\mu_p$ for some $p < 1$.

1667: \endprocl

1668: 

1669: We show that the optimal level of uniform insertion tolerance of a

1670: determinantal process $\P^f$

1671: is the {\bf harmonic mean} of $f$, defined as 

1672: $$

1673: \HM(f) := \left(\int_\Td {d\lambda_d  \over f} \right)^{-1}

1674: \,.

1675: $$

1676: Note that $\HM(f) = 0$ iff $1/f$ is not integrable.

1677: 

1678: \procl t.fulldom

1679: For any measurable $f : \Td \to [0, 1]$, we have

1680: $\mu_p\fullle\P^f$ iff $p\le \HM(f)$. Similarly,

1681: $\P^f\fullle\mu_q$ iff $q\ge 1-\HM(\bfo-f)$.

1682: \endprocl

1683: 

1684: \proof

1685: By \ref p.JNRD/ and \ref t.closest/, we have

1686: \begineqalno

1687: \essinf \P^f\Big[\eta(\bfz) = 1 \mid \others\Big] 

1688: &=

1689: \inf \Big\{ \P^f[\eta(\bfz) = 1 \mid \eta \restrict A \equiv 1] \st

1690: A \subset \Zd \setminus \{ \bfz \} \hbox{ is finite } \Big\}

1691: \cr&=

1692: \P^f[\eta(\bfz) = 1 \mid \eta \restrict B \equiv 1]

1693: =

1694: \min \left\{ \int |\bfo - T|^2 f \,d\lambda_d \st T \in [B]_f \right\}

1695: \, ,

1696: \endeqalno

1697: where $B := \Zd \setminus \{ \bfz \}$.

1698: As shown by \refbmulti{Kolmog:sshs,Kolmog:iesrs} for $d=1$, the latter equals

1699: $\HM(f)$.

1700: The proof extends immediately to general $d$.

1701: This proves the first assertion. The second follows by symmetry.

1702: \Qed

1703: 

1704: \procl r.HM

1705: Since a proof of Kolmogorov's theorem that 

1706: $$

1707: \min \left\{ \int |\bfo - T|^2 f \,d\lambda_d \st T \in [B]_f \right\}

1708: =

1709: \HM(f)

1710: \,,

1711: $$

1712: where $B := \Zd \setminus \{ \bfz \}$,

1713: is difficult to find in readily accessible sources, we

1714: provide one here.

1715: We have

1716: $$

1717: \min \left\{ \int |\bfo - T|^2 f \,d\lambda_d \st T \in [B]_f \right\}

1718: =

1719: \|u\|^2_f\,,

1720: $$

1721: where

1722: $$

1723: u :=

1724: \|P^\perp_{[B]_f} \bfo\|^2_f

1725: \,.

1726: $$

1727: Now $g \perp [B]_f$ iff $g \in L^2(f)$ and $\widehat{gf}(k) = 0$ for all $k

1728: \in B$.

1729: The latter condition holds iff $gf$ is a constant.

1730: If $g \ne \bfz$, then we deduce that $1/f \in L^2(f)$, i.e., $\HM(f) > 0$.

1731: Therefore,

1732: $$

1733: [B]^\perp_f = \cases{0 &if $\HM(f) = 0$,\cr

1734:                \C /f &if $\HM(f) > 0$.\cr

1735:                }

1736: $$

1737: Hence $u = \bfz$ if $\HM(f) = 0$, while otherwise, 

1738: $$

1739: \|u\|^2_f = |\ip{\bfo, \sqrt{\HM(f)}/f}_f|^2

1740: =

1741: \HM(f)

1742: \,,

1743: $$

1744: as desired.

1745: \endprocl

1746: 

1747: 

1748: \procl r.strongnotfull

1749: It is easy to find $f$ such that $\GM(f)> 0$ and $\HM(f)= 0$.

1750: Indeed, a natural such example is the function $g$ of \ref x.ustHx/.

1751: For any such $f$, the corresponding process

1752: $\P^f$ strongly dominates a

1753: nontrivial product measure, but does not

1754: fully dominate any nontrivial product measure. In addition, for any function

1755: $f$ such that $\HM(f) > 0$

1756: and $f$ is not constant a.e., $\GM(f)> \HM(f)$ (as a consequence of

1757: Jensen's inequality), so that $\P^f$ will strongly dominate strictly more

1758: product measures than it will fully dominate.

1759: \endprocl

1760: 

1761: \procl x.ustH-delete

1762: Let $f$ be as in \ref e.ustH-d/, so that $\P^f$ is the law of the edges of the

1763: uniform spanning forest in $\Z^d$ that lie parallel to the $x_1$-axis.

1764: Then $\P^f$ is uniformly deletion tolerant 

1765: iff $d \ge 4$.

1766: This is because $1/(\bfo - f)$ is integrable iff $d \ge 4$.

1767: For example, when $d=4$, we obtain full domination by $\mu_p$ with $p :=

1768: 0.66425^-$, where we have calculated $\HM(\bfo-f)$ as follows. First, we have

1769: $$

1770: 1/(\bfo - f) = 

1771: 1 + {\sin^2 \pi x_1 \over \sum_{j=2}^4 \sin^2 \pi x_j}

1772: \,,

1773: $$

1774: whence 

1775: \begineqalno

1776: \int_{\T^4} 1/(\bfo - f) \,d\lambda_4

1777: &=

1778: 1 + {1 \over 2} \int_{\T^3} {1 \over \sum_{j=2}^4 \sin^2 \pi x_j}

1779: \,d\lambda_3(x_2, x_3, x_4)

1780: \cr&=

1781: 1 + {1 \over 2} \int_{\T^2} {1 \over \sqrt{\left(\sum_{j=2}^3 \sin^2 \pi x_j

1782: \right) \left( 1 + \sum_{j=2}^3 \sin^2 \pi x_j\right)}} \,d\lambda_2(x_2, x_3)

1783: \,.

1784: \cr

1785: \endeqalno

1786: This last integral has no simple form and is calculated numerically.

1787: This gives the value reported for $1 - \HM(\bfo - f)$.

1788: For large $d$, we have $1 - \HM(\bfo - f) \sim 1/d$. Indeed, write

1789: $\lambda_{d-1}^*$ for Lebesgue measure on $\R^{d-1}$. Letting

1790: $$

1791: A_d:=\int_{\T^{d-1}} {d - 1 \over \sum_{j=2}^{d} 2 \sin^2 \pi x_j}

1792: \,d\lambda_{d-1}(x_2, \ldots, x_d)\,,

1793: $$

1794: we have that 

1795: $$

1796: (d-1) (1 - \HM(\bfo - f))= {A_d\over 1+A_d/(d-1)}

1797: \,.

1798: $$

1799: Hence it suffices to show that $\lim_{d\to\infty}A_d=1$.

1800: To show this, note that

1801: \begineqalno

1802: A_d

1803: &=

1804: \int_0^\infty \lambda_{d-1}\Big[ \sum_{j=2}^{d} 2 \sin^2 \pi x_j <

1805: (d-1)/t\Big]

1806: \,dt

1807: \cr&\le

1808: \int_0^2 \lambda_{d-1}\Big[ \sum_{j=2}^{d} 2 \sin^2 \pi x_j < (d-1)/t\Big] \,dt

1809: \cr&\quad+ \int_2^{10(d-1)} 

1810: \lambda_{d-1}\Big[ \sum_{j=2}^{d} 2 \sin^2 \pi x_j < (d-1)/2\Big] \,dt

1811: \cr&\quad+ \int_{10(d-1)}^\infty

1812: \lambda_{d-1}^*\Big[ \all j |x_j| < 1/2 \hbox{ and }

1813: \sum_{j=2}^{d} 8 x_j^2 < (d-1)/t\Big] \,dt

1814: \cr&\to

1815: \int_0^2 \I{[0, 1]}(t) \,dt

1816: =

1817: 1

1818: \endeqalno

1819: as $d \to \infty$, where we have used the weak law of large

1820: numbers and the bounded convergence theorem for the first piece,

1821: a standard large-deviation result for the second piece, and an easy 

1822: estimate on the third piece. The reverse inequality obtains by using only

1823: the first piece. Thus, $A_d \to 1$, as desired.

1824: This value of $1 - \HM(\bfo - f)$,

1825: which gives a full domination upper bound on $\P^f$, should be

1826: compared to $\widehat f(\bfz) = 1/d$ (by symmetry), which is the

1827: probability of a 1 at a site.

1828: On the other hand, for no $d$ is the process uniformly insertion tolerant

1829: since $1/f$ is never integrable.

1830: We remark that for the full uniform spanning forest measure on $\Zd$

1831: (considering edges in all directions), we have change intolerance, meaning

1832: that $\P[\eta(\bfz) = 1 \mid \others] \in \{ 0, 1 \}$ a.s. This follows from a

1833: result of \BLPSusf, as explained by \ref b.HeicklenLyons/.

1834: \endprocl

1835: 

1836: \procl x.ust-delete

1837: Let $g(x)$ be as in \ref x.ustHx/, so that

1838: the edges of the uniform spanning tree in the plane that

1839: lie on the $x$-axis have the law $\P^g$.

1840: We have $\P^g \fullle \mu_p$ for $p := (1+\pi)/(1+2\pi) = 0.56865^+$, and this

1841: is optimal.

1842: This is because 

1843: \begineqalno

1844: \int_\T 1/(1-g) \,d\lambda 

1845: &=

1846: \int_\T \left(1 + \sin^2 \pi x  + \sin \pi x \sqrt{1 + \sin^2 \pi x }\right)

1847: \,d\lambda(x)

1848: \cr&=

1849: 3/2 + \int_\T \sin \pi x \sqrt{1 + \sin^2 \pi x } \,d\lambda(x)

1850: \,.

1851: \cr

1852: \endeqalno

1853: An antiderivative of this integrand is 

1854: $$

1855: -{1 \over \pi} \arctan\left({\cos\pi x \over \sqrt{1 + \sin^2\pi x}}\right)

1856: -{1 \over 2\pi} \cos\pi x\sqrt{1 + \sin^2\pi x}

1857: \,,

1858: $$

1859: whence the remaining integral is $1/2+1/\pi$. Therefore $\int_\T 1/(1-g)

1860: \,d\lambda = 2+1/\pi$ and $1 - \HM(\bfo-g) = (1+\pi)/(1+2\pi)$, as desired.

1861: Observe that $\int 1/g =\infty$ and so the process is not uniformly

1862: insertion tolerant.

1863: Similarly, for the edges lying on the $x$-axis of the uniform spanning tree in

1864: 3 dimensions, we have full domination by $\mu_p$ with $p := 0.37732^+$.

1865: Here, the law of these edges is $\P^h$, where 

1866: $$

1867: h(x) :=

1868: \int_{\T^2} f(x, y, z) \,d\lambda_2(y, z)

1869: =

1870: \int_\T {\sin^2 \pi x \over \sqrt{\left(\sin^2 \pi x + \sin^2 \pi y\right)

1871: \left(1 + \sin^2 \pi x + \sin^2 \pi y\right)}} \,d\lambda_1(y)

1872: \,.

1873: $$

1874: This integral has no simpler form, so to compute $p := 1 - \HM(\bfo - h)$, we

1875: calculated $h$ numerically and used the result to calculate the harmonic mean

1876: numerically. Since $\int 1/h =\infty$, as is easily checked,

1877: this process is not uniformly insertion tolerant.

1878: Similarly, one can check that the process of edges lying on the $x$-axis of the 

1879: uniform spanning tree in $d\ge 4$ dimensions

1880: is not uniformly insertion tolerant.

1881: \endprocl

1882: 

1883: \procl x.USTzz-dom

1884: Let $f$ be as in \ref x.USTzz/.

1885: It turns out that $\GM(f) = e^{-2\cat/\pi}/\sqrt2 = 1 - \GM(\bfo-f) =   

1886: 0.39467^+$. This gives strong domination inequalities.

1887: It is easy to see that $\HM(f) = \HM(\bfo - f) = 0$, so there are no

1888: nontrivial full domination inequalities.

1889: The interest of this function is that it describes the process of edges

1890: along a zig-zag path in the plane for the uniform spanning tree, as claimed

1891: in \ref x.USTzz/. 

1892: We now sketch how to prove this.

1893: The Transfer Current Theorem of \ref b.BurPem/ allows one to calculate, via

1894: determinants, the law for any set of possible edges of the uniform spanning

1895: tree.

1896: Thus, it is enough to verify that for the edges belonging to the zig-zag

1897: path, the matrix entries are those of the Toeplitz matrix associated to

1898: the Fourier coefficients given in \ref x.USTzz/.

1899: For even $k$, the values of $\widehat f(k)$ are given in

1900: \ref b.Lyons:book/.

1901: (These values imply the astonishing

1902: fact that the edges in the plane that lie along a diagonal, e.g., the horizontal

1903: edges with left endpoints $(n, n)$ ($n \in \Z$), are independent, i.e., have

1904: law $\mu_{1/2}$.)

1905: Thus, it remains to treat the case of odd $k$.

1906: A straightforward application of the Transfer Current Theorem gives that

1907: the matrix entry corresponding to the edges $e_0$ and $e_{2k-1}$ is 

1908: \begineqalno

1909: &\int {e(s) + e(t) - e(s+t) - 1 \over 4 - e(s) - e(-s) - e(t) - e(-t)}

1910: e(k s + k t) \,d\lambda_2(s, t)

1911: \cr&\qquad\qquad=

1912: \int {e(s) + e(x-s) - e(x) - 1 \over 4 - e(s) - e(-s) - e(x-s) - e(s-x)}

1913: e(k x) \,d\lambda_2(s, x)

1914: \,,

1915: \endeqalno

1916: where $e(x) := e^{2 \pi i x}$.

1917: Evaluate the integral in $s$ for fixed $x$ by a contour integral 

1918: $$

1919: {1 \over 2 \pi i} \oint {z + e(x) z^{-1} - e(x) - 1 \over 4 - z - z^{-1} -

1920: e(x) z^{-1} - e(-x) z} {d z \over z}

1921: $$

1922: over the contour $|z| = 1$.

1923: The integrand has poles inside the unit disc at $z = 0$ and 

1924: $$

1925: z = {2-\sqrt{4-|1+e(x)|^2} \over 1+e(-x)}

1926: \,.

1927: $$

1928: After use of the residue theorem and integrating in $x$, one obtains

1929: $\widehat f(2k-1)$, as desired.

1930: \endprocl

1931: 

1932: 

1933: 

1934: 

1935: We close this section by

1936: describing how our domination results can be interpreted in terms of

1937: prediction and interpolation questions for wide-sense stationary 

1938: processes. In view of \ref t.closest/ and the well-known correspondence

1939: between prediction and Szeg\H{o} infima, for $d=1$,

1940: $\P^f[\eta(\bfz) = 1 \mid \eta \restrict \{-1,-2,\dots\} \equiv 1]$

1941: is exactly the 

1942: mean squared error for the 

1943: best linear predictor of $Y_0$ given $\Seq{Y_n}_{n\le -1}$,

1944: where 

1945: $\Seq{Y_n}$ is a wide-sense stationary process with spectral density $f$.

1946: Similarly,

1947: $\P^f[\eta(\bfz) = 0 \mid \eta \restrict \{-1,-2,\dots\} \equiv 0]$

1948: is exactly the mean squared error for the best linear predictor

1949: of $Y_0$ given $\Seq{Y_n}_{n\le -1}$, where 

1950: $\Seq{Y_n}$ is now a wide-sense stationary process with spectral density 

1951: $\bfo-f$. This gives us a correspondence between strong domination and 

1952: prediction. An analogous correspondence holds between

1953: full domination and interpolation, where one instead looks at the 

1954: mean squared error for the 

1955: best linear predictor of $Y_0$ given $\Seq{Y_n}_{n\neq 0}$.  

1956: 

1957: \bsection{Entropy}{s.ent}

1958: 

1959: We assume the reader is familiar with the definition of the entropy

1960: $H(\mu)$ of a process $\mu$, as well as basic results concerning entropy (see

1961: \ref b.Walters/ and \ref b.KatzWeiss/).  

1962: Because of Ornstein's theorem (and its generalizations, see \ref

1963: b.KatzWeiss/, \ref b.Conze/, \ref b.Thouvenot/, and \ref b.OrnWeiss/)

1964: that entropy characterizes

1965: Bernoulli shifts up to isomorphism, the following question is particularly

1966: interesting:

1967: 

1968: \procl q.entropy

1969: What is $H(\P^f)$?

1970: \endprocl

1971: 

1972: We know the answer only in the trivial case where $f$ is a constant and in

1973: case $f$ or $\bfo - f$ is the reciprocal of a trigonometric polynomial of

1974: degree 1. In principle, as we shall see, one can also determine the entropy

1975: when $f$ or $\bfo - f$ is the reciprocal of any trigonometric polynomial,

1976: but the formula would be rather unwieldy.

1977: 

1978: In general, then, we shall discuss how to estimate the entropy of $\P^f$.

1979: The definition of entropy always provides upper bounds, due to

1980: subadditivity, so the harder bound is the lower bound.

1981: As we shall see, reciprocals of trigonometric polynomials can be used to

1982: get arbitrarily close lower (and upper) bounds.

1983: Unfortunately, that method is not practical for precise computation.

1984: Nevertheless, in many cases, we have

1985: another method that appears to work quite well.

1986: Indeed, our method seems to provide

1987: upper bounds that converge more quickly than does use of the definition.

1988: 

1989: \ref b.ShiTak:II/ proved that $H(\P^f) > 0$ for all $f \ne \bfz, \bfo$.

1990: Of course, this also follows immediately from our \ref t.Bern/.

1991: 

1992: Let

1993: $$

1994: H[p]:= H(\mu_p)= -p\log p-(1-p)\log(1-p)

1995: \,.

1996: $$

1997: \ref t.dom-d/ yields easy  

1998: lower bounds on the entropy of those processes $\P^f$ such that $f(\bfo-f)$

1999: has a strictly positive geometric mean. We shall obtain more refined lower

2000: bounds later in the one-dimensional case. It is easy to see that

2001: $$

2002: \mu_p \strle \mu \strle \mu_{1-p}

2003: \implies

2004: H(\mu)\ge H[p]

2005: \,.

2006: \label e.doment

2007: $$

2008: By \ref t.dom-d/ and \ref e.doment/, we deduce the

2009: following lower bound on entropy:

2010: 

2011: \procl p.entropy-d 

2012: For any measurable $f : \Td \to [0, 1]$, we have

2013: $$

2014: H(\P^f)\ge \min\Big\{H\big[\GM(f)\big],H\big[\GM(\bfo-f)\big]\Big\}

2015: \,.

2016: $$

2017: \endprocl 

2018: 

2019: \procl r.AMGM

2020: This can be compared to the trivial upper bound $H(\P^f) \le H\big[\widehat

2021: f(0)\big]$. Note that $\widehat f(0)$ is the arithmetic mean of $f$.

2022: \ref b.ShiTak:II/ also note this upper bound and provide two lower bounds:

2023: $H(\P^f) \ge H\big[\widehat f(0)\big]/2$ and

2024: $H(\P^f) \ge \int_{\Td} H[f(x)] \,d\lambda_d(x)$.

2025: \endprocl

2026: 

2027: \procl r.support

2028: It is interesting that \ref e.doment/

2029: is not true if $\strle$ is replaced by $\preccurlyeq$.

2030: For example, let $(X_{2i},X_{2i+1})$ be, independently for 

2031: different $i$, $(1,1)$ or $(0,0)$, each with probability $1/2$.

2032: If $\nu$ is the distribution of this process, then $\nu$ is not

2033: stationary, but $\mu:=\big(\nu +T(\nu)\big)/2$ is stationary, where $T$ is the

2034: shift.

2035: Next, $H(\mu)=\log 2/2 < H[1/\sqrt{2}]$, even though 

2036: $\mu_{1-1/\sqrt{2}} \preccurlyeq \mu \preccurlyeq \mu_{1/\sqrt{2}}$,

2037: as is easily verified.

2038: We also observe that this measure does not even have full support. 

2039: \endprocl

2040: 

2041: \procl x.ustH-ent

2042: Let $f$ be as in \ref x.ustH/, so that $\P^f$ is the law of the horizontal

2043: edges of a uniform spanning tree in the square lattice $\Z^2$.

2044: It is known that the entropy of the entire uniform spanning tree measure (both

2045: horizontal and vertical edges included) is \ref e.catalan/;

2046: see \ref b.BurPem/. 

2047: This can be compared to the bound on $H(\P^f)$ that

2048: \ref p.entropy-d/ provides, together with the calculations of

2049: \ref x.ustH-dom/, namely,

2050: $

2051: H(\P^f) \ge H[e^{-4\cat/\pi}] = 0.6203^+

2052: $.

2053: (The results are much worse if one uses the bounds of \ref b.ShiTak:II/

2054: reported in \ref r.AMGM/ above.)

2055: Direct calculation using cylinder events corresponding to a 4-by-4 block

2056: gives an upper bound for the entropy $H(\P^f) \le 0.68864$.

2057: Of course, the vertical edges of the uniform spanning tree measure have the

2058: same entropy as the horizontal edges.

2059: \endprocl

2060: 

2061: \procl x.ust_H-x

2062: Let $g(x)$ be as in \ref x.ustHx/, so that

2063: the edges of the uniform spanning tree in the plane that

2064: lie on the $x$-axis have the law $\P^g$.

2065: In view of \ref x.ustH_x-dom/,

2066: \ref p.entropy-d/ implies a lower

2067: bound of $H(\P^g) \ge H[\sqrt 2 - 1] \ge 0.67835$.

2068: Also, direct calculation using cylinder events of length 16 gives an upper

2069: bound $H(\P^g) \le 0.69034$.

2070: In \ref x.ustH-x-ent/, we shall obtain more refined bounds on the entropy.

2071: It is not hard to see that the horizontal edges associated to the $y$-axis

2072: have law $\P^{\bfo - g}$.

2073: \endprocl

2074: 

2075: Although not relevant for our determinantal probability measures

2076: (where we have seen that $\strle$ and $\preccurlyeq$ are equivalent),

2077: it is interesting to ask whether 

2078: $\mu_p \preccurlyeq \mu \preccurlyeq \mu_{1-p}$ with $p > 0$

2079: yields any lower bound on the entropy $H(\mu)$ for general $\mu$. 

2080: We first ask the question whether $\mu_p \preccurlyeq \mu$

2081: with $p > 0$ implies that

2082: $H(\mu)> 0$ provided $\mu\neq \delta_\bfo$. The answer to

2083: this question is affirmative

2084: since it is known (see \ref b.furst/ for $d=1$ and \ref b.GTW/ for $d > 1$)

2085: that zero-entropy processes are {\bf disjoint} from i.i.d.\

2086: processes (meaning that there are no stationary couplings of them other than

2087: independent couplings).

2088: The next result provides an explicit lower bound on the entropy

2089: for processes trapped between two i.i.d.\ processes.

2090: The proof was obtained jointly with Chris Hoffman.

2091: 

2092: \procl p.hoffman

2093: For any $d\ge 1$, if

2094: $\mu_p \preccurlyeq \mu \preccurlyeq \mu_{1-p}$ with $p >0$, then

2095: $$

2096: H(\mu)\ge \max{\{a_p,b_p\}} >0

2097: \,,

2098: $$

2099: where

2100: $$

2101: a_p:=

2102: (1-p)\log\Big({1\over  1-p}\Big)-{1-2p \over 2}\log\Big({1\over

2103: 1-2p}\Big)

2104: $$

2105: and

2106: $$

2107: b_p:=

2108: 2(1-p)\log\Big({1\over  1-p}\Big)-(1-2p)\log\Big({1\over

2109: 1-2p}\Big)-(1-2p)\log 2.

2110: $$

2111: \endprocl

2112: 

2113: \procl r.log2

2114: Observe that $b_p$ approaches $\log 2$ as $p\to 1/2$.

2115: \endprocl

2116: 

2117: \proof

2118: Let $X$, $Y$ and $Z$ denote processes with respective distributions 

2119: $\mu_p,\mu$ and $\mu_{1-p}$. We construct a joining (stationary coupling)

2120: of all three processes as follows. Consider any joining $m_1$ of $X$ and $Y$

2121: with $X_i\le Y_i$ a.s.\ 

2122: and any joining $m_2$ of $Y$ and $Z$ with $Y_i\le Z_i$ a.s. We now pick a 

2123: realization for $Y$ according to $\mu$ and then choose $X$ and $Z$ 

2124: (conditionally) independently using $m_1$ and $m_2$ respectively (this

2125: is called the fibered product of $m_1$ and $m_2$ over $Y$). This gives

2126: us a joining $(X',Y',Z')$ of $X$, $Y$ and $Z$.

2127: We may now assume that $(X, Y, Z) = (X',Y', Z')$.

2128: 

2129: We now use standard facts about entropy.

2130: First, $H(X,Z)\ge H(X,Y,Z)- H(Y)$. We next note, using

2131: the conditional independence of $X$ and $Z$ given $Y$, that

2132: \begineqalno

2133: H(X,Y,Z)

2134: &= H(Y) + H(X,Z\mid Y)=

2135: H(Y) + H(X\mid Y)+H(Z\mid Y)

2136: \cr&\ge  H(X)+H(Z)-H(Y)\,. 

2137: \cr

2138: \endeqalno

2139: Hence

2140: $$

2141: H(X,Z)\ge H(X)+H(Z)-2H(Y)=2H[p]-2H(Y)\,.

2142: $$

2143: Since $X\le Z$, the one-dimensional marginal of $(X,Z)$ necessarily

2144: has 3 atoms of weights $p,1-2p$ and $p$. It follows that

2145: $$

2146: H(X,Z)\le 2p\log\Big({1\over p}\Big)+(1-2p)\log\Big({1\over 1-2p}\Big)\,.

2147: $$

2148: Combining this with the previous inequality shows that 

2149: $H(\mu)\ge a_p$. 

2150: 

2151: We modify the above proof to show that $b_p$ is also

2152: a lower bound. Rather than using $H(X,Z)\ge H(X,Y,Z)- H(Y)$, we

2153: use $H(X,Z)= H(X,Y,Z)- H(Y\mid X,Z)$. The earlier argument then gives

2154: $$

2155: H(X,Z)\ge 2H[p]-H(Y)- H(Y\mid X,Z)\,.

2156: $$

2157: Using the same upper bound on $H(X,Z)$ as before and the trivial upper bound

2158: for $H(Y\mid X,Z)$ of $(1-2p)\log 2$ (obtained by noting that $Y_i$ is

2159: determined by $X_i$ and $Z_i$ when $X_i=Z_i$) yields the lower bound of $b_p$.

2160: 

2161: The easiest way to check the strict positivity of $a_p$ is to

2162: observe that

2163: $$

2164: 2H[p] > 2p\log\Big({1\over p}\Big)+(1-2p)\log\Big({1\over 1-2p}\Big)

2165: \,.

2166: $$

2167: This, in turn, follows from the fact that the left-hand side is the entropy of

2168: the independent joining $\mu_p \times \mu_{1-p}$, while the right-hand side is

2169: the entropy of the monotonic coordinatewise-independent joining of $\mu_p$

2170: and $\mu_{1-p}$, whose existence follows from the inequality 

2171: $p \le 1-p$.

2172: \Qed

2173: 

2174: There are few cases where we know the entropy $H(\P^f)$ exactly.

2175: Besides the case when $f$ is constant, we can calculate the entropy when $f$

2176: is the reciprocal of a trigonometric polynomial of degree 1.

2177: We shall illustrate this in the simplest case, \ref x.renewal/.

2178: Thus, let $0 < a < 1$, $d = 1$, and $f(x) := (1 - a)^2/|e^{2\pi i x} - a|^2$.

2179: Recall that that for any invariant measure $\mu$ in one dimension, we have

2180: $$

2181: H(\mu) =

2182: \int H\big[\mu[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]\big]

2183: \,d\mu(\eta_{-1}, \eta_{-2}, \ldots)

2184: \,.

2185: \label e.ent-past

2186: $$

2187: Now by \ref e.renewal-distribution/, it is easy to check that

2188: $$

2189: P^f[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots] =

2190: {(1 - a)^2 N \over N - (N - 1) a}

2191: \,,

2192: $$

2193: where $N := \min \{ k\ge 1 \st \eta_{-k} = 1 \} $.

2194: In addition, we may easily show that for $n\ge 1$,

2195: $$

2196: \P^f[N = n] = {(1-a)(n - (n-1)a) a^{n-1} \over 1+a}

2197: \,.

2198: $$

2199: Therefore 

2200: \begineqalno

2201: H(\P^f) 

2202: &=

2203: \sum_{n=1}^\infty 

2204: {(1-a)(n - (n-1)a) a^{n-1} \over 1+a}

2205: \Big(

2206: - {(1 - a)^2 n \over n - (n - 1) a} \log {(1 - a)^2 n \over n - (n - 1) a}

2207: \cr&\hskip1in

2208: - {n - (n - 1) a - (1 - a)^2 n \over n - (n - 1) a} 

2209: \log {n - (n - 1) a - (1 - a)^2 n \over n - (n - 1) a}

2210: \Big)

2211: \cr&=

2212: {1-a \over 1+a}

2213: \sum_{n=1}^\infty a^{n-1}

2214: \Big(

2215: - (1 - a)^2 n \log [(1 - a)^2 n]

2216: \cr&\hskip1.3in

2217: - [a(1 - a) n + a] \log [a(1 - a) n + a]

2218: \cr&\hskip1.3in

2219: + [(1-a) n + a] \log [(1-a) n + a]

2220: \Big)

2221: \,.

2222: \cr

2223: \endeqalno

2224: 

2225: In principle, one can also write an infinite series of positive terms for the

2226: entropy of $\P^f$ when $f$ is equal to the reciprocal of any trigonometric

2227: polynomial, since, by \ref p.regenerate/, the process $\P^f$ is then

2228: a regenerative process.

2229: Of course, the answer will be much more unwieldy than the above formula.

2230: However, it can be used to get arbitrarily good lower bounds on the entropy of

2231: any process $\P^f$, in theory.

2232: To see this, 

2233: we use the following lemma, which is more or less Lemma 5.10 in 

2234: \ref b.Rudolph/. As our proof is somewhat shorter and gives a more precise

2235: bound, we include it. See also \ref b.BurPem/, Lemma 6.2, for a similar

2236: proof.

2237: 

2238: \procl l.rudolph

2239: For any stationary processes $\mu$ and $\nu$, we have

2240: $|H(\mu)-H(\nu)| \le H[\dbar(\mu,\nu)]$.

2241: \endprocl

2242: 

2243: \proof

2244: Consider a

2245: stationary process $(X,Y)=\Seq{(X_i,Y_i)}_{i\in\Z^d}$,

2246: where $X=\Seq{X_i}_{i\in\Z^d}$ has distribution $\mu$,

2247: $Y=\Seq{Y_i}_{i\in\Z^d}$ has distribution $\nu$, and

2248: $P(X_0\neq Y_0)=\dbar(\mu,\nu)$. 

2249: Let $Z_i := X_i + Y_i$ mod 2. Then 

2250: $$

2251: H(\mu) \le H(X, Z) = H(Y, Z) \le H(\nu) + H(Z)        

2252: \le H(\nu) + H[\dbar(\mu, \nu)]

2253: \,.

2254: $$

2255: Since the same holds with $\mu$ and $\nu$ switched, the result follows.

2256: \Qed

2257: 

2258: 

2259: 

2260: Thus, given any measurable $f : \T \to [0, 1]$ and any $\epsilon > 0$, define

2261: $\delta$ to be the smallest positive number 

2262: so that $H[\delta] = \epsilon$.

2263: Then choose a trigonometric polynomial $T \ge 1$ such that $\int_\T |T^{-1} -

2264: \max(f,\delta/2)| \,d\lambda_1 <\delta/2$.

2265: We can calculate $H(\P^{1/T})$ as closely as desired.

2266: Since $\dbar(\P^{1/T}, \P^f) < \delta$ by \ref p.L1givesdbar/,

2267: it follows by \ref l.rudolph/ that $|H(\P^f) - H(\P^{1/T})| < \epsilon$.

2268: 

2269: Unfortunately,

2270: this method of calculation is hopeless in practice since when $T$ has a high

2271: degree, it will take a very long time to see a renewal, which means that

2272: cylinder events of great length will be needed for the estimation.

2273: When a cylinder event of length $n$ is used, one must calculate $2^n$

2274: probabilities. Thus, this is completely impractical.

2275: 

2276: However, we now exhibit an alternative method that works extremely well in

2277: practice, although we cannot prove that it works well {\it a priori}.

2278: We begin with two simple examples,

2279: $$

2280: f(x) := \I{[0, 1/2]}(x)

2281: \label e.deff

2282: $$

2283: (which also occurred in \ref x.domino/)

2284: and 

2285: $$

2286: g(x) := \sin^2 \pi x = (1 - \cos 2\pi x)/2

2287: \,.

2288: \label e.defg

2289: $$

2290: Since $\widehat f(k) = \widehat g(k) = 0$ for all even $k \ne 0$,

2291: both $\P^{f}$ and $\P^{g}$ have the property that looking at only the even

2292: coordinates, we see independent fair coin flips, i.e., $\mu_{1/2}$.

2293: Therefore the entropy of both processes is at least $(1/2) \log 2$.

2294: In either case, we know of no direct method to prove that strict inequality

2295: holds, but it does.

2296: We first show this for $\P^g$. 

2297: In \ref x.ustH_x-dom/ we showed that $\GM(g) = 1/4$.

2298: By symmetry, $\GM(1-g) = 1/4$.

2299: Since $H[1/4] = 0.56^+ > (1/2) \log 2 = 0.35^-$, we obtain by

2300: \ref p.entropy-d/ that $H(\P^g) > (1/2) \log 2$.

2301: 

2302: In order to show that $H(\P^f) > (1/2) \log 2$ for $f$ as in \ref

2303: e.deff/,

2304: we need a method to

2305: obtain more refined entropy bounds. We illustrate such a method

2306: beginning with this simple function $g$.

2307: While we shall explain afterwards a method to obtain results for more general

2308: functions for the case $d=1$, this first proof contains the essential

2309: idea of this method while at the same time relying on a more elementary 

2310: calculation, and therefore, we feel, is worth including.

2311: 

2312: \procl p.entropyg

2313: With $g$ as in \ref e.defg/, we have 

2314: $$

2315: 0.63^- = {3 \over 8} H[1/4] + {5 \over 8} H[11/28]

2316: \le H(\P^g) \le

2317: {3 \over 8} H[7/20] + {5 \over 8} H[5/12] = 0.67^-

2318: \,.

2319: $$

2320: \endprocl

2321: 

2322: \procl l.GMpert

2323: Let $h : \T \to [0, 1]$ be a trigonometric polynomial of degree at most 1,

2324: i.e., $\widehat h(k) = 0$ for $|k| \ge 2$.

2325: Fix $n > 0$ and $A \subseteq \{ 1, 2, \ldots, n \}$.

2326: For $C \subseteq \{ 1, 2, \ldots, n \}$, let $\lambda(C)$ denote the sequence

2327: of lengths of consecutive intervals in $\{ 1, 2, \ldots, n \} \setminus C$.

2328: Then 

2329: $$

2330: \P^h[\eta \restrict A \equiv 0,\, \eta \restrict (\{ 1, 2, \ldots, n \}

2331: \setminus A) \equiv 1]

2332: =

2333: \sum_{C \subseteq A} (-1)^{|A \setminus C|} \prod_{i \in\lambda(C)} D_{i-1}(h)

2334: \,.

2335: $$

2336: \endprocl

2337: 

2338: \proof

2339: Whenever two sets $F_1, F_2 \subset \Z$ neither overlap nor come within

2340: distance 1 of each other, the configurations on $F_1$ and $F_2$ are

2341: $\P^h$-independent because $\widehat h(k) = 0$ for $|k| \ge 2$.

2342: Therefore $\P^h[\eta \restrict (\{ 1, 2, \ldots, n \} \setminus C) \equiv 1] =

2343: \prod_{i \in\lambda(C)} D_{i-1}(h)$. Now the desired formula follows from the

2344: inclusion-exclusion principle.

2345: \Qed

2346: 

2347: \proofof p.entropyg

2348: Write $d_n := D_{n-1}(g)$.

2349: Recall that 

2350: $$

2351: H(\P^g) =

2352: \int H\big[\P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]\big]

2353: \,d\P^g(\eta_{-1}, \eta_{-2}, \ldots)

2354: \,.

2355: \label e.ent-past

2356: $$

2357: 

2358: By the negative association property of $\P^g[\;\cdot \mid \eta_{-1} =

2359: \eta_{-2} = 1]$, we have that on the event $ \{ \eta_{-1} = \eta_{-2} = 1 \}$,

2360: \begineqalno

2361: \P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]

2362: &\ge

2363: \lim_{n \to\infty} \P^g[\eta_0 = 1 \mid \eta_{-1} = \eta_{-2}= \cdots = \eta_{-n} = 1]

2364: \cr&=

2365: \lim_{n \to\infty} {\P^g[\eta_0 = \eta_{-1} = \eta_{-2}= \cdots = \eta_{-n} = 1] \over

2366: \P^g[\eta_{-1} = \eta_{-2}= \cdots = \eta_{-n} = 1]}

2367: \cr&=

2368: \lim_{n \to\infty} d_{n+1}/d_n

2369: =

2370: \GM(g)

2371: =

2372: 1/4

2373: \,.

2374: \endeqalno

2375: Likewise, on the same event, we have 

2376: \begineqalno

2377: \P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]

2378: &\le

2379: \lim_{n \to\infty} \P^g[\eta_0 = 1 \mid \eta_{-1} = \eta_{-2}= 1,\, \eta_{-3} = 

2380: \cdots = \eta_{-n} = 0]

2381: \cr&=

2382: \lim_{n \to\infty} \P^g[\eta_0 = 0 \mid \eta_{-1} = \eta_{-2}= 0,\, \eta_{-3} = 

2383: \cdots = \eta_{-n} = 1]

2384: \cr&\hskip1in\hbox{[by symmetry]}

2385: \cr&=

2386: 1 - \lim_{n \to\infty} \P^g[\eta_0 = 1 \mid \eta_{-1} = \eta_{-2}= 0,\, \eta_{-3} = 

2387: \cdots = \eta_{-n} = 1]

2388: \cr&=

2389: 1 - \lim_{n \to\infty} {\P^g[\eta_{-1} = \eta_{-2}= 0,\, \eta_0 = \eta_{-3} = 

2390: \cdots = \eta_{-n} = 1] \over

2391: \P^g[\eta_{-1} = \eta_{-2}= 0,\, \eta_{-3} = \cdots = \eta_{-n} = 1]}

2392: \cr&=

2393: 1 - \lim_{n \to\infty} 

2394: {d_1 d_{n-2} - d_2 d_{n-2} - d_1 d_{n-1} + d_{n+1}

2395: \over

2396: d_{n-2} - d_1 d_{n-2} - d_{n-1} + d_{n}}

2397: \cr&\hskip1in\hbox{[by \ref l.GMpert/]}

2398: \cr&=

2399: 1 - 

2400: {d_1 - d_2 - d_1 \GM(g) + \GM(g)^3

2401: \over

2402: 1 - d_1 - \GM(g) + \GM(g)^2}

2403: \cr&=

2404: 7/20

2405: \endeqalno

2406: since $d_1 = 1/2$ and $d_2 = 3/16$.

2407: We may conclude that on the event $ \{ \eta_{-1} = \eta_{-2} = 1 \}$, we have 

2408: $$

2409: H\big[\P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]\big]

2410:  \in \big[H[1/4], H[7/20]\big]

2411: \,.

2412: $$

2413: By symmetry, the same holds on the event $ \{ \eta_{-1} = \eta_{-2} = 0 \}$.

2414: 

2415: Similarly, we have that on the event $ \{ \eta_{-1} = 1,\, \eta_{-2} = 0 \}$,

2416: \begineqalno

2417: \P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]

2418: &\ge

2419: \lim_{n \to\infty} \P^g[\eta_0 = 1 \mid \eta_{-2} = 0,\,

2420: \eta_{-1} = \eta_{-3}= \cdots = \eta_{-n} = 1]

2421: \cr&=

2422: \lim_{n \to\infty} {\P^g[\eta_{-2} = 0,\,

2423: \eta_0 = \eta_{-1} = \eta_{-3}= \cdots = \eta_{-n} = 1] \over

2424: \P^g[\eta_{-2} = 0,\, \eta_{-1} = \eta_{-3}= \cdots = \eta_{-n} = 1]}

2425: \cr&=

2426: \lim_{n \to\infty}

2427: {d_2 d_{n-2} - d_{n+1}

2428: \over

2429: d_1 d_{n-2} - d_n}

2430: \cr&=

2431: {d_2 - \GM(g)^3

2432: \over

2433: d_1 - \GM(g)^2}

2434: \cr&=

2435: 11/28

2436: \,.

2437: \endeqalno

2438: Likewise, on the same event, we have 

2439: \begineqalno

2440: \P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]

2441: &\le

2442: \lim_{n \to\infty} \P^g[\eta_0 = 1 \mid \eta_{-1} = 1,\, \eta_{-2}= \eta_{-3} = 

2443: \cdots = \eta_{-n} = 0]

2444: \cr&=

2445: \lim_{n \to\infty} \P^g[\eta_0 = 0 \mid \eta_{-1} = 0,\, \eta_{-2}= \eta_{-3} = 

2446: \cdots = \eta_{-n} = 1]

2447: \cr&=

2448: 1 - \lim_{n \to\infty} \P^g[\eta_0 = 1 \mid \eta_{-1} = 0,\, \eta_{-2}= \eta_{-3} = 

2449: \cdots = \eta_{-n} = 1]

2450: \cr&=

2451: 1 - \lim_{n \to\infty} {\P^g[\eta_{-1} = 0,\, \eta_0 = \eta_{-2} = 

2452: \cdots = \eta_{-n} = 1] \over

2453: \P^g[\eta_{-1} = 0,\, \eta_{-2} = \cdots = \eta_{-n} = 1]}

2454: \cr&=

2455: 1 - \lim_{n \to\infty} {d_1 d_{n-1} - d_{n+1} \over d_{n-1} - d_n}

2456: \cr&=

2457: 1 - {d_1 - \GM(g)^2 \over 1 - \GM(g)}

2458: \cr&=

2459: 5/12

2460: \,.

2461: \endeqalno

2462: We may conclude that on the event $ \{ \eta_{-1} = 1,\, \eta_{-2} = 0 \}$, we have 

2463: $$

2464: H\big[\P^g[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]\big]

2465:  \in \big[H[11/28], H[5/12]\big]

2466: \,.

2467: $$

2468: By symmetry, the same holds on the event $ \{ \eta_{-1} = 0,\, \eta_{-2} =

2469: 1 \}$.

2470: 

2471: Putting all these bounds into \ref e.ent-past/ gives the claimed bounds on the

2472: entropy of $\P^g$.

2473: \Qed

2474: 

2475: 

2476: For more general functions, we can get lower bounds on the entropy by using

2477: the asymptotics of \ref b.BumpDiaconis/ or of \ref b.TW:like/,

2478: which serve as replacements for our use of the special \ref l.GMpert/.

2479: In fact, we use an extension of the formula of \ref b.TW:like/ due to

2480: \ref b.Lyons:szego/.

2481: Denote by $\nu_{f}$ the measure on $2^\N$ obtained by the

2482: limiting condition $\lim_{n \to\infty} \P^f[\;\cdot \mid \eta_{-1} = \eta_{-2}

2483: = \cdots = \eta_{-n} = 1]$, which exists by \ref p.JNRD/.

2484: If $\GM(f) > 0$, define

2485: $$

2486: \Phi_f(z) := \exp {1 \over 2} \int_{\T} {e^{2\pi i t} + z

2487: \over e^{2\pi i t} - z} \log f(t) \,d\lambda_1(t)

2488: \label e.defPhi_f

2489: $$

2490: for $|z| < 1$.

2491: The {\bf outer} function

2492: $$

2493: \varphi_f(t) := \lim_{r \uparrow 1} \Phi_f(r e^{2 \pi i t})

2494: \label e.defvarphi_f

2495: $$

2496: exists for $\lambda_1$-a.e.\ $t \in \T$ and satisfies $|\varphi_f|^2 = f$

2497: $\lambda_1$-a.e. 

2498: This limit also holds in $L^1(\T)$.

2499: See \ref b.Rudin:RCA/, Theorem 17.11, p.~340, and Theorem 17.16, p.~343.

2500: As is easily verified,

2501: $$

2502: \log \Phi_f(z) = \widehat F(0) + 2 \sum_{k=1}^\infty \widehat F(k) z^k

2503: \label e.preexp

2504: $$

2505: when $F := (1/2) \log f$.

2506: The mean value theorem for analytic functions and the $L^1(\T)$

2507: convergence above show that

2508: $$

2509: \widehat {\varphi_f}(0) = \Phi_f(0) = \sqrt{\GM(f)}

2510: \,.

2511: $$

2512: Define $\Phi_f := \varphi_f := \bfz$ if $\GM(f) = 0$.

2513: An analytic trigonometric polynomial is (a constant times) an outer

2514: function iff its extension to the unit disc has no zeroes in the open disc.

2515: [On the one hand, the extension of 

2516: no outer function has any zeroes in the open disc since it

2517: is an exponential.  On the other hand, by factoring a polynomial, one

2518: has to check merely that $z \mapsto z-\zeta$ is a constant times

2519: an outer function for

2520: $|\zeta| \ge 1$. Indeed, $\Phi_f(z) = (-|\zeta|/\zeta)(z -\zeta)$

2521: for $f(t) := |e^{2\pi i t} -\zeta|^2$, as can be seen by the Poisson integral

2522: formula for $|\zeta|>1$ (\ref b.Rudin:RCA/, Theorem 11.9, p.~235) and by a

2523: limiting procedure for $|\zeta| = 1$.]

2524: The results of \ref b.Lyons:szego/ show that \ref t.condonB/ reduces to

2525: the following formula:

2526: 

2527: 

2528: \procl t.szego

2529: For any measurable $f : \T \to [0, 1]$, the measure $\nu_{f}$ is equal to

2530: the determinantal probability measure corresponding to the positive

2531: contraction on $\ell^2(\N)$ whose $(j, k)$-matrix entry is 

2532: $$

2533: \sum_{l=0}^{j \wedge k} \overline{\widehat{\varphi_f}(j-l)}

2534: \widehat{\varphi_f}(k-l)

2535: \,.

2536: $$

2537: \endprocl

2538: 

2539: This can be substituted in the appropriate places in the proof of \ref

2540: p.entropyg/.

2541: For example, on the event $ \{ \eta_{-1} = \eta_{-2} = 0 \}$, one has that 

2542: \begineqalno

2543: \P^f[\eta_0 = 1 \mid \eta_{-1}, \eta_{-2}, \ldots]

2544: &\ge

2545: \lim_{n \to\infty}

2546: \P^f[\eta_0 = 1 \mid \eta_{-1} = \eta_{-2}= 0,\, \eta_{-3} = \eta_{-4} =

2547: \cdots = \eta_{-n} = 1]

2548: \cr&=

2549: \nu_{f}[\eta_2=1, \eta_1=\eta_0=0]/ \nu_{f}[\eta_1=\eta_0=0]

2550: \,.

2551: \endeqalno

2552: One can use $\nu_{\bfo-f}$ in a similar fashion for estimating such

2553: probabilities above by $\eta$ that terminate in repeating 0s, rather than in

2554: repeating 1s.

2555: For example, if 

2556: $$

2557: f := |1+\ee_{1} + \ee_{2}|^2/9

2558: \,,

2559: $$

2560: then $\varphi_f = (1+\ee_{1} + \ee_{2})/3$ since the polynomial

2561: $(1+z+z^2)/3$ has no zeroes in the open unit disc.

2562: Similarly, one can show that 

2563: $$

2564: \varphi_{\bfo-f} = {\sqrt 6 + \sqrt 2 \over 6} - {\sqrt 2 \over 3} \ee_{1} 

2565: + {\sqrt 2 - \sqrt 6 \over 6} \ee_{2}

2566: \,.

2567: $$

2568: With these and \ref t.szego/, one can show by following the method of proof of

2569: \ref p.entropyg/ that 

2570: $$

2571: 0.53 \le H(\P^f) \le 0.61

2572: \,.

2573: $$

2574: This method of estimation is relatively easy to program on a computer;

2575: decomposing by cylinder events of length 8 instead of the length 2 used in the

2576: proof of \ref p.entropyg/ and using Mathematica, we find that 

2577: $$

2578: 0.601992  \le H(\P^f) \le 0.602433

2579: \,.

2580: $$

2581: % We also find for $g$ above the range $[0.659062, 0.659087]$.

2582: Similarly, using cylinder events of length 15, we obtain that 

2583: $$

2584: 0.65907716

2585: \le H(\P^g) \le

2586:   0.65907733

2587: \label e.entropy-g

2588: $$

2589: for the function $g$ in \ref e.defg/.

2590: 

2591: We now return to the problem of showing that

2592: $H(\P^f) > (1/2) \log 2$ for $f$ as in \ref e.deff/.

2593: 

2594: \procl p.entrindic

2595: If $f := \I{[0, 1/2]}$, then $H(\P^f) > (1/2) \log 2$.

2596: \endprocl

2597: 

2598: \proof

2599: Let $\widetilde{f}:= .99\I{[0, 1/2]} +.01\I{(1/2,1)}$.

2600: \ref p.L1givesdbar/ tells us that

2601: $\dbar(\P^f, \P^{\widetilde{f}}) \le .01$, and hence from

2602: \ref l.rudolph/,

2603: $H(\P^f) \ge H(\P^{\widetilde{f}})- H[.01]$. 

2604: Now $H[.01] < 0.0561$.

2605: Write $\delta := .01$ and

2606: $L := \log (\delta ^{-1} - 1) = \log 99$.

2607: A simple integration shows that 

2608: $$

2609: \skew6\widehat {\widetilde f}(k) =

2610: \cases{(1/2) &if $k = 0$,\cr

2611: \displaystyle{(-1)^k - 1 \over 2k\pi} i (1 - 2\delta)

2612: &if $k \ne 0$\cr}

2613: $$

2614: (where $i := \sqrt{-1}$).

2615: To find $\varphi_{\widetilde f}$, we proceed as follows.

2616: Let $F := (1/2) \log \widetilde f$.

2617: Then

2618: $$

2619: \widehat F(k) =

2620: \cases{(1/4) \log \delta(1-\delta)  &if $k = 0$,\cr

2621: \displaystyle{(-1)^k - 1 \over 4k\pi} i L

2622: &if $k \ne 0$.\cr}

2623: $$

2624: By using \ref e.preexp/ and exponentiating, we obtain

2625: $$

2626: \Phi_{\widetilde f}(z)

2627: =

2628: [\delta(1-\delta)]^{1/4}

2629: \left(

2630: 1

2631: -

2632: {i \over \pi} L z

2633: -

2634: {1 \over 2\pi^2} L^2 z^2

2635: -

2636: {i \over 6\pi^3} (2\pi^2 - L^2) L z^3

2637: +

2638: \cdots

2639: \right)

2640: \,.

2641: $$

2642: Since the coefficients of this power series are the Fourier coefficients of

2643: $\varphi_{\widetilde f}$,

2644: we have a procedure to compute the Fourier coefficients

2645: $\varphi_{\widetilde f}$.

2646: Since $\widetilde f(1-x) = 1-\widetilde f(x)$, we have

2647: $\overline{\Phi_{\widetilde f}(\overline z)} =

2648: \Phi_{\bfo-\widetilde f}(z)$. 

2649: Thus, the $k$th Fourier coefficient of $\varphi_{\bfo - \widetilde f}$ equals

2650: the complex conjugate of the $k$th Fourier coefficient of $\varphi_{\widetilde

2651: f}$.

2652: Using the method of \ref p.entropyg/ with cylinder events of length 3,

2653: which requires knowledge of Fourier coefficients only for $|k| \le 2$, we

2654: obtain that $H(\P^{\widetilde f}) \ge 0.4105$.

2655: Therefore $H(\P^f) > 0.4105 - 0.0561 = 0.3544 > (1/2) \log 2$.

2656: (If we use cylinder events of length 12, then we obtain the bound $H(\P^f) >

2657: 0.4442$. However, we believe that the true entropy is significantly larger

2658: still.)

2659: \Qed

2660: 

2661: 

2662: \procl x.depends

2663: We can also use these bounds to prove that $H(\P^f)$ does {\it not\/}

2664: depend {\it only\/} on the distribution of $f$.

2665: For example, consider the function 

2666: $$

2667: f(t) := \sin^2(\pi t/2)

2668: $$

2669: on $[0, 1]$, which has the same distribution as the function $g$ of \ref

2670: e.defg/. 

2671: An elementary integration shows that 

2672: $$

2673: \widehat f(k) = \cases{1/2 &if $k=0$,\cr

2674: \displaystyle {2k i \over (2k-1)(2k+1) \pi} &if $k \ne 0$\cr}

2675: $$

2676: (where $i := \sqrt{-1}$).

2677: To find $\varphi_f$, we proceed as follows.

2678: Write $F := (1/2) \log f$.

2679: The real parts of the integrals giving $\widehat F(k)$ can be found in

2680: standard tables, while the imaginary parts are derived in \ref b.LPR/,

2681: giving

2682: $$

2683: \widehat F(k) =

2684: \cases{-\log 2  &if $k=0$,\cr

2685:        \displaystyle -{1 \over 4k} + {i \over k \pi} \sum_{j=1}^k {1 \over 2j-1}

2686:                 &if $k > 0$.\cr}

2687: $$

2688: %The list of the Fourier coefficients $\widehat F(k)$ for $k = 0, 1,

2689: %\ldots, 10$ is

2690: %\begineqalno

2691: %-\log 2,\,

2692: %&-1/4 + i/\pi,\, -1/8 + 2 i/3\pi,\, -1/12 + 23 i/45\pi,\, 

2693: %\cr& -1/16 + 44 i/105\pi,\, -1/20 + 563 i/1575\pi,\, 

2694: %\cr& -1/24 + 3254 i/10395\pi,\, -1/28 + 88069 i/315315\pi,\, 

2695: %\cr& -1/32 + 11384 i/45045\pi,\, -1/36 + 1593269 i/6891885\pi,\, 

2696: %\cr& -1/40 + 15518938 i/72747675\pi

2697: %\,.

2698: %\cr

2699: %\endeqalno

2700: As in the proof of \ref p.entrindic/, we compute

2701: \begineqalno

2702: \Phi_f(z) =

2703: 1/2 &+( -1/4 + i/\pi) z +(-1/16 + i/6\pi - 1/\pi^2) z^2

2704: \cr&  -(1/32 - 19 i/360 \pi + 5/6 \pi^2 + 2 i/3\pi^3) z^3

2705: \cr&  +(-5/256 + 89 i/5040\pi - 27/40 \pi^2 - i/\pi^3 + 1/3 \pi^4) z^4

2706:   + \cdots

2707: \,,

2708: \endeqalno

2709: whose coefficients are the Fourier coefficients of $\varphi_f$.

2710: Since $1-\sin^2(\pi t/2) = \sin^2(\pi (1-t)/2)$, it follows from the

2711: definition \ref e.defPhi_f/ that $\overline{\Phi_{f}(\overline z)} =

2712: \Phi_{\bfo-f}(z)$. Since the power series coefficients of $\Phi_{\bfo - f}$

2713: are the Fourier coefficients of $\varphi_{\bfo - f}$, we obtain that the $k$th

2714: Fourier coefficient of $\varphi_{\bfo - f}$ equals the complex conjugate of

2715: the $k$th Fourier coefficient of $\varphi_f$.

2716: Using the method of \ref p.entropyg/ with cylinder events of length 8,

2717: which requires knowledge of Fourier coefficients only for $|k| \le 7$, we

2718: obtain that 

2719: $$

2720: 0.659648

2721: \le H(\P^f) \le

2722: 0.684021

2723: \,.

2724: $$

2725: Comparing with \ref e.entropy-g/, we see that $H(\P^f) > H(\P^g)$.

2726: \endprocl

2727: 

2728: Note that even when neither $\widehat f$ nor $\widehat{\varphi_f}$ can be

2729: found explicitly, they can always be found by numerical integration and

2730: exponentiation, and then one can follow the procedure we have used in 

2731: this last example. Indeed, that will be done for part of our final example.

2732: 

2733: \procl x.ustH-x-ent

2734: As in \ref x.ust_H-x/, let 

2735: $$

2736: g(x) := 

2737: {\sin \pi x \over \sqrt{1+\sin^2 \pi x}}

2738: \,.

2739: $$

2740: Then the edges of the uniform spanning tree measure in the plane that lie on

2741: the $x$-axis have the law $\P^g$.

2742: Write $g_1(x) := \sin^2 \pi x$.

2743: The calculations in \ref x.ustHx/ show that $\Phi_{g_1}(z) = (1-z)/2$ and

2744: $\Phi_{\bfo+g_1}(z) = [\sqrt 2 + 1 - (\sqrt 2 - 1)z]/2$, whence

2745: $$

2746: \Phi_g(z)

2747: =

2748: \sqrt{\Phi_{g^2}(z)}

2749: =

2750: \sqrt{{\Phi_{g_1} \over \Phi_{\bfo+g_1}}}

2751: =

2752: \sqrt{1-z \over \sqrt 2 + 1 - (\sqrt 2 - 1)z}

2753: \,.

2754: $$

2755: Expansion of $\Phi_g(z)$ in a Maclaurin series gives 

2756: $$

2757: \frac{1}{({1 + {\sqrt{2}}})^{1/2}} - \frac{z}{( 1 + {\sqrt{2}} )

2758: ^{\nonfrac{3}{2}}} - \frac{(-1+2\sqrt{2})\,z^2}{2\,(1 +\sqrt{2})^{5/2}} -

2759: \frac{( 5 - 2\,{\sqrt{2}} ) \,z^3}{2\,( 1 + {\sqrt{2}}

2760: ) ^{\nonfrac{7}{2}}} - \frac{( -27 + 28\,{\sqrt{2}} )

2761: \,z^4}{8\,( 1 + {\sqrt{2}} ) ^{\nonfrac{9}{2}}} +

2762: \cdots

2763: \,,

2764: $$

2765: which tells us $\widehat{\varphi_g}$.

2766: The transfer currents, which can be calculated by the method in \ref

2767: b.BurPem/, tell us $\widehat g$; for example, $\widehat g(k)$ for $k=0, 1, 2,

2768: 3, 4$ is 

2769: $$

2770: \frac{1}{2}\,,\qquad \frac{1}{2} - \frac{2}{\pi

2771: }\,,\qquad\frac{5}{2} - \frac{8}{\pi }\,,\qquad\frac{25}{2} - \frac{118}{3\,\pi

2772: }\,,\qquad

2773: \frac{129}{2} - \frac{608}{3\,\pi }

2774: \,.

2775: $$

2776: We use numerical integration to find $\widehat{\varphi_{\bfo-g}}$.

2777: Then if we use cylinder events of length 8, we find that 

2778: $$

2779: 0.69005 \le H(\P^g) \le 0.69013

2780: \,.

2781: $$

2782: It is interesting how close this is to $\log 2 = 0.69315^-$.

2783: \endprocl

2784: 

2785: 

2786: We close our treatment of entropy with some elementary observations.

2787: If $f_n \to 1/2$ weak${}^*$ (meaning that we have weak${}^*$

2788: convergence of the measures having these densities),

2789: must $H(\P^{f_n}) \to \log 2$?

2790: The answer is ``no", as we now demonstrate.

2791: Given any $f : \T \to [0, 1]$ and any integer $n$, let $\mult{n} f$ denote the

2792: function 

2793: $$

2794: \mult{n} f(x) := f(nx)

2795: \,.

2796: $$

2797: Let $\Seq{\eta^{(n)}_k \st k \in \Z}$ have the distribution $\P^{\mult{n}f}$.

2798: Since $f(x) = \sum_{k \in \Z} \widehat{f}(k) e^{2\pi i k x}$ in $L^2(\T)$, the

2799: Fourier expansion of $\mult{n}f$ is $\mult{n}f(x) = \sum_{k \in \Z}

2800: \widehat{f}(k) e^{2\pi i k n x}$ for $n \ne 0$.

2801: In particular, for any $n \ne 0$ and any $r$, the processes

2802: $\Seq{\eta^{(n)}_{nk+r} \st k \in \Z}$ 

2803: each have distribution $P^f$ and are independent of each other 

2804: as $r$ ranges from 0 to $n-1$.

2805: Therefore $H(\P^{\mult{n}f}) = H(\P^f)$.

2806: Note that unless $f$ is constant, 

2807: $H(\P^{\mult{n}f}) = H(\P^f) < H\big[\widehat f(0)\big]$, even

2808: though $\mult{n}f$ tends weak$^*$ to the constant function $\widehat f(0)

2809: \cdot \bfo$.

2810: One can show that a similar phenomenon holds in higher dimensions, that is,

2811: if $f : \Td \to [0, 1]$ and $A: \Td \to \Td$ is a group epimorphism, then

2812: $H(\P^f) = H(\P^{f \circ A})$.

2813: 

2814: 

2815: \bsection {Phase Multiplicity}{s.sk}

2816: 

2817: 

2818: In this section, we classify exactly the set of functions $f$ for 

2819: which $\P^f$ satisfies a

2820: strong full $K$ property. This property, which we now describe, is an 

2821: essential strengthening of the 

2822: usual Kolmogorov or $K$ property. One of the reasons this property 

2823: is interesting is that it is closely connected to the notion

2824: of multiplicity for Gibbs states in 

2825: statistical mechanics; in particular, it corresponds to 

2826: {\it uniqueness}. In the next section, we

2827: classify exactly the set of functions $f$ for which $\P^f$ satisfies a

2828: (1-sided) strong $K$ property.

2829: 

2830: {\bf In this section and the next, we always assume that $f$ is not}

2831: {\bf identically  0 nor identically 1.}

2832: 

2833: 

2834: To begin, we first recall the $K$ property for stationary processes indexed 

2835: by $\Z$. 

2836: There are many equivalent formulations, of which we choose an appropriate one.

2837: 

2838: For $S \subseteq \Z^d$, write $\F(S)$ for the $\sigma$-field on $2^{\Z^d}$

2839: generated by $\eta(e)$ for $e \in S$.

2840: 

2841: \procl d.K

2842: A translation-invariant probability measure $\mu$ on

2843: $2^{\Z}$ is {\bf $K$\/} (or a {\bf Kolmogorov automorphism})

2844: if for any finite cylinder event $E$ and 

2845: for all $\epsilon>0$ there exists an $N$ such that 

2846: $$

2847: \mu\Big[ \big|\mu\big(E \mid \F(\OC{-\infty,-N}\cap \Z)\big)-\mu(E)\big| \ge \epsilon \Big] \le

2848: \epsilon

2849: \,.

2850: $$

2851: \endprocl

2852: 

2853: It is well known that the above definition of $K$ is equivalent 

2854: to having a trivial (1-sided) tail $\sigma$-algebra in the sense that

2855: the $\sigma$-algebra $\bigcap_{m \ge 1} \F\big(\OC{- \infty,

2856: -m}\big)$ is trivial

2857: (see page 120 of \ref b.Georgii/ for a version of this).

2858: The $K$ property is known to be an isomorphism invariant and is

2859: also known to be equivalent to the

2860: property that all nontrivial factors have strictly positive entropy.

2861: The latter implies that a process is $K$ iff its time reversal is $K$,

2862: something which is not immediate from the definition.

2863: 

2864: In order to give a complete discussion of the points that we wish to make,

2865: we need to introduce three further properties, which are respectively

2866: called the full $K$ property, the strong full $K$ property, and the

2867: (1-sided) strong $K$ property, the last property given only for $d=1$.

2868: 

2869: While the notion of the $K$ property generalizes 

2870: to $\Z^d$ (see \ref b.Conze/),

2871: there is a slight strengthening of the definition that 

2872: has a more aesthetic extension to

2873: $\Z^d$. To give this, define $B^d_n:=[-n,n]^d\cap\Z^d$.

2874: 

2875: \procl d.2K

2876: A translation-invariant probability measure $\mu$ on

2877: $2^{\Z^d}$ is {\bf full $K$\/} if for any finite cylinder event $E$ and 

2878: for all $\epsilon>0$ there exists an $N$ such that 

2879: $$

2880: \mu\Big[ \big|\mu\big(E \mid \F((B^d_N)^c)\big)-\mu(E)\big| \ge \epsilon \Big] \le

2881: \epsilon

2882: \,.

2883: $$

2884: \endprocl

2885: 

2886: Analogously to an earlier statement, the

2887: full $K$ property is equivalent (see again page 120 of \ref b.Georgii/)

2888: to having a 

2889: {\bf trivial full tail}, which means that % $\F_{\pm \infty}$

2890: $\bigcap_{m\ge 1} \F\big((B^d_m)^c\big)$ is trivial.

2891: It is well known that $K$ does not imply 

2892: full $K$ even for $d=1$ (see, for example, the bilaterally deterministic

2893: Bernoulli shift processes constructed in \ref b.OW:bilat/).

2894: However, for Markov random fields, the two notions are equivalent

2895: (see \refbmulti{HS,HS2}).

2896: \ref b.L:det/ proved that all (not necessarily $\Z^d$-invariant) 

2897: determinantal probability measures satisfy

2898: the full $K$ property. 

2899: For Gibbs states arising in statistical mechanics, this property 

2900: is equivalent

2901: to an ``extremality'' property of the Gibbs state 

2902: (see page 118 of \ref b.Georgii/ for precisely 

2903: what this means and this equivalence). 

2904: 

2905: The following definitions strengthen the $K$ property in an even more 

2906: essential way. We begin with the full version, which seems to us more 

2907: natural.

2908: 

2909: \procl d.2sK

2910: A translation-invariant probability measure $\mu$ on

2911: $2^{\Z^d}$

2912: is {\bf strong full $K$\/} if for any finite cylinder event $E$ and 

2913: for all $\epsilon>0$ there exists an $N$ such that 

2914: $$

2915: \big|\mu\big(E \mid \F((B^d_N)^c)\big)-\mu(E)\big| <\epsilon

2916: \quad \mu\hbox{-a.e.}

2917: $$

2918: \endprocl

2919: 

2920: For the full $K$ property, ``$\mu$-most'' conditionings far away have

2921: little effect on a ``local event'', while in the strong full $K$ case,

2922: {\it all\/} conditionings far away have little effect on a ``local event''.

2923: This is a substantial difference. An example that illustrates this 

2924: difference is the Ising model in $\Z^2$. 

2925: The plus state for the Ising model at high temperatures

2926: is strong full $K$, while at low temperatures,

2927: it is full $K$ (because it is extremal) but not strong full $K$.

2928: 

2929: Finally, if $d=1$, we define

2930: strong $K$ in the following way (which the reader can presumably

2931: anticipate). One extension of this definition 

2932: to $\Z^d$ (among various possible) will be given in \ref s.1side/.

2933: 

2934: \procl d.1sK

2935: A translation-invariant probability measure $\mu$ on $2^{\Z}$ 

2936: is {\bf strong $K$\/} if for any finite cylinder event $E$ and 

2937: for all $\epsilon>0$ there exists an $N$ such that 

2938: $$

2939: \big|\mu\big(E \mid \F(\OC{-\infty,-N}\cap \Z)\big)-\mu(E)\big| <\epsilon

2940: \quad\mu\hbox{-a.e.}

2941: $$

2942: \endprocl

2943: 

2944: We note that this definition is closely related to, but weaker than,

2945: the $\psi$-mixing property (see \ref b.Bradley/). The difference is that the

2946: event $E$ here is specified in advance, rather than having a uniformity 

2947: for all events $E$, provided only that they depend 

2948: on the random variables

2949: with positive index. If we extend the notion of strong $K$ 

2950: in the obvious way to the case when $\{0,1\}$ is replaced by a countable set,

2951: then the example in \ref b.Bradley/ of a process that is 

2952: $\psi$-mixing, but whose time reversal is not $\psi$-mixing, yields an example

2953: of a strong $K$ process whose time reversal is not.

2954: This cannot occur for our measures $\P^f$ since they are all time reversible.

2955: 

2956: The following development has an analogy with the ``plus and minus states''

2957: in the Ising model from statistical mechanics; however, such knowledge 

2958: is not needed by the reader.

2959: 

2960: Consider a function $f:\Td\to [0,1]$ and consider the corresponding probability measure

2961: $\P^f$ on $2^{\Z^d}$. We shall define a probability measure

2962: $(\P^f)^+$ on $2^{\Z^d}$ which will be ``$\P^f$ conditioned 

2963: on all 1s at $\infty$'';

2964: more specifically (but still not precisely), we want to define $(\P^f)^+$ by

2965: $$

2966: \Pfp := \lim_{n\to \infty} \P^f[\;\cdot \mid  \eta\equiv 1 \hbox{ on }

2967: (B^d_n)^c]\,.

2968: $$

2969: To make sure this is well defined, we proceed in stages. 

2970: Let

2971: $$

2972: (\Pfp)_n := \lim_{k\to\infty} \P^f[\;\cdot \mid  \eta\equiv 1 \hbox{ on

2973: }B^d_{n+k}\backslash B^d_n]\,.

2974: $$

2975: This limit is taken in the weak$^{*}$-topology. \ref p.JNRD/ implies

2976: that, when restricted to $B^d_n$, 

2977: this sequence is stochastically decreasing and hence necessarily converges. 

2978: $(\Pfp)_n$ is therefore well defined. One next defines

2979: $$

2980: \Pfp := \lim_{n\to \infty} (\Pfp)_n

2981: \,,

2982: $$

2983: where again the limit is taken in the weak$^{*}$-topology. 

2984: \ref p.JNRD/ again implies

2985: that for fixed $k$,

2986: $(\Pfp)_n$ restricted to $B^d_k$, is, for $n>k$,

2987: stochastically increasing in $n$ and hence converges.

2988: This implies that its limit $\Pfp$ is well

2989: defined and completes the definition of $\Pfp$.

2990: The stochastic monotonicity results also imply that 

2991: $\Pfp \preccurlyeq  \P^f$ and that

2992: $\Pfp$ is translation invariant.

2993: 

2994: In exactly the same way (using 0 instead of 1 boundary conditions), one defines

2995: $\Pfm$, which satisfies $\P^f  \preccurlyeq \Pfm$. Analogy with the

2996: ferromagnetic Ising model in statistical mechanics leads us to the

2997: following definition.

2998: 

2999: \procl d.pt

3000: The probability measure $\P^f$ has {\bf phase multiplicity} if $\Pfm\neq \Pfp$, and otherwise has {\bf phase uniqueness}.

3001: \endprocl

3002: 

3003: \procl l.pt

3004: $\P^f$ has phase uniqueness if and only if it is strong full $K$.

3005: \endprocl

3006: 

3007: The reason this is true is that \ref p.JNRD/ implies

3008: that the most

3009: extreme boundary conditions are when all 1s or all 0s are used and the

3010: measures corresponding to any other boundary conditions are

3011: ``stochastically trapped'' between these two special cases. 

3012: A detailed proof follows straightforwardly from the stochastic 

3013: monotonicity arguments above and

3014: is left to the reader.

3015: 

3016: We shall now

3017: obtain necessary and sufficient conditions for the equality of $\P^f$ and

3018: $\Pfp$. From these, a necessary and sufficient condition for the strong full $K$ property will easily emerge. 

3019: 

3020: 

3021: Let $\calT$ denote the set of trigonometric polynomials on $\T^d$.

3022: Let $L^2(1/f)$ denote the set 

3023: $$

3024: \Big\{h:\Td\to \C \st \int_\Td {|h|^2 \over f} \, d\lambda_d <\infty\Big\}\,. 

3025: $$

3026: Here we use the convention that $0/0:=0$. Note also that

3027: $h$ needs to vanish where $f$ does.

3028: 

3029: Our main result on phase multiplicity is the following.

3030: 

3031: \procl t.sn

3032: Assume that $f:\Td\to [0,1]$ is measurable. %positive a.e. 

3033: The following are equivalent. 

3034: \beginitems

3035: \itemrm{(i)} $\Pfp=\P^f$; 

3036: \itemrm{(ii)} $f$ is in the closure in $L^2(1/f)$ of $\calT\cap L^2(1/f)$;

3037: \itemrm{(iii)} There exists a nonzero trigonometric polynomial $T$ such that 

3038: ${|T|^2 \over f}\in L^1(\Td)$; i.e., $\calT\cap L^2(1/f)\neq {0}$.

3039: \enditems

3040: Moreover, if $\Pfp\neq \P^f$, then $\Pfp= \delta_\bfz$.

3041: \endprocl

3042: 

3043: 

3044: \proof

3045: We shall first show that (i) and (ii) are equivalent and then

3046: that (ii) and (iii) are equivalent.

3047: 

3048: {\bf (i) implies (ii):} Let $u_n$ be the element in $[B]_f$ achieving the

3049: minimum in \ref e.closest/ for $B := (B^d_n)^c$.

3050: Then 

3051: $$

3052: \|\bfo\|^2_f = \|\bfo - u_n\|^2_f + \|u_n\|^2_f

3053: \,.

3054: $$

3055: By (i) and \ref e.closest/, we have $\|\bfo - u_n\|_f \to \|\bfo\|_f$ as $n

3056: \to\infty$, or in other words, $\|u_n\|_f \to 0$ as $n \to\infty$.

3057: Furthermore, $\bfo - u_n \perp [(B^d_n)^c]_f$ in $L^2(f)$, which is the same

3058: as $(\bfo - u_n) f$ being a trigonometric polynomial $T_n$ with $\widehat

3059: {T_n}$ supported in $B^d_n$.

3060: We have 

3061: $$

3062: \|f - T_n\|^2_{(1/f)}

3063: =

3064: \int |f - T_n|^2 {1 \over f} \,d\lambda_d

3065: =

3066: \int \left|\bfo - {T_n \over f}\right|^2 f \,d\lambda_d

3067: =

3068: \int |u_n|^2 f \,d\lambda_d

3069: =

3070: \|u_n\|^2_f

3071: \label e.twoways

3072: $$

3073: tends to 0 as $n \to\infty$, which proves (ii).

3074: 

3075: {\bf (ii) implies (i):}

3076: Given $\epsilon > 0$, let

3077: $T$ be a trigonometric polynomial with $\|f - T\|_{(1/f)} < \epsilon$.

3078: Let $n$ be such that $\widehat T$ is supported in $B^d_n$.

3079: Define $v := T/f$, which is in $L^2(f)$.

3080: Then $v \perp [(B^d_n)^c]_f$ in $L^2(f)$ and $\|\bfo - v\|_f = \|f -

3081: T\|_{(1/f)} < \epsilon$, so 

3082: $\|P_{[(B^d_n)^c]_f}^\perp \bfo\|_f^2 > {\|\bfo\|^2_f -\epsilon^2}$.

3083: Combining this with \ref e.byproj/, we see that

3084: $\Pfp[\eta(\bfz) = 1] > \P^f[\eta(\bfz) = 1] - \epsilon^2$. Since

3085: this holds for any $\epsilon > 0$, we obtain that $\Pfp[\eta(\bfz) = 1] \ge

3086: \P^f[\eta(\bfz) = 1]$. However, 

3087: since $\Pfp \preccurlyeq \P^f$ and both are translation invariant, 

3088: the two measures are actually equal by \ref l.weak/, which proves (i).

3089: 

3090: {\bf (ii) implies (iii):} 

3091: This is immediate.

3092: 

3093: {\bf (iii) implies (ii):}

3094: Assume there exists a nonzero

3095: trigonometric polynomial $T \in L^2(1/f)$. 

3096: Let $\epsilon > 0$. 

3097: Choose $B\subseteq \Td$ such that

3098: $$

3099: \int_B \left({|T|^2 \over f} + 2|T| +f\right) < \epsilon 

3100: $$

3101: and $B$ is an open set containing the zero set of $T$ (which is a set of

3102: measure 0, as remarked in the proof of \ref t.support/). Let 

3103: $g := T  \I B + f  \I {B^c}$. Then $g/T \in L^\infty(\Td)$ since $T$ 

3104: is bounded away from 0 on $B^c$ and $0\le f \le 1$. We can now choose 

3105: trigonometric 

3106: polynomials $p_n$ on $\Td$ such that $p_n \to g/T$ a.e.\ 

3107: on $\Td$ with $\|p_n\|_\infty \le \|g/T\|_\infty$ for all $n$ (just use

3108: $p_n:=(g/T) * K^d_n$, where $K^d_n$ is, as before, the Fej\'er kernel for

3109: $\Td$).  

3110: Since $|T|^2/f \in L^1(\Td)$, it follows from the 

3111: Lebesgue dominated convergence theorem that

3112: $$

3113: \lim_{n\to\infty}\int_\Td {|T|^2 \over f}  \left|{g \over T}-p_n\right|^2 =0\,.

3114: $$

3115: Hence there exists a trigonometric polynomial $h$ such that

3116: $$

3117: \int_\Td {|T|^2 \over f}  \left|{g \over T}-h\right|^2 < \epsilon\,.

3118: \label e.1

3119: $$

3120: Minkowski's inequality (applied to $L^2(|T|^2/f)$) yields

3121: $$

3122: \left(\int_\Td {|T|^2 \over f}  \left|{f \over T}-h\right|^2\right)^{{1\over

3123: 2}} \le

3124: \left(\int_\Td {|T|^2 \over f}  \left|{f \over T}-{g \over

3125: T}\right|^2\right)^{{1\over 2}} +

3126: \left(\int_\Td {|T|^2 \over f}  \left|{g \over T}-h\right|^2\right)^{{1\over

3127: 2}}\,.

3128: $$

3129: The second summand is at most $\epsilon^{{1/ 2}}$ by \ref e.1/. 

3130: The first summand is 

3131: $$

3132: \left(\int_B {|T|^2 \over f}  \left|{f \over T}-1\right|^2\right)^{{1\over 2}}

3133: \le 

3134: \left(\int_B f+ 2|T| + {|T|^2 \over f}\right)^{{1\over 2}} <\epsilon^{{1\over

3135: 2}}

3136: $$

3137: by choice of $B$.

3138: Hence 

3139: $\int_\Td |f-Th|^2 {1 \over f} \, dx < 4\epsilon$. Since $Th$ is a 

3140: trigonometric polynomial, (ii) is proved.

3141: 

3142: Finally, assume that $\Pfp\neq \P^f$. Since (iii) fails,

3143: $[(B^d_n)^c]_f^\perp = \bfz$ for all $n$, which means by \ref e.byproj/

3144: that $\Pfp[\eta(\bfz) = 1] = 0$, or, in other words, $\Pfp = \delta_\bfz$.

3145: \Qed

3146: 

3147: \procl x.continuous-finite

3148: If $f:\T\to [0,1]$ is continuous and has a finite number of 0s with $f$

3149: approaching each of these 0s at most polynomially quickly, then $\Pfp=\P^f$

3150: since it is easy to construct a trigonometric polynomial $T$ with

3151: the same 0s as $f$ and approaching 0 at least as quickly as $f$

3152: (for example, $T$ could be of the form 

3153: $\prod_{i=1}^k \sin^n 2\pi(x-a_i)$).

3154: In particular, if $f$ is real analytic and not $\bfz$, 

3155: then $\Pfp = \P^f$. 

3156: \endprocl

3157: 

3158: \procl x.trivial

3159: If $f$ vanishes on a set of positive measure, then $\Pfp=\delta_\bfz$.

3160: \endprocl

3161: 

3162: \procl x.cont

3163: If $f:\T\to [0,1]$ is a continuous function with a single zero at $x_0\in \T$

3164: and $f(x) = e^{-1/|x-x_0|}$ in some neighborhood of $x_0$,

3165: then $\Pfp=\delta_\bfz$. Indeed, there is no nonzero 

3166: trigonometric polynomial $T$ with

3167: $|T|^2/f\in L^1(\T)$ since the rate at which a trigonometric polynomial 

3168: approaches 0 is at most polynomially quickly.  

3169: \endprocl

3170: 

3171: \procl x.depends-distribution

3172: We give an example to show that the property $\Pfp = \P^f$ does {\it not\/}

3173: depend only on the distribution of $f$, even among real-analytic functions

3174: $f$.

3175: In 2 dimensions, the function 

3176: $$

3177: f(x, y) := \sin^2 (2\pi y - \cos 2\pi x)

3178: $$

3179: generates a system for which $\Pfp \ne \P^f$.

3180: This is because $f$ vanishes (even to second order) on a curve 

3181: $$

3182: y = {1 \over 2 \pi} \cos 2 \pi x + \Z

3183: \label e.curve

3184: $$

3185: that is not in the zero set of any trigonometric polynomial in two

3186: variables except the zero polynomial.

3187: However, $f$ has the same distribution as the trigonometric polynomial

3188: $g(x, y) := \sin^2 2 \pi y$ and

3189: $(\P^g)^+ = \P^g$.

3190: To see that the curve \ref e.curve/ does not lie in the zero set of any

3191: nonzero trigonometric polynomial, suppose that $T(x, y)$ is a trigonometric

3192: polynomial that vanishes there. Write $w := e^{2\pi i x}$ and $z := e^{2\pi

3193: i y}$. By multiplying by a suitable complex

3194: exponential, we may assume that $P(w, z) = T(x, y)$ is an analytic

3195: polynomial. 

3196: Our assumption is that 

3197: $$

3198: h(w) := P(w, e^{i(w+w^{-1})/2})

3199: $$

3200: satisfies $h(w) = 0$ for all $|w|=1$.

3201: Since $h$ is an analytic function for $w \ne 0$, it follows that $h(w) = 0$

3202: for all $w \ne 0$.

3203: Now for each $z \ne 0$, there are an infinity of $w$ such that

3204: $e^{i(w+w^{-1})/2} = z$. Therefore for each $z \ne 0$,

3205: there are an infinity of $w$ such that $P(w, z) = 0$.

3206: Since a polynomial has only a finite number of zeroes if it is not

3207: identically zero, this means that for each $z \ne 0$, $P(w, z) = 0$ for all

3208: $w \in \C$.

3209: Hence $P \equiv 0$, as desired.

3210: \endprocl

3211: 

3212: \procl x.algvar

3213: For $d = 2$, if $f$ is real analytic on a neighborhood of its 

3214: zero set, then $\Pfp = \P^f$ iff its zero set is contained in a 

3215: (nontrivial) algebraic variety, where we view $\T^2$ as $ \{ (z_1, 

3216: z_2) \in \C^2 \st |z_1| = |z_2| = 1 \}$.

3217: This is because the slowest $f$ can vanish at a point is of order $x^2+y^2$

3218: (since the constant and linear terms of $f$ must vanish)

3219: and $1/(x^2+y^2)$ is not integrable. Therefore, all the zeroes of $f$ must be

3220: cancelled by those of $T$.

3221: Conversely, if the zero set of $f$ is contained in the zero set of a

3222: trigonometric polynomial $T$, then by {\L}owasiewicz's inequality (see,

3223: e.g., \ref b.BierMil/, Theorem 6.4), for a sufficiently large $n$, we have

3224: $|T|^n/f$ is bounded, whence integrable.

3225: \endprocl

3226: 

3227: 

3228: \procl x.depends-distribution1

3229: Here, we give a 1-dimensional example showing that the property $\Pfp =

3230: \P^f$ does {\it not\/} depend only on the distribution of $f$, even among

3231: continuous $f$.

3232: Let $f(x) := \sqrt{x} |\sin (\pi/x)|$ on $[0, 1]$.

3233: Then $\Pfp \ne \P^f$ since there are an infinite number of first-order zeroes.

3234: However, let $g$ be the increasing rearrangement of $f$ on $[0, 1]$ and define

3235: $$

3236: h(x) := \cases{g(2x) &if $x \le 1/2$,\cr

3237:                g(2-2x) &if $x \ge 1/2$.\cr}

3238: $$

3239: Then $h$ is continuous and has the same distribution as $f$, yet

3240: $(\P^h)^+ = \P^h$ by \ref x.continuous-finite/ together with an easy

3241: computation.

3242: Of course, there is no such example for real-analytic $f$ in one dimension

3243: by \ref x.continuous-finite/.

3244: \endprocl

3245: 

3246: We finally state and prove our necessary and sufficient condition for

3247: the strong full $K$ property.

3248: 

3249: \procl c.sn

3250: Consider $f:\Td\to [0,1]$ and the corresponding probability measure

3251: $\P^f$. Then

3252: $\P^f$ is strong full $K$ if and only if there is a nonzero

3253: trigonometric polynomial $T$ such that 

3254: ${|T|^2 \over f (\bfo-f)} \in L^1(\Td)$.

3255: \endprocl

3256: 

3257: \proof

3258: According to \ref t.sn/, 

3259: $\Pfp=\P^f$ iff there exists a nonzero

3260: trigonometric polynomial $T_1$ such that

3261: ${|T_1|^2 \over f} \in L^1(\Td)$. The same argument applied to $\bfo - f$

3262: tells us that $\Pfm=\P^f$ iff there exists a nonzero

3263: trigonometric polynomial $T_2$ such that

3264: ${|T_2|^2 \over \bfo-f} \in L^1(\Td)$. \ref l.pt/ now completes the proof.

3265: (Take $T := T_1 T_2$.)

3266: \Qed

3267: 

3268: \procl r.analytic

3269: Observe that by trivial scaling, we could have an $f$ whose Fourier

3270: coefficients go to zero slowly but which is bounded away from 0 and 1.

3271: Hence slow decay of Fourier coefficients does not imply phase multiplicity.

3272: On the other hand, if the coefficients decay exponentially, then 

3273: $\sum_{n \in \Z} \widehat{f}(n) z^n$ is complex-analytic in an annulus around

3274: the unit circle, which implies that its restriction to the unit circle is

3275: real-analytic and hence there is phase uniqueness.

3276: \endprocl

3277: 

3278: \procl r.folner

3279: We mention that if $\Pfp=\P^f$, then this measure is

3280: F{\o}lner independent in the sense of \ref b.Adams/ and

3281: if in addition $\Pfm=\P^f$, then this measure is even strong

3282: F{\o}lner independent in the sense that any conditioning outside

3283: a box yields a measure which is $\dbar$ close to the unconditioned

3284: process. The arguments for proving these facts are analogous to those

3285: of \ref b.OW:ising/ (see \ref b.Adams/ for a published version), where it is

3286: proved that the plus state for the Ising model is a Bernoulli shift.

3287: (The concept of F{\o}lner independence has been used also in \ref b.HS/ and

3288: \ref b.hoffman:MRF/.)

3289: \endprocl  

3290: 

3291: We close this section with an interpretation of

3292: our phase multiplicity results in terms 

3293: of interpolation questions for wide-sense stationary 

3294: processes. The following comments can be proved from \ref t.closest/ 

3295: together with the well-known correspondence 

3296: between interpolation and Szeg\H{o} infima.

3297: First, $\Pfp= \delta_\bfz$ is equivalent to being able to 

3298: interpolate perfectly from information far away for 

3299: a wide-sense stationary process $\Seq{Y_n}$  with spectral density $f$,

3300: which means that for every $n$, $Y_0$ is in the closed linear

3301: subspace spanned by  $\{Y_k\}_{|k|\ge  n}$. For $d=1$, it

3302: is stated on p.~102 of \ref b.Roz/ that this latter condition is 

3303: equivalent to the negation of condition (iii) in \ref t.sn/.

3304: A similar statement holds for $\Pfm$ with $f$ replaced by $\bfo-f$.

3305: The proof of \ref t.sn/ implies (via this whole correspondence) that, 

3306: for spectral measures that are

3307: absolutely continuous with a bounded nonnegative density,

3308: if perfect linear interpolation fails, then

3309: our ability to interpolate as $n$ gets large goes to 0; i.e.,

3310: the length of the projection of $Y_0$ onto the closed linear

3311: subspace spanned by  $\{Y_k\}_{|k|\ge  n}$ goes to 0 as $n\to\infty$.

3312: 

3313: \bsection{1-Sided Phase Multiplicity}{s.1side}

3314: 

3315: We first study the notion of strong $K$ for $d=1$. 

3316: This is natural since the

3317: structure of ``one-sided'' behavior can be very different from

3318: ``two-sided'' behavior in various situations; for example, the existence

3319: of bilaterally deterministic Bernoulli shift processes demonstrates this.

3320: In contrast with \ref x.depends-distribution1/, we shall see that whether

3321: $\P^f$ is strong $K$ depends only on the distribution of $f$.

3322: 

3323: Consider a function $f:\T\to [0,1]$ and the 

3324: corresponding probability measure

3325: $\P^f$ on $2^{\Z}$. We define a probability measure

3326: $(\P^f)^{+,1}$ on $2^{\Z}$ thought of as ``$\P^f$ conditioned 

3327: on all 1s at $-\infty$'' (the superscript ``1'' refers to the fact that we are

3328: doing this on 1 side); namely,

3329: $$

3330: (\P^f)^{+,1}:= \lim_{n\to \infty} \lim_{k \to\infty} 

3331: \P^f\big[\;\cdot \mid \eta\equiv 1 \hbox{ on }

3332: [-n-k,-n]\cap \Z\big]

3333: \,.

3334: $$

3335: The existence of the limit and its translation invariance follow from

3336: stochastic monotonicity, as did that of $(\P^f)^+$.

3337: As before, we also have $(\P^f)^{+,1} \preccurlyeq  \P^f$. 

3338: In the analogous way (using 0 instead of 1 boundary conditions), one defines

3339: $(\P^f)^{-,1}$, which then satisfies $\P^f  \preccurlyeq (\P^f)^{-,1}$. 

3340: 

3341: \procl d.pt1

3342: The probability measure $\P^f$ has a {\bf 1-sided phase multiplicity} if 

3343: $(\P^f)^{-,1}\neq(\P^f)^{+,1}$, and otherwise has {\bf 1-sided phase

3344: uniqueness}.

3345: \endprocl

3346: 

3347: Note the following analogue (whose proof is left to the reader) 

3348: of \ref l.pt/ for the 1-sided case.

3349: 

3350: \procl l.pt1

3351: $\P^f$ has 1-sided phase uniqueness if and only if it 

3352: is strong $K$.

3353: \endprocl

3354: 

3355: Our main result for 1-sided phase multiplicity is the following. 

3356: 

3357: \procl t.sn1GM

3358: If $f:\T\to [0,1]$, then $(\P^f)^{+,1}=\P^f$ iff $\GM(f) > 0$. 

3359: Moreover, if $(\P^f)^{+,1}\neq \P^f$, then $(\P^f)^{+,1}= \delta_\bfz$.

3360: \endprocl

3361: 

3362: \proof

3363: By \ref t.szego/, we have

3364: $$

3365: \lim_{k \to\infty} \P^f[\eta(0) = 1 \mid \eta_{-n} = \eta_{-n-1} =

3366: \cdots = \eta_{-n-k} = 1]

3367: =

3368: \sum_{l=0}^{n-1} |\widehat{\varphi_f}(l)|^2

3369: \,.

3370: $$

3371: (In fact, via \ref t.closest/, this special case of \ref t.szego/ is due to

3372: Kolmogorov and Wiener; see \ref b.GrenanderSzego/, Section 10.9.)

3373: Taking $n \to\infty$, we find that

3374: $$

3375: (\P^f)^{+,1}\big[\eta(0) = 1\big] = \|\widehat{\varphi_f}\|^2_2 = \|\varphi_f\|^2_2

3376: \,.

3377: $$

3378: If $\GM(f) > 0$, then we obtain

3379: $$

3380: (\P^f)^{+,1}\big[\eta(0) = 1\big] = \|f\|_1 = \widehat

3381: f(0) = \P^f\big[\eta(0) = 1\big]

3382: \,.

3383: $$

3384: Since $(\P^f)^{+,1} \preccurlyeq \P^f$ and both probability measures are

3385: $\Z$-invariant, it follows by \ref l.weak/ that $(\P^f)^{+,1} = \P^f$.

3386: On the other hand, if $\GM(f) = 0$, then 

3387: $$

3388: (\P^f)^{+,1}\big[\eta(0) = 1\big] = 0

3389: \,,

3390: $$

3391: so

3392: $(\P^f)^{+,1} = \delta_\bfz$.

3393: \Qed

3394: 

3395: Our necessary and sufficient condition for the strong $K$ property 

3396: now follows immediately.

3397: 

3398: \procl c.sn1

3399: For $f:\T\to [0,1]$, the corresponding probability measure $\P^f$

3400: is strong $K$ if and only if $\GM(f) \GM(\bfo-f) > 0$.

3401: \endprocl

3402: 

3403: 

3404: 

3405: We now construct a process $\P^f$ that is strong $K$ but not strong full $K$.

3406: 

3407: \procl t.diff

3408: There exists an $f:\T\to [0,1]$ such that $\P^f$ is strong $K$ but not

3409: strong full $K$.

3410: \endprocl

3411: 

3412: \proof

3413: Choose $f$ such that $f$ is continuous, bounded away from 1, vanishes at a

3414: single point $x_0$, and such that $f(x) = 

3415: e^{-1/(|x-x_0|)^{1/2}}$ in some neighborhood of $x_0$.

3416: Since a trigonometric polynomial vanishes at its zeroes at most

3417: polynomially quickly, there cannot exist a nonzero

3418: trigonometric polynomial $T$

3419: such that $|T|^2/f \in L^1(\T)$. Hence by \ref c.sn/,

3420: $\P^f$ is not strong full $K$. 

3421: On the other hand, it is clear that $\GM(f) \GM(\bfo-f) > 0$, 

3422: so by \ref c.sn1/, we know that $\P^f$ is strong $K$.

3423: \Qed

3424: 

3425: 

3426: \procl r.plus

3427: By combining Theorems 

3428: \briefref t.dom/ and \briefref t.sn1GM/, we see that

3429: $\P^f \succcurlyeq \P^g$ implies 

3430: $(\P^f)^{+,1} \succcurlyeq (\P^g)^{+,1}$.

3431: However, 

3432: it is not necessarily the case that $(\P^f)^+ \succcurlyeq (\P^g)^+$.

3433: For example, let $g \equiv 1/2$ and $f$ be a function as described in the last

3434: result, but which has a geometric mean larger than $1/2$

3435: (recall \ref t.dom/). 

3436: This example also shows that if $\P^g \preccurlyeq \P^f$,

3437: there do not necessarily exist $g'\le f'$ 

3438: such that $\P^{g'} = \P^g$ and $\P^{f'} = \P^f$.  To see this,

3439: let $f$ and $g$ be as above. Then $\P^{g'} = \P^g$ easily implies that

3440: $g' \equiv 1/2$, which in turn (using $g'\le f'$) implies that

3441: $(\P^{f'})^+ =  \P^{f'}$. Since

3442: $\Pfp \neq  \P^f$, this contradicts the fact that $\P^f = \P^{f'}$.

3443: \endprocl

3444: 

3445: 

3446: It may be of interest to see another proof (which we only sketch)

3447: of \ref t.sn1GM/ that does not

3448: depend on the asymptotics of \ref t.szego/, but rather follows

3449: the lines of the proof of \ref t.sn/.

3450: The appropriate replacement of the trigonometric polynomials

3451: for this question is the set 

3452: $$

3453: \calA :=

3454: \{h\in L^2(\T) \st \hbox{ there exists } \ell  \hbox{ such that }

3455: \widehat{h}(k) =0 \hbox{ for } k< \ell\}\,.

3456: $$

3457: Note that when $\ell$ is taken to be 0, we get the {\bf Hardy space}

3458: of analytic functions, denoted $H^2(\T)$.

3459: 

3460: 

3461: \procl t.sn1

3462: Assume that $f:\T\to [0,1]$ and let $\calA$ be as defined as above.

3463: The following are equivalent. 

3464: \beginitems

3465: \itemrm{(i)} $(\P^f)^{+,1}=\P^f$; 

3466: \itemrm{(ii)} $f$ is in the closure in $L^2(1/f)$ of $\calA\cap L^2(1/f)$;

3467: \itemrm{(iii)} There exists a nonzero element $T\in \calA$ such that 

3468: ${|T|^2 \over f}\in L^1(\T)$; i.e., $\calA\cap L^2(1/f)\neq {0}$;

3469: \itemrm{(iv)} $\GM(f) > 0$. 

3470: \enditems

3471: Moreover, if $(\P^f)^{+,1}\neq \P^f$, then $(\P^f)^{+,1}= \delta_\bfz$.

3472: \endprocl

3473: 

3474: \proof

3475: {\bf (i) iff  (ii)} and {\bf (ii) implies (iii):} These are proved

3476: exactly as in \ref t.sn/.

3477: 

3478: {\bf (iii) implies (ii):}

3479: Fix a nonzero $T \in \calA$ with ${|T|^2 \over f}\in L^1(\T)$.

3480: Given any $\epsilon$, there exists $\delta>0$ such that if $B$ is any

3481: subset of $\T$ with measure less than $\delta$, then

3482: $$

3483: \int_B \left({|T|^2 \over f} + 2|T| +f\right) < \epsilon\,.

3484: $$

3485: Since $T$ vanishes only on a set of measure 0

3486: (see \ref b.Rudin:RCA/, Theorem 17.18, p.~345),  

3487: there is a $\gamma>0$ such that

3488: $\{x\in \T \st |T(x)| < \gamma\}$ has measure less than $\delta$.

3489: If we take $B$ to be this latter set, then we can proceed exactly 

3490: as in \ref t.sn/ since now $T$ is bounded away from 0 on $B^c$. 

3491: 

3492: {\bf (iii) implies (iv):}

3493: Any function in $\calA$ is a complex exponential times a function in the

3494: Hardy space $H^2(\T)$. Hence there is 

3495: a nonzero function $T \in H^2(\T)$ with $|T|^2/f \in

3496: L^1(\T)$. Now

3497: $\GM(|T|) > 0$  (see \ref b.Rudin:RCA/, Theorem 17.17, p.~344). 

3498: Letting $g := |T|^2/f$, we have that

3499: $g \in L^1(\T)$ and hence $\GM(g) \le \int g < \infty$.

3500: Therefore $\GM(f) = \GM(|T|^2/g) = \GM(|T|)^2/\GM(g) > 0$, as desired.

3501: 

3502: {\bf (iv) implies (iii):}

3503: If $\GM(f) > 0$, then take $T := \varphi_f$, defined in

3504: \ref s.dom/.

3505: For this $T$, we have $|T|^2/f = \bfo$, which is trivially integrable.

3506: 

3507: The final statement is proved as in \ref t.sn/.

3508: \Qed

3509: 

3510: 

3511: 

3512: 

3513: We now extend \ref t.sn1GM/ and \ref c.sn1/ to higher dimensions.

3514: Fix a choice of ordering of $\Zd$ given by a set $\ord$ 

3515: and a sequence $\Seq{k_n}\in \Z^d$ such that the distance from

3516: $\bfz$ to $-(\ord+k_n)$ goes to $\infty$.

3517: 

3518: \procl d.1sK

3519: A translation-invariant probability measure $\mu$ on $2^{\Z^d}$ 

3520: {\bf has phase uniqueness 

3521: relative to the ordering induced by $\ord$ and the sequence $\Seq{k_n}$ \/} 

3522: as above if for any finite cylinder event $E$ and 

3523: for all $\epsilon>0$ there exists an $N$ such that for all $n\ge N$

3524: $$

3525: \big|\mu\big(E \mid \F(-(\ord+k_n)\big)-\mu(E)\big| <\epsilon

3526: \quad\mu\hbox{-a.e.}

3527: $$

3528: \endprocl

3529: 

3530: It is easy to check using \ref p.JNRD/

3531: that this definition {\it does not depend\/} on the choice

3532: of the sequence $\Seq{k_n}$. However,

3533: it turns out that the choice of ordering, $\ord$, of $\Zd$ can make a 

3534: difference. In particular, if the ordering is archimedean

3535: (meaning that for any two positive elements $a$ and $b$, 

3536: there exists an integer $n$ such that $na > b$), 

3537: such as\ftnote{${}^*$}{In fact, all archimedean orders arise in this way; 

3538: see \ref b.Teh/, \ref b.Zaiceva/, or \ref b.Trevisan/.} when the ordering is

3539: induced by $\ord := \{ k \in \Z^d \st k \cdot x > 0 \}$, 

3540: where $x \in \R^d$ is a fixed vector having two coordinates with an

3541: irrational ratio, then we have a complete characterization

3542: of 1-sided phase multiplicity in terms of the geometric mean.  For the standard

3543: lexicographic ordering, however, 1-sided phase multiplicity cannot be

3544: characterized by the geometric mean alone, as shown by the

3545: example $f :=   \I{[0,1]\times[0, 1/2]}$ in two dimensions.

3546: Here, both $f$ and $\bfo - f$ have 0 geometric mean, but the columns of a

3547: configuration are independent under $\P^f$, so that conditioning on the remote

3548: past has no effect on the present.

3549: On the other hand, positivity of the geometric means will still {\it suffice\/}

3550: for phase uniqueness 

3551: relative to the ordering induced by $\ord$,

3552: as we now prove. Analogously to the

3553: one-dimensional case, we let, for our given ordering $\ord$,

3554: $$

3555: (\P^f)^{+,1}:= \lim_{n\to \infty}

3556: \P^f\big[\;\cdot \mid \eta\equiv 1 \hbox{ on } -(\ord+k_n)\big]

3557: \,.

3558: $$

3559: By \ref p.JNRD/, this limit exists and is independent of $\Seq{k_n}$.

3560: As before, $(\P^f)^{-,1}$ is defined analogously using boundary

3561: conditions of 0s, one-sided phase multiplicity is defined

3562: by $(\P^f)^{-,1}\neq(\P^f)^{+,1}$, and the property

3563: in \ref d.1sK/

3564: is equivalent to one-sided phase uniqueness.

3565: 

3566: \procl t.sn1GM-d

3567: Let $f:\Td\to [0,1]$. If $\GM(f) > 0$, then $(\P^f)^{+,1}=\P^f$.

3568: If $\GM(f) = 0$ and the ordering is archimedean, then

3569: $(\P^f)^{+,1}= \delta_\bfz$. In particular, if the 

3570: ordering is archimedean, then $\P^f$ has phase 

3571: uniqueness (relative to this ordering) if and only if 

3572: $\GM(f) \GM(\bfo-f) > 0$.

3573: \endprocl

3574: 

3575: We begin with the background results on which we rely for the proof.

3576: For an ordering induced by $\ord$, define the 

3577: corresponding {\bf Helson-Lowdenslager space}

3578: $$

3579: \hl^2 := \hl^2(\T^d, \ord) := \Big \{ f \in L^2(\T^d) \st \supp \widehat f

3580: \subset \ord \cup \{ \bfz \} \Big \}

3581: \,.

3582: $$

3583: For $0 \le f \in L^1(\lambda)$, let $[\ord \cup \{ \bfz \}]$ be the linear span

3584: of $ \{ \ee_k \st k \in \ord \cup \{ \bfz \} \}$ and $[\ord \cup \{ \bfz

3585: \}]_f$ be its closure in $L^2(f)$.

3586: The replacement for outer functions is the class of {\bf spectral factors},

3587: which are the functions $\varphi \in \hl^2$ with the additional properties 

3588: $$

3589: \widehat\varphi(\bfz) > 0

3590: \label e.at0

3591: $$

3592: and

3593: $$

3594: 1/\varphi \in [\ord \cup \{ \bfz \}]_{|\varphi|^2}

3595: \,.

3596: \label e.inverse

3597: $$

3598: \ref b.HelLow:I/ show that for any $\ord$ and for any

3599: $0 \le f \in L^1(\T^d)$, the condition $\GM(f) >

3600: 0$ is equivalent to the existence of a spectral factor $\varphi_f$ such that

3601: $|\varphi_f|^2 = f$ a.e. [More precisely, they prove $\GM(f) > 0$ iff $\exists

3602: \varphi \in \hl^2$ satisfying $|\varphi|^2 = f$ a.e.\ and %\ref e.support/ and 

3603: \ref e.at0/. Their proof

3604: shows that in this case, $\varphi$ can be chosen so that also \ref e.inverse/

3605: holds.] Furthermore, \ref b.Lyons:szego/ shows that $\varphi_f$ is then unique.

3606: 

3607: Denote by $\nu_{f}$ the measure on $2^{\ord \cup \{ \bfz \}}$ given

3608: by $\P^f[\;\cdot \mid \eta \restrict (-\ord) \equiv 1]$. This is, as usual,

3609: defined by conditioning on more and more 1s in $-\ord$ and using

3610: \ref p.JNRD/. The results of \ref b.Lyons:szego/ imply the following.

3611: 

3612: \procl t.szego-d 

3613: Fix an ordering induced by a set $\ord$.

3614: Let $f : \Td \to [0, 1]$ be measurable. If $\GM(f) > 0$, define

3615: $\varphi_f$ to be its spectral factor (with respect to $\ord$).

3616: Otherwise, define $\varphi_f := \bfz$.

3617: If $\GM(f) > 0$ or the ordering of $\Zd$ given by $\ord$ is archimedean, then

3618: the measure $\nu_{f}$ is equal to

3619: the determinantal probability measure corresponding to the positive

3620: contraction on $\ell^2(\ord \cup \{ \bfz \})$ whose $(j, k)$-matrix entry is 

3621: $$

3622: \sum_{l\in \ord \cup \{ \bfz \},\; l \preccurlyeq j,k} 

3623: \overline{\widehat{\varphi_f}(j-l)}\widehat{\varphi_f}(k-l)

3624: \,.

3625: $$

3626: \endprocl

3627: 

3628: \proofof t.sn1GM-d

3629: Suppose first that $\GM(f) > 0$ or the ordering given by $\ord$ is archimedean.

3630: By \ref t.szego-d/, we have

3631: $$

3632: \P^f[\eta(\bfz) = 1 \mid \eta \restrict -(\ord+k_n) \equiv 1]

3633: =

3634: \P^f[\eta(k_n) = 1 \mid \eta \restrict (-\ord) \equiv 1]

3635: =

3636: \sum_{l \in \ord \cup \{ \bfz \},\; l \preccurlyeq k_n} 

3637: |\widehat{\varphi_f}(k_n - l)|^2

3638: \,.

3639: $$

3640: Taking $n \to\infty$, we find that

3641: $$

3642: (\P^f)^{+,1}\big[\eta(\bfz) = 1\big] = \|\widehat{\varphi_f}\|^2_2 =

3643: \|\varphi_f\|^2_2

3644: \,.

3645: $$

3646: If $\GM(f) > 0$, then we obtain

3647: $$

3648: (\P^f)^{+,1}\big[\eta(\bfz) = 1\big] = \|f\|_1 = \widehat

3649: f(\bfz) = \P^f\big[\eta(\bfz) = 1\big]

3650: \,.

3651: $$

3652: Since $(\P^f)^{+,1} \preccurlyeq \P^f$ and both probability measures are

3653: $\Zd$-invariant, it follows that $(\P^f)^{+,1} = \P^f$.

3654: However, if $\GM(f) = 0$ and the ordering is archimedean, then 

3655: $$

3656: (\P^f)^{+,1}\big[\eta(\bfz) = 1\big] = 0

3657: \,,

3658: $$

3659: so

3660: $(\P^f)^{+,1} = \delta_\bfz$.

3661: The last statement of the theorem follows as before.

3662: \Qed

3663: 

3664: 

3665: Analogously to \ref s.sk/, our one-sided phase multiplicity results 

3666: for $d=1$ can be translated to known results for prediction (rather than

3667: interpolation) questions for wide-sense stationary 

3668: processes. We leave these observations to the reader. 

3669: 

3670: \bsection{Open Questions}{s.open}

3671: 

3672: \procl q.1

3673: Calculate $H(\P^f)$.

3674: \endprocl

3675: 

3676: We conjecture that entropy is concave:

3677: 

3678: \procl g.concave

3679: For any $f$ and $g$, we have $H\big(\P^{(f+g)/2}\big) \ge \big(H(\P^f) +

3680: H(\P^g)\big)/2$.

3681: \endprocl

3682: 

3683: 

3684: \procl q.6

3685: Letting $A$ vary over all sets of measure 1/2, how do we get the largest

3686: and the smallest entropies for $\P^{\I A}$?

3687: More generally, given $f$, which $g$ with the same distribution as $f$

3688: maximize or minimize $H(\P^g)$?

3689: \endprocl

3690: 

3691: 

3692: \procl q.bounds 

3693: If $d=1$ and $\GM(f) \GM(\bfo - f) > 0$, do we get arbitrarily close lower

3694: bounds (in principle) on $H(\P^f)$ by the method in \ref s.dom/?

3695: How can one get close lower bounds on the entropy for higher-dimensional

3696: processes?

3697: \endprocl

3698: 

3699: \procl q.block

3700: Suppose $f : \T \to [0, 1]$ is a trigonometric polynomial $f$ of

3701: degree $m$.

3702: Then $\P^f$ is $m$-dependent, as are all $(m+1)$-block factors of

3703: independent processes (as defined by \ref b.Hed/).

3704: Is it the case that $\P^{f}$ is an $(m+1)$-block factor of an i.i.d.\ process?

3705: Erik Broman (personal communication) has shown this when $m = 1$.

3706: If one can find such block factors sufficiently explicitly for

3707: trigonometric polynomials, then one could find explicit factors of i.i.d.\

3708: processes that give any process $\P^f$. This would enable one to use more

3709: standard probabilistic techniques to study $\P^f$.

3710: More generally, in higher dimensions,

3711: if $f : \Td \to [0, 1]$ is a trigonometric polynomial, is $\P^f$ a block

3712: factor?

3713: \endprocl

3714: 

3715: 

3716: \procl q.otherPT

3717: Given an ordering $\ord$ on $\Zd$ and an increasing sequence of finite sets

3718: $\ord_n \subset \ord$ whose union is all of $\ord$, when is there phase

3719: multiplicity in the sense that 

3720: $$

3721: \lim_{n\to\infty}

3722: \P^f[\eta(\bfz) = 1 \mid \eta \restrict (\ord \setminus \ord_n) \equiv 1]

3723: \ne

3724: \lim_{n\to\infty}

3725: \P^f[\eta(\bfz) = 1 \mid \eta \restrict (\ord \setminus \ord_n) \equiv 0]

3726: \,?

3727: $$

3728: This clearly does not depend on the choice of $\ord_n$.

3729: In one dimension, this is the same as 1-sided phase multiplicity, where we know

3730: the answer.

3731: \endprocl

3732: 

3733: 

3734: \procl q.3

3735: Suppose that $d = 1$.

3736: Note that translation and flip of $f$ yield the same measure $\P^{f}$.

3737: Does $\P^{f}$ determine $f$ up to translation and

3738: flip? 

3739: \endprocl

3740: 

3741: We now ask about some properties that are important for models

3742: in statistical physics.

3743: Let $\others$ be as in \ref d.fulldom/.

3744: A stationary process $\nu$ is called {\bf quasi-local} if $\nu[\eta(\bfz) =

3745: 1 \mid \others]$ has a continuous version, where the product topology is

3746: used on $2^{\Zd \setminus \{ \bfz \}}$.

3747: It is called {\bf almost surely quasi-local} if $\nu[\eta(\bfz) =

3748: 1 \mid \others]$ has an almost surely continuous version.

3749: 

3750: \procl q.q-local

3751: For which $f$ is $\P^f$ quasi-local or almost surely quasi-local?

3752: When there is phase multiplicity, then

3753: $\P^f$ is as far as possible from this since then each version of

3754: $\P^f[\eta(\bfz) = 1 \mid \others]$ is nowhere continuous.

3755: In one dimension,

3756: there is a natural definition of one-sided quasi-locality, which is the same

3757: as what is sometimes referred to as the uniform martingale property

3758: (see \ref b.kalikow/). It is

3759: not hard to see that this property, or even its almost sure version, would

3760: imply an affirmative answer to \ref q.bounds/.

3761: \endprocl

3762: 

3763: We say that a stationary process $\nu$ on $2^\Zd$ is {\bf insertion

3764: tolerant} if $\nu[\eta(\bfz) = 1 \mid \others] > 0$ $\nu$-a.s.

3765: Similarly, $\nu$ is {\bf deletion tolerant} if $\nu[\eta(\bfz) = 0 \mid

3766: \others] > 0$ $\nu$-a.s.

3767: 

3768: \procl g.tolerance

3769: Every $\P^f$ is insertion and deletion tolerant, other than the trivial cases

3770: $f = \bfz$ and $f = \bfo$.

3771: \endprocl

3772: 

3773: \ref t.fulldom/ implies a positive solution to \ref g.tolerance/ when

3774: $1/[f(\bfo-f)]$ is integrable. Indeed, we see by \ref t.fulldom/ that this is

3775: the exact criterion for uniform insertion tolerance and uniform

3776: deletion tolerance.

3777: 

3778: \medbreak

3779: \noindent {\bf Acknowledgements.}\enspace

3780: We thank Chris Hoffman for his help with \ref p.hoffman/.

3781: We thank Eric Bedford and Jeff Geronimo for important references and Jean-Paul

3782: Thouvenot for discussions.

3783: We are grateful to David Wilson for permission to include Figures \briefref

3784: f.window-ust/ and \briefref

3785: f.window-tiling/ and to Peter Duren and J.M.~Landsberg for helpful remarks.

3786: \ref f.window-tiling/ was created using the linear algebraic techniques of

3787: \ref b.wilson:planar/ for generating domino tilings; the needed

3788: matrix inversion was

3789: accomplished using the formulas of \ref b.Kenyon:local/. The resulting tiling

3790: gives dual spanning trees by the bijection of Temperley (see \ref

3791: b.KPW:tr-mat/). One of the trees is \ref f.window-ust/.  

3792: 

3793: 

3794: \bibfile{\jobname}

3795: \def\noop#1{\relax}

3796: \input \jobname.bbl

3797: 

3798: 

3799: \filbreak

3800: \begingroup

3801: \eightpoint\sc

3802: \parindent=0pt\baselineskip=10pt

3803: 

3804: Department of Mathematics,

3805: Indiana University,

3806: Bloomington, IN 47405-5701

3807: \emailwww{rdlyons@indiana.edu}

3808: {http://php.indiana.edu/\string~rdlyons/}

3809: 

3810: and

3811: 

3812: School of Mathematics,

3813: Georgia Institute of Technology,

3814: Atlanta, GA 30332-0160

3815: \email{rdlyons@math.gatech.edu}

3816: 

3817: Department of Mathematics,

3818: Chalmers University of Technology,

3819: S-41296 Gothenburg, Sweden 

3820: \email{steif@math.chalmers.se}

3821: 

3822: and

3823: 

3824: 

3825: School of Mathematics,

3826: Georgia Institute of Technology,

3827: Atlanta, GA 30332-0160

3828: \email{steif@math.gatech.edu}

3829: 

3830: \endgroup

3831: 

3832: 

3833: \bye

3834: