1: \documentclass[11pt]{article} \usepackage{amsfonts} \usepackage{latexsym} \usepackage{graphicx}
2:
3: \newcommand{\breath}{\vspace{3mm}}
4: \newcommand{\pause}{\breath \hrulefill }
5: \newcounter{thmcount}
6:
7: \newcounter{excount}
8:
9: \newcounter{claimcount}[thmcount]
10:
11: \newcounter{subclaimcount}[claimcount]
12:
13: \newcounter{subsubclaimcount}[subclaimcount]
14:
15: \newcounter{prereqcounter}[subsubsection] \setcounter{claimcount}{0} \setcounter{subclaimcount}{0} \setcounter{subsubclaimcount}{0}
16:
17: \newcounter{myeqncount}[thmcount]
18:
19: \newsavebox{\thmstyle}
20:
21: \newfont{\fibfont}{cmfib8}
22:
23: \newcommand{\thmfont}[1]{{\sl #1}}
24: \newcommand{\Theorem}[2]{ \refstepcounter{thmcount} \subparagraph{{\normalsize Theorem \thethmcount:}} {\em #1} \savebox{\thmstyle}{{\tt Theorem}} \thmfont{ #2 }}
25: \newcommand{\Corollary}[2]{ \refstepcounter{thmcount} \subparagraph{{\normalsize Corollary \thethmcount:}} {\em #1} \savebox{\thmstyle}{{\tt Corollary}} \thmfont{ #2 }}
26: \newcommand{\Proposition}[2]{ \refstepcounter{thmcount} \subparagraph{{\normalsize Proposition \thethmcount:}} {\em #1} \savebox{\thmstyle}{{\tt Proposition}} \thmfont{ #2 }}
27: \newcommand{\Definition}[2]{ \refstepcounter{thmcount} \subparagraph{{\normalsize Definition \thethmcount:}} {\em #1} \savebox{\thmstyle}{{\tt Definition }} {\begin{list}{} {\setlength{\leftmargin}{2em} \setlength{\rightmargin}{0em}} \item \thmfont{ #2 } \end{list} } \breath}
28: \newcommand{\Exercise}{ \refstepcounter{excount} {\bf Exercise [\theexcount]}}
29: \newcommand{\Claim}[1]{\refstepcounter{claimcount} {\bf Claim \theclaimcount: \ }\thmfont{ #1}}
30: \newcommand{\claim}{\Claim}
31: \newcommand{\thmpart}[1]{{\bf Part #1}}
32: \newcommand{\Proofof}[1]{{\bf \hspace{-1em} Proof of #1: \ \ }}
33: \newcommand{\proofof}[1]{\Proofof{#1}}
34: \newenvironment{prf}[1]{\begin{list}{} {\setlength{\leftmargin}{1em} \setlength{\rightmargin}{0em}} \item {\bf \hspace{-1em} #1 \ \ }} {\end{list} }
35:
36: \newenvironment{thmproof}{ \begin{prf}{Proof:}} { {\tt \hrulefill $\Box$} \end{prf} \breath }
37:
38: \newenvironment{claimproof}{ \begin{prf}{Proof:}} { {\tt \dotfill $\Box$[Claim \theclaimcount]} \end{prf}}
39:
40: \newenvironment{subclaimproof}{ \begin{prf}{Proof:}} { {\tt \dotfill $\Box$[Claim \theclaimcount.\thesubclaimcount]} \end{prf} }
41:
42: \newenvironment{subsubclaimproof}{ \begin{prf}{Proof:}} { {\tt \dotfill $\Box$[Claim \theclaimcount.\thesubclaimcount.\thesubsubclaimcount]} \end{prf}}
43:
44: \newcommand{\Proof}{ {\bf Proof:}}
45: \newcommand{\proof}{\Proof}
46: \newcommand{\QED}{\hrulefill\ensuremath{\Box}}
47: \newcommand{\qed}{\QED}
48: \newenvironment{commenteqnarray} { \[ \begin{array}{rclr} } { \end{array} \] }
49:
50: \newcommand{\btxt}[1]{\mbox{#1}}
51: \newcommand{\etxt}{}
52: \newcommand{\beq}{\begin{eqnarray*}}
53: \newcommand{\eeq}{\end{eqnarray*}}
54: \newcommand{\beqn}{\begin{equation}}
55: \newcommand{\eeqn}{\end{equation}}
56: \newcommand{\blist}{\begin{enumerate}}
57: \newcommand{\elist}{\end{enumerate}}
58: \newcommand{\bitem}{\begin{itemize}}
59: \newcommand{\eitem}{\end{itemize}}
60: \newcommand{\bquote}{\begin{quotation}}
61: \newcommand{\equote}{\end{quotation}}
62: \newcommand{\Cesaro}{Ces\`aro }
63: \newcommand{\done}{\ensuremath{\mathsf{ 1\!\!1}}}
64: \newcommand{\dC}{\ensuremath{\mathbb{C}}}
65: \newcommand{\dG}{\ensuremath{\mathbb{G}}}
66: \newcommand{\dH}{\ensuremath{\mathbb{H}}}
67: \newcommand{\dM}{\ensuremath{\mathbb{M}}}
68: \newcommand{\dN}{\ensuremath{\mathbb{N}}}
69: \newcommand{\dP}{\ensuremath{\mathbb{P}}}
70: \newcommand{\dR}{\ensuremath{\mathbb{R}}}
71: \newcommand{\dS}{\ensuremath{\mathbb{S}}}
72: \newcommand{\dT}{\ensuremath{\mathbb{T}}}
73: \newcommand{\dU}{\ensuremath{\mathbb{U}}}
74: \newcommand{\dY}{\ensuremath{\mathbb{Y}}}
75: \newcommand{\dZ}{\ensuremath{\mathbb{Z}}}
76: \newcommand{\barb}{\ensuremath{{\bar{b}}}}
77: \newcommand{\barc}{\ensuremath{{\bar{c}}}}
78: \newcommand{\baru}{\ensuremath{{\bar{u}}}}
79: \newcommand{\barchi}{\ensuremath{{\bar{\chi }}}}
80: \newcommand{\bC}{\ensuremath{\mathbf{ C}}}
81: \newcommand{\bL}{\ensuremath{\mathbf{ L}}}
82: \newcommand{\bN}{\ensuremath{\mathbf{ N}}}
83: \newcommand{\bP}{\ensuremath{\mathbf{ P}}}
84: \newcommand{\bQ}{\ensuremath{\mathbf{ Q}}}
85: \newcommand{\bR}{\ensuremath{\mathbf{ R}}}
86: \newcommand{\bS}{\ensuremath{\mathbf{ S}}}
87: \newcommand{\bT}{\ensuremath{\mathbf{ T}}}
88: \newcommand{\bU}{\ensuremath{\mathbf{ U}}}
89: \newcommand{\bZ}{\ensuremath{\mathbf{ Z}}}
90: \newcommand{\ba}{\ensuremath{\mathbf{ a}}}
91: \newcommand{\bb}{\ensuremath{\mathbf{ b}}}
92: \newcommand{\bc}{\ensuremath{\mathbf{ c}}}
93: \newcommand{\be}{\ensuremath{\mathbf{ e}}}
94: \newcommand{\bg}{\ensuremath{\mathbf{ g}}}
95: \newcommand{\bh}{\ensuremath{\mathbf{ h}}}
96: \newcommand{\bi}{\ensuremath{\mathbf{ i}}}
97: \newcommand{\bk}{\ensuremath{\mathbf{ k}}}
98: \newcommand{\bl}{\ensuremath{\mathbf{ l}}}
99: \newcommand{\bm}{\ensuremath{\mathbf{ m}}}
100: \newcommand{\bn}{\ensuremath{\mathbf{ n}}}
101: \newcommand{\bo}{\ensuremath{\mathbf{ o}}}
102: \newcommand{\bp}{\ensuremath{\mathbf{ p}}}
103: \newcommand{\bq}{\ensuremath{\mathbf{ q}}}
104: \newcommand{\br}{\ensuremath{\mathbf{ r}}}
105: \newcommand{\bt}{\ensuremath{\mathbf{ t}}}
106: \newcommand{\bu}{\ensuremath{\mathbf{ u}}}
107: \newcommand{\bv}{\ensuremath{\mathbf{ v}}}
108: \newcommand{\bw}{\ensuremath{\mathbf{ w}}}
109: \newcommand{\bx}{\ensuremath{\mathbf{ x}}}
110: \newcommand{\by}{\ensuremath{\mathbf{ y}}}
111: \newcommand{\sA}{\ensuremath{\mathcal{ A}}}
112: \newcommand{\sB}{\ensuremath{\mathcal{ B}}}
113: \newcommand{\sC}{\ensuremath{\mathcal{ C}}}
114: \newcommand{\sE}{\ensuremath{\mathcal{ E}}}
115: \newcommand{\sF}{\ensuremath{\mathcal{ F}}}
116: \newcommand{\sI}{\ensuremath{\mathcal{ I}}}
117: \newcommand{\sL}{\ensuremath{\mathcal{ L}}}
118: \newcommand{\sM}{\ensuremath{\mathcal{ M}}}
119: \newcommand{\sN}{\ensuremath{\mathcal{ N}}}
120: \newcommand{\sP}{\ensuremath{\mathcal{ P}}}
121: \newcommand{\sR}{\ensuremath{\mathcal{ R}}}
122: \newcommand{\sS}{\ensuremath{\mathcal{ S}}}
123: \newcommand{\sT}{\ensuremath{\mathcal{ T}}}
124: \newcommand{\sU}{\ensuremath{\mathcal{ U}}}
125: \newcommand{\sV}{\ensuremath{\mathcal{ V}}}
126: \newcommand{\sW}{\ensuremath{\mathcal{ W}}}
127: \newcommand{\sX}{\ensuremath{\mathcal{ X}}}
128: \newcommand{\fra}{\ensuremath{{\fbox{a}}}}
129: \newcommand{\frb}{\ensuremath{{\fbox{b}}}}
130: \newcommand{\bre}{\ensuremath{\overline{\underline{e}}}}
131: \newcommand{\gA}{\ensuremath{\mathfrak{ A}}}
132: \newcommand{\gB}{\ensuremath{\mathfrak{ B}}}
133: \newcommand{\gH}{\ensuremath{\mathfrak{ H}}}
134: \newcommand{\gR}{\ensuremath{\mathfrak{ R}}}
135: \newcommand{\gT}{\ensuremath{\mathfrak{ T}}}
136: \newcommand{\gV}{\ensuremath{\mathfrak{ V}}}
137: \newcommand{\gW}{\ensuremath{\mathfrak{ W}}}
138: \newcommand{\ga}{\ensuremath{\mathfrak{ a}}}
139: \newcommand{\go}{\ensuremath{\mathfrak{ o}}}
140: \newcommand{\gr}{\ensuremath{\mathfrak{ r}}}
141: \newcommand{\alp }{\ensuremath{\alpha}}
142: \newcommand{\bet }{\ensuremath{\beta}}
143: \newcommand{\gam }{\ensuremath{\gamma}}
144: \newcommand{\del }{\ensuremath{\delta}}
145: \newcommand{\eps }{\ensuremath{\epsilon}}
146: \newcommand{\kap }{\ensuremath{\kappa}}
147: \newcommand{\h}[1]{{\ensuremath{\widehat{#1}}}}
148: \newcommand{\hT}{\ensuremath{{\widehat{T}}}}
149: \newcommand{\hX}{\ensuremath{{\widehat{X}}}}
150: \newcommand{\ha}{\ensuremath{{\widehat{a}}}}
151: \newcommand{\hb}{\ensuremath{{\widehat{b}}}}
152: \newcommand{\hd}{\ensuremath{{\widehat{d}}}}
153: \newcommand{\hg}{\ensuremath{{\widehat{g}}}}
154: \newcommand{\hl}{\ensuremath{{\widehat{l}}}}
155: \newcommand{\hm}{\ensuremath{{\widehat{m}}}}
156: \newcommand{\hn}{\ensuremath{{\widehat{n}}}}
157: \newcommand{\hr}{\ensuremath{{\widehat{r}}}}
158: \newcommand{\hs}{\ensuremath{{\widehat{s}}}}
159: \newcommand{\hsX}{\ensuremath{{\widehat{\mathcal{ X}}}}}
160: \newcommand{\hdel }{\ensuremath{\widehat{\delta}}}
161: \newcommand{\hmu }{\ensuremath{\widehat{\mu}}}
162: \newcommand{\hnu }{\ensuremath{\widehat{\nu}}}
163: \newcommand{\nvb}{\ensuremath{{b^{-1}}}}
164: \newcommand{\prpb}{\ensuremath{{b^\perp}}}
165: \newcommand{\fA}{\ensuremath{\mathsf{ A}}}
166: \newcommand{\fB}{\ensuremath{\mathsf{ B}}}
167: \newcommand{\fC}{\ensuremath{\mathsf{ C}}}
168: \newcommand{\fa}{\ensuremath{\mathsf{ a}}}
169: \newcommand{\fb}{\ensuremath{\mathsf{ b}}}
170: \newcommand{\fc}{\ensuremath{\mathsf{ c}}}
171: \newcommand{\fd}{\ensuremath{\mathsf{ d}}}
172: \newcommand{\fo}{\ensuremath{\mathsf{ o}}}
173: \newcommand{\fr}{\ensuremath{\mathsf{ r}}}
174: \newcommand{\tl}[1]{\ensuremath{{\widetilde{#1}}}}
175: \newcommand{\tlS}{\ensuremath{{\widetilde{S}}}}
176: \newcommand{\tla}{\ensuremath{{\widetilde{a}}}}
177: \newcommand{\tlb}{\ensuremath{{\widetilde{b}}}}
178: \newcommand{\tlc}{\ensuremath{{\widetilde{c}}}}
179: \newcommand{\tlg}{\ensuremath{{\widetilde{g}}}}
180: \newcommand{\tlm}{\ensuremath{{\widetilde{m}}}}
181: \newcommand{\tln}{\ensuremath{{\widetilde{n}}}}
182: \newcommand{\tls}{\ensuremath{{\widetilde{s}}}}
183: \newcommand{\tlu}{\ensuremath{{\widetilde{u}}}}
184: \newcommand{\tlbC}{\ensuremath{{\widetilde{\mathbf{ C}}}}}
185: \newcommand{\tlba}{\ensuremath{{\widetilde{\mathbf{ a}}}}}
186: \newcommand{\tlbc}{\ensuremath{{\widetilde{\mathbf{ c}}}}}
187: \newcommand{\tlbg}{\ensuremath{{\widetilde{\mathbf{ g}}}}}
188: \newcommand{\tlsM}{\ensuremath{{\widetilde{\mathcal{ M}}}}}
189: \newcommand{\tlsU}{\ensuremath{{\widetilde{\mathcal{ U}}}}}
190: \newcommand{\tlsV}{\ensuremath{{\widetilde{\mathcal{ V}}}}}
191: \newcommand{\tlmu }{\ensuremath{{\widetilde{\mu}}}}
192: \newcommand{\tlnu }{\ensuremath{{\widetilde{\nu}}}}
193: \newcommand{\undb}{\ensuremath{{\underline{b}}}}
194: \newcommand{\unde}{\ensuremath{{\underline{e}}}}
195: \newcommand{\undu}{\ensuremath{{\underline{u}}}}
196: \newcommand{\va}{\ensuremath{{\vec{a}}}}
197: \newcommand{\vb}{\ensuremath{{\vec{b}}}}
198: \newcommand{\vd}{\ensuremath{{\vec{d}}}}
199: \newcommand{\ve}{\ensuremath{{\vec{e}}}}
200: \newcommand{\vs}{\ensuremath{{\vec{s}}}}
201: \newcommand{\cvb}{\ensuremath{{\covec{b}}}}
202: \newcommand{\lb}{\ensuremath{\left}}
203: \newcommand{\rb}{\ensuremath{\right}}
204: \newcommand{\sub}[2]{ \raisebox{-0.6ex} { $\stackrel{#1}{{\scriptscriptstyle #2}}$ } }
205: \newcommand{\maketall}{\rule[-0.5cm]{0cm}{1cm}}
206: \newcommand{\implies}{\mbox{$\Longrightarrow$}}
207: \newcommand{\map}{\ensuremath{\longrightarrow}}
208: \newcommand{\maps}{\ensuremath{\longrightarrow}}
209: \newcommand{\goto}{\ensuremath{\rightarrow}}
210: \newcommand{\into}{\ensuremath{\map}}
211: \newcommand{\seilpmi}{\ensuremath{\Longleftarrow}}
212: \newcommand{\statement}[1]{\lb( \ \maketall \begin{minipage}{40em} \begin{tabbing} #1 \end{tabbing} \end{minipage} \ \rb)}
213: \newcommand{\oo}{\ensuremath{\infty}}
214: \newcommand{\X}{\ensuremath{\times}}
215: \newcommand{\x}{\ensuremath{\X}}
216: \newcommand{\tensor}{\ensuremath{\otimes}}
217: \newcommand{\Tensor}{\ensuremath{\bigotimes}}
218: \newcommand{\dirsum}{\ensuremath{\oplus}}
219: \newcommand{\intsct}{\ensuremath{\cap}}
220: \newcommand{\Intsct}{\ensuremath{\bigcap}}
221: \newcommand{\disj}{\ensuremath{\sqcup}}
222: \newcommand{\Disj}{\ensuremath{\bigsqcup}}
223: \newcommand{\set}[2]{\ensuremath{\left\{ #1 \; ; \; #2 \right\} }}
224: \newcommand{\seq}[2]{\ensuremath{ \lb\{#1 |_{_{{#2}}} \rb\} }}
225: \newcommand{\supp}[1]{ \ensuremath{\mathbf{ supp}\lb[#1\rb]}}
226: \newcommand{\norm}[2]{\ensuremath{\left\| #1 \right\|_{{#2}} } }
227: \newcommand{\inn}[1]{\ensuremath{\left\langle #1 \right\rangle }}
228: \newcommand{\Id}[1]{\ensuremath{\mathbf{ Id}_{{#1}}}}
229: \newcommand{\pr}[1]{\ensuremath{\mathbf{ pr}_{{#1}}}}
230: \newcommand{\inc}[1]{\ensuremath{\mathbf{ inc}_{{#1}}}}
231: \newcommand{\chr}[1]{\ensuremath{{\done}_{{#1}}}}
232: \newcommand{\choice}[1]{\ensuremath{\lb\{ \begin{array}{lcr} #1 \end{array} \rb. }}
233: \newcommand{\cl}[1]{\ensuremath{{\bullet\!\!-#1-\!\!\bullet}}}
234: \newcommand{\op}[1]{\ensuremath{{\leftarrow#1\rightarrow}}}
235: \newcommand{\interior}[1]{\ensuremath{\bi\bn\bt\lb[ #1 \rb]}}
236: \newcommand{\shift}[1]{ {{\sS}_{^{\!\!h\!i\!f\!t}}^{#1}} }
237: \newsavebox{\gaussopen} \savebox{\gaussopen}{\begin{picture}(5,5) \put(0,5){\oval(12,12)[br]} \put(12,4){\oval(12,12)[tl]} \end{picture} }
238:
239: \newsavebox{\gaussclose} \savebox{\gaussclose}{\begin{picture}(5,5) \put(9,5){\oval(12,12)[bl]} \put(-3,4){\oval(12,12)[tr]} \end{picture} }
240:
241: \newcommand{\re}[1]{\ensuremath{{\br\be\lb[#1\rb]}}}
242: \newcommand{\im}[1]{\ensuremath{{\bi\bm\lb[#1\rb]}}}
243: \newcommand{\Ces}[1]{\ensuremath{\sum\!\!\!-\!\!\lim_{#1 \goto \oo}}}
244: \newcommand{\Lbsg}{{\ensuremath{\sL^{_{\!b\!s\!g}}}}}
245: \newcommand{\card}[1]{\ensuremath{{\sC_{^{\!\!a\!r\!d}}\lb[#1\rb]}}}
246: \newcommand{\Meas}[1]{\ensuremath{\sM_{^{\!\sE\!\!\sA\!\sS}} \lb[#1\rb] }}
247: \newcommand{\invMeas}[2]{\ensuremath{\sM^{#1}_{^{\!\sE\!\!\sA\!\sS}} \lb[#2\rb] }}
248: \newcommand{\marg}[2]{\ensuremath{\pr{#1}^* \lb[#2\rb]}}
249: \newcommand{\Four}{\ensuremath{{\sF_{^{\!\!\!o\!u\!r}}}}}
250: \newcommand{\SU}[1]{{\ensuremath{\dS\dU^{\,#1} }}}
251: \newcommand{\U}[1]{{\ensuremath{\dU^{\,#1} }}}
252: \newcommand{\M}[1]{\fbox{\(#1\)}\,}
253: \newcommand{\mtrx}[3]{\ensuremath{\lb[#1 |_{#2}^{#3} \rb]}}
254: \newcommand{\covec}[1]{\ensuremath{{\vec{#1}^{\!*}}}}
255: \newcommand{\Real}{\dR}
256: \newcommand{\Natur}{\dN}
257: \newcommand{\Zahl}{\dZ}
258: \newcommand{\Cplx}{\dC}
259: \newcommand{\Torus}[1]{\ensuremath{{\dT}^{#1}}} \setlength{\unitlength}{0.3mm}
260: \newsavebox{\cubebox} \savebox{\cubebox}{\begin{picture}(12,8) \put(0,-5){\framebox(10,10)} \put(3,-2){\framebox(10,10)} \end{picture} }
261:
262: \newsavebox{\ballbox} \savebox{\ballbox}{\begin{picture}(9,8) \put(5,3){\circle{15}} \put(5,3){\circle{12}} \end{picture} }
263:
264: \newsavebox{\cycopen} \savebox{\cycopen}{\begin{picture}(5,5) \put(0,4){\oval(12,12)[l]} \put(1,-2){\vector(1,0){5}} \end{picture} }
265:
266: \newsavebox{\cycclose} \savebox{\cycclose}{\begin{picture}(5,5) \put(9,4){\oval(12,12)[r]} \put(7,10){\vector(-1,0){5}} \end{picture} }
267:
268: \newcommand{\CC}[1]{\ensuremath{{\lb[ #1 \rb]}}}
269: \newcommand{\CO}[1]{\ensuremath{{\lb[ #1 \rb)}}}
270: \newcommand{\pb}[1]{#1^{\!\nwarrow}}
271: \newlength{\oldlength}
272:
273: \newcommand{\tlgW}{\widetilde{\gW}}
274: \newcommand{\Latt}{{\Zahl^D}}
275: \newcommand{\statMeas}[1]{\invMeas{s\!t\!a\!t}{#1}}
276: \newcommand{\extMeas}[1]{\invMeas{e\!x\!t}{#1}}
277: \newcommand{\ergMeas}[1]{\invMeas{e\!r\!g}{#1}}
278: \newcommand{\dual}[1]{\widehat{\gA^{#1}}}
279: \newcommand{\hgA}{\widehat{\gA}}
280: \newcommand{\Freq}[3]{{\bf Freq}\lb[ #1 \subset #2; \ #3\rb] }
281: \newcommand{\Fin}[1]{\sF_{\!i\!n}\lb[#1\rb]}
282: \newcommand{\tilebox}[2]{\fbox{$\begin{array}{#1} #2 \end{array}$}}
283: \newcommand{\connect}{\!\!\!\!\!\longleftrightarrow\!\!\!\!\!\!}
284: \newcommand{\minitilebox}[4]{\fbox{${^{#1}_{#3} \; ^{#2}_{#4} }$}}
285: \newcommand{\B}{\bullet}
286: \newcommand{\RReal}{\Real_\dagger}
287: \newcommand{\RMeas}[1]{{\sM_{^{\!\sE\!\!\sA\!\sS}\dagger} \lb[#1\rb] }}
288: \newcommand{\RinvMeas}[2]{{\sM^{#1}_{^{\!\sE\!\!\sA\!\sS}\dagger} \lb[#2\rb] }}
289: \newcommand{\RstatMeas}[1]{\RinvMeas{s\!t\!a\!t}{#1}}
290: \newcommand{\RextMeas}[1]{\RinvMeas{e\!x\!t}{#1}}
291:
292:
293: \begin{document}
294:
295: \title{Building a Stationary Stochastic Process from a Finite-dimensional
296: Marginal}
297:
298: \author{Marcus Pivato}
299:
300: \maketitle
301:
302:
303: \abstract{If $\gA$ is a finite alphabet, $\sU \subset \Latt$,
304: and $\mu_\sU$ is a probability measure on $\gA^\sU$ that ``looks like''
305: the marginal projection of a stationary stochastic process on $\gA^\Latt$,
306: then can we ``extend'' $\mu_\sU$ to such a process? Under what conditions
307: can we make this extension ergodic, (quasi)periodic, or (weakly) mixing?
308: After surveying classical work on this problem when $D = 1$,
309: we provide some sufficient conditions and some necessary conditions for
310: $\mu_\sU$ to be extendible for $D > 1$, and show that, in general, the
311: problem is not formally decidable.}
312:
313: \subsubsection*{Mathematics Subject Classification Number:}
314:
315: \subparagraph{Primary:} 37A50, 60G10 (Ergodic Theory of Stationary Stochastic Processes)
316: \subparagraph{Secondary:} 37B10 (Symbolic Dynamics)
317:
318: \newpage
319:
320:
321: \section{Introduction}
322:
323: \subsection{The Markov Extension in $\Zahl$}
324:
325: Let $\gA$ be a finite alphabet, and let $\gA^\Zahl$ be the space of
326: {\em bi-infinite sequences} on $\gA$. A {\bf stationary stochastic
327: process} is a probability measure $\mu$ on $\gA^\Zahl$ so that,
328: for any $V \in \Natur, \ \ b_0,b_1,\ldots,b_V \in \gA$, and any $k \in \Zahl$
329:
330: \beq
331: \lefteqn{
332: \mu\set{\ba \in \gA^\Zahl}{a_0 = b_0,\ldots,a_V = b_V }
333: }\\
334: &\hspace{5em} = &
335: \mu\set{\ba \in \gA^\Zahl}{a_k = b_0,\ldots,a_{k+V} = b_{V} } \\
336: \eeq
337:
338: Let $\sU$ be the interval $\CC{0...U} \subset \Zahl$. The {\bf projection
339: map} $\pr{\sU} : \gA^\Zahl \into \gA^\sU$ is the map sending the
340: sequence $\mtrx{a_n}{n \in \Zahl}{}$ to the sequence $\mtrx{a_n}{n \in \sU}{}$.
341: With this map, we can project $\mu$ down to a {\bf marginal} measure,
342: $\mu_\sU \ := \ \marg{\sU} \mu$, on the space $\gA^\sU$. This marginal
343: is then {\bf locally stationary}: for any $V < U$, any
344: $b_0,b_1,\ldots,b_V \in \gA$, and any $k \in \Zahl$ so that $V + k \leq U$
345: also,
346:
347: \beq
348: \lefteqn{
349: \mu_\sU\set{\ba \in \gA^\sU}{a_0 = b_0,\ldots,a_V = b_V }
350: }\\
351: &\hspace{5em} = &
352: \mu_\sU\set{\ba \in \gA^\sU}{a_k = b_0,\ldots,a_{k+V} = b_{V} } \\
353: \eeq
354:
355: Can we {\em reverse} this process? Given a locally stationary
356: measure $\mu_\sU$ upon $\gA^\sU$, can we {\bf extend} it to a stationary
357: stochastic process $\mu$ on $\gA^\Zahl$, so that
358: $\marg{\sU} \mu \ = \ \mu_\sU$? Yes, and furthermore, we can do so
359: in a canonical fashion, via the so-called {\bf Markov Extension}.
360:
361: An intuitive description of the Markov Extension is this:
362: We randomly ``choose'' the coordinates $a_0,\ldots,a_U$ according
363: to the probability measure $\mu_\sU$. We then randomly chose the
364: coordinate $a_{U+1}$, again according to $\mu_\sU$ (now treated
365: as a probability measure on $\gA^{\sU+1}$), but {\em conditioned}
366: upon the fact that we have already fixed coordinates $a_1,\ldots,a_U$.
367: Next, we randomly chose the
368: coordinate $a_{U+2}$, again according to $\mu_\sU$ (now treated
369: as a probability measure on $\gA^{\sU+2}$), but {\em conditioned}
370: upon the fact that we have already fixed coordinates $a_2,\ldots,a_{U+1}$.
371: Inductively, we get a $U-$step Markov process
372: on $\gA$.
373:
374: To formally construct the Markov Extension, we need a bit
375: of notation:
376:
377: \bitem
378: \item If $\ba \ = \ \mtrx{a_n}{n\in\Zahl}{}$ is an element of $\gA^\Zahl$,
379: and $\sV \subset \Zahl$, then let $\ba_\sV \ := \ \mtrx{a_v}{v \in \sV}{}$
380:
381: \item If $\mu$ is a measure upon $\gA^\Zahl, \ \sV \subset \Zahl$,
382: and $\bb \in \gA^\sV$, then let ``$\mu[\bb]$'' denote the measure
383: of the associated cylinder set:
384:
385: \[ \mu[\bb] \ := \ \mu\set{\ba \in \gA^\Zahl}{\ba_\sV \ = \ \bb } \]
386:
387: \item Suppose $\sV \subset \Zahl$ and $k \in
388: \Zahl$ are such that $(k + \sV) \ \subset \ \sU$. If $\bb \ := \ \mtrx{b_v}{v \in \sV}{}$
389: is an element of $\gA^\sV$, then let $\bb'$ be the ``shift'' of $\bb$
390: by $k$: that is, $\bb' \ := \ \mtrx{b'_v}{v \in (k+\sV)}{}$, where,
391: for all $v \in \sV, \ \ b'_v \ = \ b_{v-k}$. Then define:
392:
393: \[ \mu_\sU[\bb] \ := \
394: \mu_\sU\set{\ba \in \gA^\sU}{\ba_{(k + \sV)} \ = \ \bb'} \]
395:
396: (because $\mu_\sU$ is locally stationary, it doesn't matter which $k$ we use
397: in this definition, if more than one $k$ is available)
398:
399: \eitem
400:
401: The {\bf Markov Extension} of $\mu_\sU$ is the probability
402: measure $\mu_{m\!r\!k}$, where, for any $N \geq U$, and $\bb \in \gA^\CC{0..N}$,
403:
404: \[ \mu_{m\!r\!k}[\bb] \ := \ \mu_\sU\lb[\bb_\sU\rb] \cdot
405: \prod_{k=1}^{N-U}
406: \mu_\sU\lb[\frac{\bb_\CC{k...U\!+\!k}}{ \bb_\CO{k...U\!+\!k} }\rb] \]
407:
408: Here, $\CO{k...U\!+\!k} \ := \ \{k,\ k\!+\!1,\ \ldots,\ k\!+\!U\!-\!1\}$, while
409: $\CC{k...U\!+\!k} \ := \ \{k,\ k\!+\!1,\ \ldots,\ k\!+\!U\}$, and
410: $\mu_\sU\lb[\frac{\bb_\CC{k...U\!+\!k} }{ \bb_\CO{k...U\!+\!k} }\rb]$ is
411: the {\bf conditional probability}:
412:
413: \[ \mu_\sU\lb[\frac{\bb_\CC{k...U+k}}{ \bb_\CO{k...U+k} }\rb]
414: \ \ := \ \ \frac{ \mu_\sU\lb[\bb_\CC{k...U+k} \rb] }
415: { \mu_\sU\lb[\bb_\CO{k...U+k} \rb] } \]
416:
417: $\mu_{m\!r\!k}$ is a stationary probability measure on $\gA^\Natur$.
418: Define the probabilities of cylinder sets indexed by negative
419: coordinates by simply {\em shifting} them into the positive
420: domain. Thus, $\mu_{m\!r\!k}$ is defined on all cylinder sets in $\gA^\Zahl$.
421: It is straightforward to check that the probability measure thus
422: defined is stationary, and that its marginal projection upon $\gA^\sU$
423: is equal to $\mu_\sU$.
424:
425: \breath
426:
427: This construction indicates that a stationary extension of the
428: measure $\mu_\sU$ always exists. In general, there may be many such
429: extensions. Intuitively, $\mu_{m\!r\!k}$ is an extension built so as
430: to provide the maximum amount of ``random choice'' at each successive
431: coordinate. Hence, the following variational principle is not too
432: surprising:
433:
434: \subparagraph{Theorem:} {\tt Maximal Entropy Property}
435:
436: {\em
437: Of all the different stationary extensions of $\mu_\sU$ that exist,
438: $\mu_{m\!r\!k}$ is the one possessing the largest {\bf process entropy},
439: which we define as:
440:
441: \[ H\lb(\mu_{m\!r\!k}\rb) \ \ := \ \
442: \lim_{N \goto \oo} \frac{-1}{N} \sum_{\ba \in \gA^\CC{1...N}}
443: \mu_{m\!r\!k}[\ba] \log_2 \lb(\mu_{m\!r\!k}[\ba]\rb) \]
444: }
445: \Proof \ \ \ See, for example, \cite{Schlijper}. \qed
446:
447:
448: \breath
449:
450: Under what circumstances do {\em ergodic} extensions of $\mu_\sU$
451: exist? Can we build an extension measure which is supported only on
452: {\em periodic} words of some fixed periodicity? Also, what happens if
453: $\sU$ is {\em not} just an interval inside $\Zahl$?
454:
455: \subsection{Extension on Lattices}
456:
457: Now, let $D > 0$, and let $\Zahl^D$ be a $D-$dimensional lattice.
458: Then $\gA^{\Zahl^D}$ is the space of
459: {\em $D-$dimensional configurations} on $\gA$. If $\bk \in \Zahl^D$,
460: then the {\bf shift by $\bk$} is the map $\shift{\bk}:\gA^{\Zahl^D} \into
461: \gA^{\Zahl^D}$ so that, if $\ba := \mtrx{a_\bn}{\bn \in \Zahl^D}{}$,
462: then $\shift{\bk} \ba \ \ := \ \mtrx{a'_\bn}{\bn \in \Zahl^D}{}$,
463: where $a'_\bn \ = \ a_{\bn-\bk}, \ \ \forall \bn \in \Zahl^D$.
464: \label{shift.defn}
465:
466: A {\bf stationary stochastic
467: process} is a probability measure $\mu$ on $\gA^{\Zahl^D}$ that
468: is {\bf invariant} under all shift maps. That is, if
469: $\sV \subset \Zahl^D$ is any finite subset, and $\bb \in \gA^\sV$,
470: then for any $\bk \in \Zahl^D$,
471:
472: \[ \mu\lb[\shift{\bk}(\bb) \rb] \ \ = \ \ \mu[\bb] \]
473:
474: If $\sU \subset \Zahl^D$, and $\bk \in \Zahl^D$,
475: then define $\shift{\bk}\sU \ = \ \sU+ \bk$, and
476: define $\shift{\bk}:\gA^{\sU} \into
477: \gA^{\sU+\bk}$ so that, if $\ba := \mtrx{a_\bn}{\bn \in \sU}{}$,
478: then $\shift{\bk} \ba \ \ := \ \mtrx{a'_\bn}{\bn \in \sU+\bk}{}$,
479: where $a'_\bn \ = \ a_{\bn-\bk}, \ \ \forall \bn \in \sU+\bk$.
480: A probability measure $\mu_\sU$ on $\gA^\sU$
481: is {\bf locally stationary} if for any $\sV \subset \sU$, any
482: $\bb \in \gA^\sV$, and any $\bk \in \Zahl^D$ so that $\shift{\bk}{\sV} \subset \sU$
483: also,
484:
485: \[ \mu_\sU\lb[\shift{\bk}(\bb) \rb] \ \ = \ \ \mu_\sU[\bb] \]
486:
487:
488: \subparagraph{The Extension Problem:}
489: {\em Given a locally stationary
490: measure $\mu_\sU$ upon $\gA^\sU$, can we {\bf extend} it to a stationary
491: stochastic process $\mu$ on $\gA^{\Zahl^D}$, so that $\marg{\sU}{\mu} \ \ = \ \ \mu_\sU$?}
492:
493: \breath
494:
495: The Extension Problem does not always have solutions, as examples in
496: Section \ref{counter} will show. If we {\em can} solve the Extension
497: Problem, can we construct an extension which is ergodic? (quasi)
498: Periodic or (weakly) mixing?
499:
500:
501:
502:
503:
504: \subsection{Extension on Group Modules}
505:
506: Now, let $\dG$ be an arbitrary group, and let $\sM$ be a {\bf
507: $\dG$-module}: \ an arbitrary set equipped with a $\dG-$action.
508: A few examples of this to keep in mind:
509:
510: \bitem
511: \item $\sM \ := \ \Zahl^D$ and $\dG \ := \ \Zahl^D$, also, acting upon $\sM$
512: by {\bf translation}.
513:
514: \item $\sM \ := \ (\Zahl/P_1) \dirsum (\Zahl/P_2) \dirsum \ldots
515: (\Zahl/P_D)$, and $\dG \ := \ \Zahl^D$ acts upon $\sM$ by {\bf translation}
516: with periodic boundary conditions.
517:
518: \item $\dG$ is an arbitrary group, $\dH$ an arbitrary subgroup, and
519: $\sM \ := \ \dG/\dH$ is the set of {\bf right cosets}. $\dG$ acts upon
520: $\sM$ by multiplication: if $\bg \in \dG$ and $(\bk\dH) \in \sM$, then
521: $\bg.(\bk\dH) \ := \ (\bg.\bk)\dH$. (Every {\bf transitive}
522: $\dG-$module is of this type, and {\em every} $\dG-$module can be written
523: as a disjoint union of transitive $\dG$-modules.)
524: \eitem
525:
526: Let $\gA^\sM$ be the space of
527: {\em $\sM-$indexed configurations} on $\gA$. If $\bg \in \dG$
528: then the {\bf shift by $\bg$} is the map $\shift{\bg}:\gA^{\sM} \into
529: \gA^{\sM}$ so that, if $\ba := \mtrx{a_m}{m \in \sM}{}$,
530: then $\shift{\bg} \ba \ \ := \ \mtrx{a'_{m}}{m \in \sM}{}$,
531: where $a'_m \ = \ a_{\bg^{-1}.m}, \ \forall m \in \sM$.
532:
533: A {\bf $\dG$-invariant stochastic
534: process} is a probability measure $\mu$ on $\gA^{\sM}$ that
535: is {\bf invariant} under the shift action of $\dG$. That is, if
536: $\sV \subset \sM$ is any finite subset, and $\bb \in \gA^\sV$,
537: then for any $\bg \in \dG$,
538:
539: \[ \mu\lb[\shift{\bg}(\bb) \rb] \ \ = \ \ \mu[\bb] \]
540:
541: If $\sU \subset \sM$ and $\bg \in \dG$, then define
542: $\shift{\bg}\sU \ = \ \bg.\sU \ = \ \set{\bg.u}{u\in\sU}$,
543: and define $\shift{\bg}:\gA^{\sU} \into
544: \gA^{\bg.\sU}$ so that, if $\ba := \mtrx{a_u}{u \in \sU}{}$,
545: then $\shift{\bg} \ba \ \ := \ \mtrx{a'_{u}}{u \in \bg.\sU}{}$,
546: where $a'_u \ = \ a_{\bg^{-1}.u}, \ \forall u \in \bg.\sU$.
547: A probability measure $\mu_\sU$ on $\gA^\sU$
548: is {\bf locally stationary} if for any $\sV$ subset $\sU$, any
549: $\bb \in \gA^\sV$, and any $\bg \in \dG$ so that $\shift{\bg}{\sV} \subset \sU$
550: also,
551:
552: \[ \mu_\sU\lb[\shift{\bg}(\bb) \rb] \ \ = \ \ \mu_\sU[\bb] \]
553:
554: Again, we ask:
555:
556: \subparagraph{The (group module) Extension Problem:}
557: {\em Given a locally stationary
558: measure $\mu_\sU$ upon $\gA^\sU$, can we {\bf extend} it to a stationary
559: stochastic process $\mu$ on $\gA^{\sM}$, so that $\marg{\sU}{\mu} \ \ = \ \ \mu_\sU$?}
560:
561: \breath
562:
563: If $\sM = \Zahl^D = \dG$, then this is just the Extension Problem on a
564: $D-$dimensional lattice. If $\sM := (\Zahl/P_1) \dirsum (\Zahl/P_2)
565: \dirsum \ldots (\Zahl/P_D)$ and $\dG := \Zahl^D$, then a $\dG-$invariant
566: measure on $\gA^\sM$ is ``equivalent'' to a stationary stochastic
567: process on $\gA^{\Zahl^D}$ which is supported only on {\em periodic}
568: configurations with fundamental domain $\CO{0...P_1} \x
569: \CO{0...P_2} \x \ldots \x \CO{0...P_D}$. In Section \ref{envelope}, we will
570: demonstrate that, if $\sU \subset
571: \CO{0...P_1} \x
572: \CO{0...P_2} \x \ldots \x \CO{0...P_D} \subset \Zahl^D$ is some
573: ``small enough'' domain, then any locally stationary measure
574: $\mu_\sU$ can be identified with a locally invariant measure $\mu_{\sU'}$,
575: where $\sU' \subset \sM$ is the obvious ``representation'' of $\sU$
576: inside $\sM$.
577:
578: \subsection{Organization of this paper}
579:
580: In \S\ref{sect.appl}, we motivate the Extension Problem by
581: discussing applications to the {\bf Invariant Measure Problem} for
582: subshifts of finite type and cellular automata. In \S
583: \ref{counter}, we show that the Extension Problem is not trivial
584: by providing examples of locally stationary measures which {\em
585: cannot} be extended. These examples imply two necessary
586: conditions for extendibility: the {\bf Entropy Condition} and the {\bf
587: Tiling Condition}.
588:
589: In \S\ref{sect.harm}, we review basic harmonic analysis on
590: configuration space, treating it as a compact abelian group, and
591: characterise the Extension Problem in terms of
592: constructing a suitable set of Fourier coefficients. We use this
593: in \S\ref{finite}, where we consider extension on {\em finite}
594: $\dG$-modules, and show that, if $\nu$ is an extendible measure with
595: full support, and $\mu$ is ``close enough'' to $\nu$, then $\mu$ is
596: also extendible. A similar result can be developed for
597: constructing {\em periodic} extensions, but first we need a
598: tool to ``reduce'' the Extension Problem on an infinite
599: module to an extension problem on a suitably chosen finite module,
600: which we develop in \S\ref{envelope}, via the concept of ``envelopes''.
601:
602: In \S\ref{embed}, we show that an extendible, locally
603: stationary measure with full support can be ``embedded'' in any
604: ergodic $\Latt$-dynamical system, in the sense that it is a marginal
605: projection of a stationary $\Latt$-process generated by a partition on
606: that system.
607:
608: In \S\ref{sect.per}, we combine the results of \S\ref{finite} and
609: \S\ref{envelope} to investigate when a measure has an
610: almost-surely periodic extension, and provide examples of measures
611: which {\em never} have periodic extensions, as well as measures which
612: {\em only} have periodic extensions. Then we use the results of
613: \S\ref{embed} to show that ``almost all'' extendible measures
614: have extensions which are ergodic, mixing, weakly mixing, or
615: quasiperiodic.
616:
617: In \S \ref{decide}, we show that the Extension Problem is,
618: in general, formally undecidable.
619:
620:
621: \subsection{Preliminaries and Notation}
622:
623: If we treat $\gA$ as a {\bf discrete} topological space, and endow
624: $\gA^\sM$ with the Tychonoff product topology, then $\gA^\sM$ is a
625: compact, metrizable space. If $\sM$ is finite, then $\gA^\sM$ is
626: finite and discrete. If $\sM$ is infinite, then $\gA^\sM$ is
627: uncountable and {\bf totally disconnected}.
628:
629: The topology on $\gA^\sM$ is generated by
630: {\bf cylinder sets}. If $\sU \subset \sM$ is finite, and $\bb \in \gA^\sU$,
631: then the associated cylinder set is:
632:
633: \[ \set{\ba \in \gA^\sM}{\ba_\sU \ = \ \bb } \]
634:
635: Here, by ``$\ba_\sU$'' we mean the element $\mtrx{a_u}{u \in \sU}{}$,
636: where $\ba = \mtrx{a_m}{m \in \sM}{}$. Normally, we will use the
637: symbol ``$\bb$'' to denote both the word $\bb$ and the cylinder set
638: it induces ---the distinction will be clear from context. For example,
639: if $\mu$ is some measure, then ``$\mu[\bb]$'' indicates the measure
640: of the cylinder set defined by $\bb$.
641:
642: Whenever we speak of measures on $\gA^\sM$, we will mean measures
643: on the {\bf Borel sigma-algebra} generated by the product topology.
644:
645: \breath
646:
647: If $\sM$ is a $\dG-$module, then $\invMeas{\dG}{\gA^\sM}$ is the space of all
648: $\dG-$invariant probability measures on $\gA^\sM$. This is a convex
649: subset of $\Meas{\gA^\sM}$, the space of {\em all} {\bf probability
650: measures} on $\gA^\sM$, which, in turn, is a convex subset of
651: the real vector space $\Meas{\gA^\sM; \ \Real}$ of {\bf real-valued
652: measures} on $\gA^\sM$.
653:
654: The elements of $\Meas{\gA^\sM; \ \Cplx}$ ({\bf complex-valued
655: measures} on $\gA^\sM$) act as linear functionals on $\bC(\gA^\sM; \ \Cplx)$
656: (the Banach space of {\bf complex-valued, continuous functions}).
657: This induces a {\bf weak$-*$ topology} on
658: $\Meas{\gA^\sM; \ \Cplx}$, making it into a locally
659: convex topological vector space.
660:
661: $\invMeas{\dG}{\gA^\sM}$ is a compact subset of $\Meas{\gA^\sM; \ \Cplx}$
662: under this topology.
663:
664: When $\dG = \sM = \Latt$, we will refer to $\invMeas{\dG}{\gA^\sM}$ as
665: ``$\statMeas{\gA^\Latt}$''.
666:
667: \breath
668:
669: If $\sU \subset \sM$, then $\invMeas{\dG}{\gA^\sU}$ is the space of all
670: locally $\dG-$invariant probability measures on $\gA^\sU$.
671: $\extMeas{\gA^\sU}$ is the set of all {\bf extendable probability measures}:
672: measures which can be extended to a $\dG-$invariant measure on $\gA^\sM$.
673: Notice that:
674:
675: \begin{quote}
676: {\em $\extMeas{\gA^\sU}$ is a compact, convex subset of
677: $\Meas{\gA^\sU; \Cplx}$.}
678: \end{quote}
679:
680: This is because
681: the marginal projection map $\pr{\sU}^*:\Meas{\gA^\sM; \ \Cplx} \into
682: \Meas{\gA^\sU; \ \Cplx}$ is {\bf linear} and {\bf continuous},
683: and $\extMeas{\gA^\sU}$ is simply the image of the compact,
684: convex subset $\invMeas{\dG}{\gA^\sM}$ under $\pr{\sU}^*$.
685:
686:
687:
688:
689:
690:
691:
692:
693: \section{Applications \label{sect.appl}}
694:
695: \subsection{Subshifts of Finite Type \label{finite.type}}
696:
697: Let $\sU \subset \Latt$ be finite, and suppose that $\gW \subset \gA^\sU$ is some
698: set of ``admissible'' $\sU$-words. The {\bf subshift of finite type}
699: defined by $\gW$ is the closed, shift-invariant subset of
700: $\gA^\Latt$:
701:
702: \[ \inn{\gW} \ := \ \set{\ba \in \gA^\Latt}
703: { \forall \bn \in \Latt, \ \ \ba_{\sU+\bn} \in \gW} \]
704:
705: One-dimensional subshifts of finite type were first studied by Parry
706: \cite{ParrySoFT} and Smale \cite{Smale}; \ excellent recent
707: introductions are \cite{LindMarcus} and \cite{Kitchens}. Higher
708: dimensional subshifts are closely related to tilings \cite{Mozes1},
709: \cite{Mozes2},\cite{Radin}, and involve many additional subtleties; \ see,
710: for example \cite{MarkleyPaul1},\cite{MarkleyPaul2}. Of particular
711: interest is
712:
713: \subparagraph{The Nontriviality Problem:}
714: {\em For a given set $\gW$, is the corresponding
715: set $\inn{\gW}$ is even nonempty? }
716:
717: \breath
718:
719: The Nontriviality Problem is known to be {\bf formally
720: undecidable}; see \cite{RMR}, \cite{BerDomino}, or \cite{KitS}.
721:
722: \Theorem{}
723: { Let $\sU$ and $\gW$ be as above. $\inn{\gW}$ is nontrivial if and
724: only if there is some locally stationary probability measure $\mu_\sU$
725: on $\gA^\sU$, with $\supp{\mu_\sU} \subset \inn{\gW}$, such that
726: $\mu_\sU$ has a stationary extension.}
727:
728: \begin{thmproof}
729: Suppose that such a $\mu_\sU$ existed, and let $\mu$ be a stationary extension.
730: Clearly, any $\mu-$generic configuration in $\gA^\Latt$ must satisfy the
731: membership criteria of $\inn{\gW}$. Hence, $\inn{\gW}$ must be nonempty.
732:
733: Conversely, if $\inn{\gW}$ was nonempty, then by the Krylov-Bogoliov theorem
734: \cite{Walters}, there are stationary probability measures whose
735: support is contained in $\inn{\gW}$. Let $\mu$ be one of these measures, and
736: let $\mu_\sU := \marg{\sU} \mu$. Then $\supp{\mu} \subset \gW$.
737: \end{thmproof}
738:
739: Let $\extMeas{\gW}$ be the set of extendible measures supported
740: on $\gW$.
741:
742: \Corollary{}
743: { It is formally undecidable whether, for a given subset $\gW \subset \gA^\sU$,
744: the set $\extMeas{\gW}$ is nonempty.}
745: \qed
746:
747: \breath
748:
749: However, it is easily decidable whether $\statMeas{\gW}$ itself is
750: nonempty. The set of all real-valued measures supported on $\gW$ is
751: a finite-dimensional vector space, and the stipulation that an element
752: of this vector space be a locally stationary probability measure takes
753: the form of a finite system of linear equations and inequalities; \
754: solving such a system is a decidable problem.
755:
756: \subsection{Cellular Automata \label{CA} }
757:
758: Let $\sU \subset \Latt$ be finite (metaphorically speaking, $\sU$ is
759: a ``neighbourhood of zero'') and let $\phi:\gA^\sU \into \gA$. For
760: every $\bn \in \Latt$, define $\phi_\bn := \phi \circ \shift{-\bn} :
761: \gA^{\sU+\bn} \into \gA$.
762:
763: The {\bf cellular automata} determined by $\phi$ is then the function
764: $\Phi:\gA^\Latt \into \gA^\Latt$ sending
765: $\mtrx{a_\bn}{\bn \in \Latt}{} \mapsto
766: \mtrx{\phi_\bn \lb(\ba_{\sU+\bn} \rb)}{\bn \in \Latt}{}$.
767: $\phi$ is called the {\bf local transformation rule} for $\Phi$.
768: Cellular automata were first investigated by Von Neumann \cite{vonNeumannCA}
769: and Ulam \cite{UlamCA}, and later extensively studied by Hedlund
770: \cite{HedlundCA}, Wolfram \cite{WolframBook}, and others; \ more
771: recent surveys are \cite{PhysicaCA1},\cite{PhysicaCA2},\cite{DelormeMazoyer},
772: \cite{GolesMartinez}.
773:
774: \breath
775:
776: Any cellular automaton on $\Latt$ can be represented by a subshift of finite
777: type on $\Latt \x \Zahl$. Simply define
778:
779: \[ \tlsU \ := \ \lb( \sU \x \{0\} \rb) \disj
780: \lb\{ ( \underbrace{0,0,\ldots,0}_D , \ 1 ) \rb\} \]
781:
782: and then set $\tlgW \ := \ \set{\ba \in \gA^\tlsU}{
783: a_{( 0,0,\ldots,0, \ 1 )} \ = \
784: \phi \lb( \ba_{\lb( \sU \x \{0\} \rb)} \rb)}$
785:
786: If $\ba \in \gA^{\Latt \x \Zahl}$, then $\ba$ can be seen as a $\Zahl-$indexed
787: sequence of configurations in $\gA^\Latt$. Clearly, $\ba$ is in $\inn{\tlgW}$ if and
788: only if this sequence describes the $\Phi-$orbit of some point in $\gA^\Latt$.
789:
790: Of course, unless $\Phi$ is surjective, not every element of $\gA^\Latt$ will
791: necessarily have a $\Phi-$preimage, and thus, not every element can
792: appear in such a $\Zahl-$indexed sequence of configurations. We can
793: obviate this difficulty by concentrating on the {\bf center} of
794: the dynamical system $(\gA^\Latt, \ \Phi)$.
795:
796: If $X$ is any compact space, and $T:X \into X$ continuous, then
797: the {\bf nonwandering set}, $\Omega(X,T)$ is the set of
798: all points $x \in X$ which are {\bf regionally recurrent}: for
799: any neighbourhood $U$ of $x$, there is some $n \in \Natur$ so that
800: $T^n(U) \intsct U \not= \emptyset. \ \ \ \Omega(X,T)$ is a compact
801: $T-$invariant subset, so we can look at the restricted dynamical system
802: $\lb(\Omega(X,T), \ T_{|\Omega(X,T)} \rb)$ ---however, not all
803: elements of $\Omega(X,T)$ will be regionally recurrent under
804: $T$, when seen in the subspace topology (see \cite{Walters} for
805: an example) ---hence, $\Omega^2(X,T) :=
806: \Omega\lb(\Omega(X,T), \ T_{|\Omega(X,T)} \rb)$ may be a proper
807: subset.
808:
809: By transfinite induction, for any countable ordinal number $\alpha$,
810: define \ $\Omega^{\alp+1}(X,T) :=
811: \Omega\lb(\Omega(X,T), \ T_{|\Omega^\alp(X,T)} \rb)$,
812: and, if $\gamma$ is a limit ordinal, define $\Omega^{\gam}(X,T)$\ $:=
813: \Intsct_{\alp < \gam} \Omega^{\alp}(X,T)$. Since $X$ is compact,
814: this descending sequence of compact subsets must become constant at
815: some countable ordinal $\alp$, so that $\Omega^{\alp+1}(X,T)
816: =\Omega^{\alp}(X,T)$. The {\bf center} of $(X,T)$, defined $\bZ(X,T)
817: := \Omega^{\alp}(X,T)$, is nonempty, compact, and $T-$invariant. If
818: $\mu$ is any $T-$invariant Radon measure on $X$, then $\supp{\mu}
819: \subset \bZ(X,T)$.
820:
821: So, treat $(\gA^\Latt, \ \Phi)$ as a compact topological dynamical
822: system, and let $\bZ(\Phi)$ be its center.
823: The restricted
824: map $\Phi_|:\bZ(\Phi) \into \bZ(\Phi)$ is surjective, so every element
825: in $\bZ(\Phi)$ appears in some $\Zahl-$indexed sequence of $\gA^\Latt-$configurations
826: admissable to $\tlgW$.
827:
828: \subparagraph{The Invariant Measure Problem:}
829: {\em Given a local transformation rule $\phi:\gA^\sU \into \gA$,
830: describe the set of $\Phi-$invariant, stationary measures on $\gA^\Latt$.}
831:
832: \breath
833:
834: Suppose that we represent the cellular automata as a subshift of
835: finite type in the aforementioned way, and suppose that $\mu_\tlsU$ is a
836: locally stationary probability measure on $\gA^\tlsU$. It is easy to
837: verify that a stationary extension of $\mu_\tlsU$ to $\gA^{\Latt \x \Zahl}$
838: is equivilant to a $\Phi-$invariant, stationary measure on $\gA^\Latt$.
839:
840:
841:
842:
843:
844:
845:
846:
847:
848:
849:
850:
851:
852:
853:
854:
855:
856:
857:
858:
859:
860:
861:
862:
863:
864:
865:
866:
867:
868:
869:
870:
871:
872:
873:
874:
875:
876:
877:
878:
879:
880:
881:
882:
883:
884:
885:
886:
887:
888:
889:
890:
891:
892:
893: \section{Caveats and Counterexamples \label{counter}}
894:
895: \subsection{Nonextendability in $\Zahl$; The Entropy Metric}
896:
897: The following counterexample, which first appeared in \cite{DJRW},
898: shows that, even in $\Zahl$, locally stationary measures are not
899: always extendible, when the initial domain is ``disconnected''.
900:
901: Suppose that $\sU := \{0,1,3\}$. If $\mu_\sU$ is a probability
902: measure on $\gA^\sU$, then we can treat the functions $\pr{0},
903: \pr{1}$, and $\pr{3}$ as {\em random variables} ranging over the domain
904: $\gA$. So, let $\mu_\sU$ be any probability measure
905: on $\gA^\sU$ such that:
906:
907: \bitem
908: \item (A) $\pr{0} = \pr{1}, \ \mu_\sU-$almost-surely.
909: \item (B) $\pr{0}$ and $\pr{3}$ are {\em independent} as random variables.
910: (thus $\pr{1}$ and $\pr{3}$ are also independent.)
911: \eitem
912:
913: To ensure $\mu_\sU$ is locally stationary, it suffices to require
914: only that the random variables $\pr{0}$, $\pr{1}$, and $\pr{3}$ are
915: {\em identically distributed}.
916:
917: The measure $\mu_\sU$ cannot be extended even to a locally
918: stationary measure on $\gA^\CC{0..3}$, much less a stationary measure
919: on $\gA^\Zahl$. To see this, suppose that $\mu_\CC{0..3}$ was a
920: locally stationary extension. Then condition (A) defining $\mu_\sU$
921: implies that, as random variables on the probability space
922: $\lb(\gA^\CC{0..3}, \ \mu_\CC{0..3} \rb), \ \ \pr{0} = \pr{1} \ =
923: \pr{2} \ = \ \pr{3}$. But by condition (B),
924: $\pr{0}$ and $\pr{3}$ are independent ---a contradiction.
925:
926: \breath
927:
928: This example can be understood as part of a more general phenomenon.
929: If $\sS$ is any set, and $\mu$ is any probability measure on
930: $\gA^\sS$, then $\mu$ induces an {\bf entropy metric}, $D_\mu$, on the
931: set $\Fin{\sS}$ of all finite subsets of $\sS$. If $\sU, \sV \subset
932: \sS$ are finite, then define
933:
934: \[ H_\mu [\sU | \sV] \ :=
935: \ - \sum_{\bb \in \gA^\sV} \sum_{\ba \in \gA^\sU} \mu[\ba|\bb] \log_2(\mu[\ba|\bb]), \]
936:
937: \[ \btxt{ where \etxt} \ \ \mu[\ba|\bb] \ := \
938: \frac{\mu\set{\bc \in \gA^\sS}{\bc_\sU = \ba \btxt{ \ and \ \etxt}
939: \bc_\sV = \bb }}
940: {\mu\set{\bc \in \gA^\sS}{\bc_\sV = \bb}}. \]
941:
942: \[ \btxt{ Then define: \etxt} \ \
943: D_\mu[\sU, \ \sV] \ := \ H_\mu [\sU | \sV] + H_\mu [\sV | \sU]. \]
944:
945: It is easy to check that $D_\mu$ is a {\em metric} on $\Fin{\sS}$. Furthermore,
946: if $\sS$ is a $\dG-$module, and $\mu$ is a $\dG-$invariant measure, then
947: $D_\mu$ is a $\dG-$action invariant metric. If $\sS$ is a {\em subset}
948: of some $\dG-$module, and $\mu$ is a {\em locally} $\dG-$invariant
949: measure, then $D_\mu$ is a ``locally'' $\dG$-invariant metric, in the obvious
950: sense.
951:
952: \breath
953:
954: Now, suppose $\sM$ is a $\dG-$module, $\sU \subset \sM$, and $\mu_\sU$ is a
955: locally $\dG-$invariant
956: measure on $\gA^\sU$. If $\mu$ is to be an invariant extension of
957: $\mu_\sU$, then it must satisfy the condition:
958:
959: \bquote
960: {\em For every $\sV, \sW \in \Fin{\sU}$, and every $\bg \in \dG, \ \
961: D_\mu[\bg.\sV, \ \bg.\sW] \ = \ D_{\mu_\sU} [\sV, \sW].$}
962: \equote
963:
964: Hence, $D_\mu$ is forced to take certain values on a subset of
965: $\Fin{\sM}$. The question is: can we define $D_\mu$ in the {\em rest}
966: of $\Fin{\sM}$ so that it is a metric? If we cannot, then it is
967: impossible to extend $\mu$.
968:
969: \breath
970:
971: In the aforementioned counterexample, $D_{\mu_\sU} \lb[ \{0\},\ \{1\} \rb]
972: = 0$. Thus, if $\mu$ was an extension of $\mu_\sU$, we would have:
973:
974: \[ D_\mu \lb[ \{0\},\ \{1\} \rb] \ = \
975: D_\mu \lb[ \{1\},\ \{2\} \rb] \ = \
976: D_\mu \lb[ \{2\},\ \{3\} \rb] \ = \ 0 \]
977:
978: and hence, $D_\mu \lb[ \{0\},\ \{3\} \rb] \ = \ 0$. But we know
979: that $D_\mu \lb[ \{0\},\ \{3\} \rb] \ > \ 0$, because $\pr{0}$ and
980: $\pr{3}$ are independent random variables. Hence, no such
981: extension $\mu$ can exist.
982:
983: \subsection{Nonextendability in $\Zahl^D$; The Tiling Condition
984: \label{tiling.condition}}
985:
986: In the previous counterexample, it seems the problem was that
987: the domain $\sU$ was not ``connected''. However, in $\Zahl^2$,
988: extendability can fail even when $\sU$ is a $2 \x 2$ box.
989:
990: Suppose $\sU \subset \Latt$, and $\mu_\sU \in \statMeas{\gA^\sU}$.
991: The {\bf support} of $\mu_\sU$ is some subset $\supp{\mu_\sU} \subset \gA^\sU$;
992: \ let $\inn{\supp{\mu_\sU}}$ be the {\bf subshift of finite type} defined
993: by $\supp{\mu_\sU}$. If
994: $\mu \in \statMeas{\gA^\Latt}$ is a stationary extension of $\mu_\sU$,
995: then any $\mu-$generic configuration $\ba \in \gA^\Latt$ must
996: be an element of $\inn{\supp{\mu_\sU}}$.
997:
998: Thus, we have:
999:
1000: \subparagraph{The Tiling Condition:}
1001: {\em $\mu_\sU$ cannot be extendible unless $\inn{\supp{\mu_\sU}}$ is
1002: nontrivial.}
1003:
1004: \breath
1005:
1006: Intuitively, the configuration $\ba$ determines a
1007: {\bf tiling} of $\Latt$ by elements in $\supp{\mu_\sU}$: \
1008: for any $\bk \in \Latt,
1009: \ \ \ba_{(\bk+\sU)}$ is an element of $\supp{\mu_\sU}$.
1010:
1011: For example, suppose that $D := 2, \ \ U := \CC{0..1} \x \CC{0..1}$,
1012: and $\gA := \{0,\ 1,\ 2\}$. Elements of $\gA^\sU$ are thus $2 \x 2$ words
1013: in $\gA$.
1014:
1015: \begin{figure}[htbp]
1016: \[\begin{array}{ccccccc}
1017: \ddots & \vdots & \vdots & \vdots & \vdots & \vdots & \\
1018: \ldots &0 &0 &2 &1 &1 & \ldots \\
1019: \ldots &1 &0 &1 &0 &1 & \ldots \\
1020: \ldots &2 &1 &2 &1 &2 & \ldots \\
1021: \ldots &0 &1 &1 &1 &1 & \ldots \\
1022: \ldots &0 &1 &0 &0 &1 & \ldots \\
1023: & \vdots & \vdots & \vdots & \vdots & \vdots & \ddots \\
1024: \end{array}\]
1025: \caption{A configuration of letters \label{fig1}}
1026: \end{figure}
1027:
1028:
1029: \begin{figure}[htbp]
1030: \[\begin{array}{cc cc cc ccc}
1031: & \updownarrow & & \updownarrow & & \updownarrow & & \updownarrow\\
1032: \connect & \tilebox{cc}{ 0 & 0 \\ 1 & 0 } &
1033: \connect & \tilebox{cc}{ 0 & 2 \\ 0 & 1 } &
1034: \connect & \tilebox{cc}{ 2 & 1 \\ 1 & 0 } &
1035: \connect & \tilebox{cc}{ 1 & 1 \\ 0 & 1 } &
1036: \connect \\
1037: & \updownarrow & & \updownarrow & & \updownarrow & & \updownarrow\\
1038: \connect & \tilebox{cc}{ 1 & 0 \\ 2 & 1 } &
1039: \connect & \tilebox{cc}{ 0 & 1 \\ 1 & 2 } &
1040: \connect & \tilebox{cc}{ 1 & 0 \\ 2 & 1 } &
1041: \connect & \tilebox{cc}{ 0 & 1 \\ 1 & 2 } &
1042: \connect \\
1043: & \updownarrow & & \updownarrow & & \updownarrow & & \updownarrow\\
1044: \connect & \tilebox{cc}{ 2 & 1 \\ 0 & 1 } &
1045: \connect & \tilebox{cc}{ 1 & 2 \\ 1 & 1 } &
1046: \connect & \tilebox{cc}{ 2 & 1 \\ 1 & 1 } &
1047: \connect & \tilebox{cc}{ 1 & 2 \\ 1 & 1 } &
1048: \connect \\
1049: & \updownarrow & & \updownarrow & & \updownarrow & & \updownarrow\\
1050: \connect & \tilebox{cc}{ 0 & 1 \\ 0 & 1 } &
1051: \connect & \tilebox{cc}{ 1 & 1 \\ 1 & 0 } &
1052: \connect & \tilebox{cc}{ 1 & 1 \\ 0 & 0 } &
1053: \connect & \tilebox{cc}{ 1 & 1 \\ 0 & 1 } &
1054: \connect \\
1055: & \updownarrow & & \updownarrow & & \updownarrow & & \updownarrow\\
1056: \end{array}\]
1057: \caption{The corresponding assignment of matrices. \label{fig2}}
1058: \end{figure}
1059:
1060: Choosing a configuration in $\gA^{\Zahl^2}$ is equivalent to
1061: assigning a $2 \x 2$ matrix to each point in the lattice, so that
1062: adjacent sides agree. For example, the configuration in Figure \ref{fig1}
1063: is equivalent to the assignment of Figure \ref{fig2}
1064:
1065: We will define a locally stationary measure $\mu_\sU$ so that
1066: $\supp{\mu_\sU}$ cannot tile $\Zahl^2$ in this manner. We will
1067: do this by explicitly constructing $\supp{\mu_\sU}$ to tile
1068: a {\em different} space instead ---a kind of ``pseudolattice''
1069: (see Figure \ref{Fig:pseudolattice}).
1070:
1071: Stack two $3 \x 3$ grids on top of one another, and
1072: then ``break'' the connection between the central element of each
1073: level, and its southern, eastern, and western neighbours.
1074: Cross-connect the eastern and western neighbours with each other.
1075: Connect the southern neighbour to the central element of the
1076: level {\em above}, and we connect the central element of this level to the
1077: southern element of the level {\em below}. We also maintain the
1078: connection between the central element and its northern neighbour,
1079:
1080:
1081: \begin{figure}
1082: \includegraphics[scale=2]{pseudolattice.eps}
1083: \caption{A ``pseudolattice''.\label{Fig:pseudolattice}}
1084: \end{figure}
1085:
1086:
1087: \begin{figure}
1088: \[\begin{array}{cc ccc cc}
1089: & \updownarrow & & \updownarrow & & \updownarrow & \\
1090: \connect & \tilebox{cc}{ 9 & 1 \\ 10 & 0 } & \connect &
1091: \tilebox{cc}{ 1 & 4 \\ 0 & 0 } & \connect &
1092: \tilebox{cc}{ 4 & 9 \\ 0 & 10 } & \connect \\
1093: & \updownarrow & & \updownarrow & & \updownarrow & \\
1094: \connect & \tilebox{cc}{ 10 & 0 \\ 11 & 7 } & &
1095: \tilebox{cc}{ 0 & 0 \\ 6 & 6 } & &
1096: \tilebox{cc}{ 0 & 10 \\ 7 & 11 } & \connect \\
1097: & \updownarrow & & & & \updownarrow & \\
1098: \connect & \tilebox{cc}{ 11 & 7 \\ 9 & 1 } & \connect &
1099: \tilebox{cc}{ 7 & 7 \\ 1 & 4 } & \connect &
1100: \tilebox{cc}{ 7 & 11 \\ 4 & 9 } & \connect \\
1101: & \updownarrow & & \updownarrow & & \updownarrow & \\
1102: \end{array}\]
1103:
1104:
1105: \[\begin{array}{ccccccc}
1106: & \updownarrow & & \updownarrow & & \updownarrow & \\
1107: \connect & \tilebox{cc}{ 13 & 12 \\ 14 & 2 } & \connect &
1108: \tilebox{cc}{ 12 & 5 \\ 2 & 2 } & \connect &
1109: \tilebox{cc}{ 5 & 13 \\ 2 & 14 } & \connect \\
1110: & \updownarrow & & \updownarrow & & \updownarrow & \\
1111: \connect & \tilebox{cc}{ 14 & 2 \\ 15 & 6 } & &
1112: \tilebox{cc}{ 2 & 2 \\ 7 & 7 } & &
1113: \tilebox{cc}{ 2 & 14 \\ 6 & 15 } & \connect \\
1114: & \updownarrow & & & & \updownarrow & \\
1115: \connect & \tilebox{cc}{ 15 & 6 \\ 13 & 12 } & \connect &
1116: \tilebox{cc}{ 6 & 6 \\ 12 & 5 } & \connect &
1117: \tilebox{cc}{ 6 & 15 \\ 5 & 13 } & \connect \\
1118: & \updownarrow & & \updownarrow & & \updownarrow & \\
1119: \end{array}\]
1120: \caption{A configuration on the pseudolattice \label{fig4}}
1121: \end{figure}
1122:
1123: Now we'll form a locally stationary measure which tiles {\em this}
1124: space instead. Consider the tiling portrayed in Figure \ref{fig4}.
1125: Count every element of $\gA^{2 \x 2}$ as many times as it
1126: appears in these two pictures. There are $18$ tiles, and each one
1127: appears exactly once. Thus, each of the tiles shown gets a
1128: probability of $\frac{1}{18}$.
1129:
1130: To show that $\mu_\sU$ is locally stationary, it suffices to check
1131: that the {\em left columns} and {\em right columns}
1132: have the same probability distribution, and that the {\em top} and {\em bottom
1133: rows} have the same probability distribution. This is easy to confirm.
1134:
1135:
1136: We claim that one simply {\em cannot} tile $\Zahl^2$ with
1137: this collection of blocks. For example, as soon as one lays down a
1138: tile of the form $\minitilebox{0}{0}{6}{6}$, one is forced to place a
1139: tile $\minitilebox{ 1 }{ 4 }{ 0 }{ 0 }$ immediately above it, since
1140: this is the only tile which will ``match''. Once one has done this,
1141: one {\em must} place the tile $\minitilebox{ 9 }{ 1 }{ 10 }{ 0 }$ to the
1142: left of $\minitilebox{ 1 }{ 4 }{ 0 }{ 0 }$, and the tile
1143: $\minitilebox{ 4 }{ 9 }{ 0 }{ 10 }$ to its right. So far, all the
1144: tiles are compatible. However, now, what tile shall we lay down below
1145: $\minitilebox{ 9 }{ 1 }{ 10 }{ 0 }$? To be compatible with
1146: $\minitilebox{ 9 }{ 1 }{ 10 }{ 0 }$, this tile's top row should read
1147: $\minitilebox{10 }{ 0 }{}{}$. However, to be compatible with the tile
1148: $\minitilebox{ 0 }{ 0 }{ 6 }{ 6 }$ to its immediate right, the tile's
1149: right-hand side should read $\minitilebox{\ \ }{ 0 }{\ \ }{ 6}$. There is no
1150: tile in our collection which meets these two criteria.
1151:
1152:
1153: \breath
1154:
1155: The Tiling Condition is necessary, but not sufficient. To see this,
1156: recall that the set $\extMeas{\sU}$ is {\bf closed} as a subset of
1157: $\statMeas{\sU}$. Thus, its complement is {\bf open}. Hence, every
1158: nonextendible measure is surrounded by a neighbourhood of
1159: nonextendible measures.
1160:
1161: If $\bet$ is the {\em equidistributed} measure (assigning equal
1162: probability to {\em every} element of $\gA^\sU$), and $\eps > 0$ is
1163: small, then consider the measure:
1164:
1165: \[ \mu_\eps \ \ \ := \ \ (1 - \eps)\cdot \mu_\sU \ + \ \eps \cdot \bet \]
1166:
1167: $\mu_\eps$ is a convex combination of $\mu_\sU$ and $\bet$.
1168: Since $\eps > 0$, the support of $\mu_\eps$ is {\em all} of $\gA^\sU$.
1169: Thus, $\mu_\eps$ always satisfies the Tiling Condition.
1170: However, if $\eps$ is ``sufficiently small'', the measure $\mu_\eps $
1171: will be inside the neighbourhood of nonextendible measures
1172: around $\mu_\sU$.
1173:
1174:
1175:
1176:
1177:
1178:
1179:
1180:
1181:
1182:
1183:
1184:
1185:
1186:
1187:
1188:
1189:
1190:
1191:
1192:
1193:
1194:
1195:
1196:
1197:
1198:
1199:
1200:
1201:
1202:
1203:
1204:
1205:
1206:
1207:
1208:
1209:
1210:
1211:
1212:
1213:
1214:
1215:
1216:
1217:
1218:
1219:
1220:
1221:
1222:
1223:
1224:
1225:
1226:
1227:
1228:
1229:
1230:
1231:
1232:
1233:
1234:
1235:
1236:
1237:
1238:
1239:
1240:
1241:
1242:
1243:
1244:
1245:
1246:
1247:
1248:
1249:
1250:
1251:
1252:
1253:
1254:
1255:
1256:
1257:
1258:
1259:
1260:
1261:
1262:
1263: \section{Harmonic Analysis of Extensions \label{sect.harm}}
1264:
1265: \subsection{Configuration Space as a Compact Group}
1266:
1267: Solving the Extension Problem requires a good way of describing
1268: measures, and Harmonic Analysis provides one.
1269: To employ this approach, we must reconceive the configuration
1270: space as a {\bf compact abelian topological group}. Hence,
1271: from now on, we will operate under the assumption that:
1272:
1273: \bquote
1274: {\em The alphabet $\gA$ is a finite abelian group.}
1275: \equote
1276:
1277: The choice of group structure on $\gA$ is unimportant ---if $\gA$ has
1278: $A$ elements, then the simplest choice is to let $\gA := \Zahl/A$.
1279:
1280: If we endow $\gA^\sM$ with the product group structure,
1281: it is a compact abelian topological group. What is its dual group?
1282:
1283: Let $\hgA$ be the dual group of $\gA$. If $\sV \subset \sM$ is finite,
1284: and, for all $v \in \sV, \ \ \chi_v \in \hgA$, then $\chi_v \circ \pr{v}: \gA^\sM \into \Torus{1}$ is the map taking the configuration $\mtrx{a_m}{m \in \sM}{}$
1285: to the value $\chi_v(a_v)$. (Here ``$\Torus{1}$'' is the {\bf unit circle}
1286: group.)
1287:
1288:
1289:
1290: \[ \btxt{ We will use the notation \ ``\etxt} \Tensor_{v \in \sV} \chi_v
1291: \btxt{'' to refer to the map: \ \ \etxt} \]
1292:
1293: \beq
1294: \lb( \prod_{v \in \sV} \chi_v \circ \pr{v} \rb) : \gA^\sM & \into & \Torus{1} \\
1295: \mtrx{a_m}{m \in \sM}{} & \mapsto &
1296: \prod_{v \in \sV} \chi_v(a_v) \\
1297: \eeq
1298:
1299: It is easy to verify the next theorem:
1300:
1301: \Theorem{}
1302: { Let $\sM$ be any set. The dual group of $\gA^\sM$ is the set:
1303:
1304: \[ \set{ \Tensor_{v \in \sV} \chi_v }
1305: { \sV \subset \sM \btxt{ \ is any finite subset, and, for all \ \etxt} v \in \sV, \ \chi_v \in \hgA. } \]
1306: }
1307: \qed
1308:
1309:
1310: \subsection{The Fourier Transform}
1311:
1312: Now, if $\mu$ is a measure on $\gA^\sM$ , and $\chi \in \dual{\sM}$, then
1313: the {\bf Fourier Coefficient} of $\mu$ at $\chi$ is defined:
1314:
1315: \[ \hmu_\chi \ \ = \ \
1316: \inn{\mu,\ \chi} \ \ := \ \ \int_{\gA^\sM} \barchi \ d\mu \]
1317:
1318: The {\bf Fourier Transform} of $\mu$ is the function: \
1319: $ \hmu: \dual{\sM} \into \Cplx$ so that
1320: $\hmu_\chi \ \ = \ \ \inn{\mu,\ \chi}$.
1321:
1322: If $\Meas{\gA^\sM; \ \Cplx}$ is endowed with the {\bf total variation norm},
1323: and $\bC( \dual{\sM}; \ \Cplx)$ is endowed with the {\bf uniform norm},
1324: then the map
1325:
1326: \beq
1327: \Four: \Meas{\gA^\sM; \ \Cplx} & \into & \bC( \dual{\sM}; \ \Cplx) \\
1328: \mu & \mapsto & \hmu \\
1329: \eeq
1330:
1331: is an {\bf injective, bounded linear function} of norm 1
1332: \cite{Katznelson}. Thus, the Fourier transform of $\mu$
1333: totally characterizes it: if $\mu$ and $\nu$ are two measures, and
1334: $\hmu = \hnu$, then $\mu = \nu$.
1335:
1336:
1337:
1338:
1339:
1340:
1341:
1342:
1343:
1344:
1345:
1346:
1347:
1348:
1349:
1350:
1351:
1352:
1353: \subsection{Fourier Theory and (local) Stationarity}
1354:
1355: The shift action of $\dG$ upon $\gA^\sM$ induces a right action of
1356: $\dG$ upon $\dual{\sM}$. If $\bg \in \dG$, and $\chi \in \dual{\sM}$,
1357: then define:
1358:
1359: \beqn
1360: \label{char.action}
1361: \chi.\bg \ = \ \chi \circ \shift{\bg^{-1}}
1362: \eeqn
1363:
1364: \[ \btxt{ Note that, if \ \ \etxt}
1365: \chi \ = \ \prod_{v \in \sV} \lb( \chi_v \circ \pr{v} \rb), \btxt{ \ \ then
1366: \ \ \etxt}
1367: \chi.\bg \ = \ \prod_{v \in \sV} \lb( \chi_v \circ \pr{\bg.v} \rb) \]
1368:
1369: \breath
1370:
1371: If $\sU \subset \sM$ is not closed under the $\dG-$action, then
1372: there is no ``shift action'' on $\gA^\sU$. However, we can still treat
1373: $\dG$ as ``acting'' upon $\dual{\sU}$ in a certain limited capacity, as follows:
1374:
1375: Suppose $\sV \subset \sU$, and $\chi \ = \ \prod_{v \in \sV}\ \lb(\chi_v
1376: \circ \pr{v}\rb)$. Suppose that $\bg \in \dG$ is
1377: such that $\bg.\sV \subset \sU$ also. Then $\chi.\bg \ = \ \prod_{v
1378: \in \sV} \lb( \chi_v \circ \pr{\bg.v} \rb)$ is still an element of
1379: $\dual{\sU}$.
1380:
1381: \Theorem{\label{stat.four}}
1382: {
1383: \blist
1384: \item If $\mu \in \Meas{\gA^\sM}$, then $\mu$ is $\dG-$invariant if and only if,
1385: for every $\chi \in \dual{\sM}$ and every $\bg \in \dG$, \ \
1386: $ \inn{\mu, \ \chi} \ \ = \ \ \inn{\mu, \ \chi.\bg}$
1387:
1388: \item If $\sU \subset \sM$, and $\mu \in \Meas{\gA^\sU}$, then $\mu$ is
1389: {\em locally} $\dG-$invariant if and only if,
1390: for every $\chi \in \dual{\sU}$ and every $\bg \in \dG$ so that $\chi.\bg$ is
1391: also in $\dual{\sU}$, \ \ $\inn{\mu, \ \chi} \ \ = \ \ \inn{\mu, \ \chi.\bg}$.
1392: \elist
1393: }
1394: \begin{thmproof}
1395: We will prove \thmpart{2}, since \thmpart{1} clearly follows.
1396:
1397: \proofof{``$\implies$''}
1398: Let $\chi = \Tensor_{v \in \sV} \chi_v$, for some $\sV \subset \sU$.
1399: Then a simple computation reveals:
1400:
1401: \[ \inn{\mu, \chi} \ \ =
1402: \ \ \sum_{\ba \in \gA^\sV} \mu[\ba] \cdot
1403: \barchi(\ba) \]
1404:
1405: Where, by ``$\mu[\ba]$'', we mean
1406: $\mu\set{\bb \in \gA^\sU}{\bb_{\sV} \ = \ \ba}$. Thus,
1407:
1408: \beq
1409: \inn{\mu, \chi.\bg}
1410: & =_{(1)} &
1411: \sum_{\ba \in \gA^{\bg.\sV}}
1412: \mu[\ba] \cdot \lb(\barchi\circ\shift{\bg^{-1}} (\ba)\rb) \\
1413: & =_{(2)} &
1414: \sum_{\ba \in \gA^{\sV}}
1415: \mu\lb[\shift{\bg}\ba\rb] \cdot \lb(\barchi\circ
1416: \shift{\bg^{-1}}\circ\shift{\bg} (\ba) \rb) \\
1417: & = &
1418: \sum_{\ba \in \gA^{\sV}}
1419: \mu\lb[\shift{\bg}\ba\rb] \cdot \barchi (\ba) \\
1420: & =_{(3)} &
1421: \sum_{\ba \in \gA^{\sV}}
1422: \mu\lb[\ba\rb] \cdot \barchi (\ba) \\
1423: & = &
1424: \inn{\mu,\ \chi}
1425: \eeq
1426:
1427: (1) By definition of $\chi.\bg$ (equation (\ref{char.action})).
1428:
1429: (2) Because $\shift{\bg}:\gA^{\sU} \into \gA^{\bg.\sU}$ is an isomorphism.
1430:
1431: (3) Because $\mu$ is locally $\dG-$invariant.
1432:
1433: \proofof{``$\seilpmi$''}
1434: If $\sV \subset \sU$ is finite, then for any $\ba \in \gA^\sV$,
1435: then it is easy to verify that:
1436:
1437: \[ \mu[\ba] \ \ = \ \
1438: \marg{\sV}\mu [\ba]
1439: \ \ = \ \
1440: \sum_{\chi \in \dual{\sV}} \hmu_\chi \cdot \chi(\ba) \]
1441:
1442: The argument is then very similar to that of ``$\implies$''.
1443: \end{thmproof}
1444:
1445: \subsection{Fourier Properties of Stationary Extensions}
1446:
1447: Suppose that $\sU \subset \sM$, and $\sV \subset \sU$ is a finite subset,
1448: and suppose that $\chi \ \ := \ \ \Tensor_{v \in \sV} \ \chi_v$
1449: is some element of $\dual{\sU}$. Then we can also think of $\chi$ as
1450: an element of $\dual{\sM}$. In other words, $\dual{\sU}$ embeds canonically
1451: in $\dual{\sM}$. We will ``abuse notation'', and identify elements of
1452: $\dual{\sU}$ with their images in $\dual{\sM}$. The following theorem
1453: is a straightforward computation:
1454:
1455: \Theorem{\label{harm.ext}}
1456: {
1457: Let $\mu_\sU \in \Meas{\gA^\sU; \ \Cplx }$, and let $\mu \in
1458: \Meas{\gA^\sM; \ \Cplx }$. Then
1459: $\displaystyle \statement{$\marg{\sU} \mu \ = \ \mu_\sU$} \ \ \iff
1460: \ \ \statement{$\forall \chi \in \dual{\sU}, \ \
1461: \inn{\mu,\ \chi} \ = \ \inn{\mu_\sU, \ \chi} $}$.
1462: }
1463:
1464: \qed
1465:
1466: Thus, we have reduced the Extension Problem to finding a measure
1467: $\mu$ on $\Meas{\gA^\sM}$ whose Fourier coefficients agree with those
1468: of $\mu_\sU$ on $\dual{\sU}$. However, we can't just ``fill in'' the
1469: remaining Fourier coefficients in an arbitrary way. First of all,
1470: we must produce something which is $\dG-$invariant. Second of all,
1471: we want to end up with a probability measure.
1472:
1473: \Theorem{}
1474: {
1475: Let $\mu_\sU \in \invMeas{\dG}{\gA^\sU}$, and let $\mu \in \Meas{\gA^\sM}$.
1476: Then $\mu$ is a stationary extension of $\mu_\sU$ if and only if the
1477: following two conditions are satisfied:
1478:
1479: \bitem
1480: \item For every $\chi \in \dual{\sU}$, and every $\bg \in \dG$, \ \
1481: $ \inn{\mu, \ \chi.\bg} \ \ = \ \ \inn{\mu_\sU, \ \chi }$.
1482:
1483: (This equation must be true even when $\chi.\bg$ is no longer in $\dual{\sU}$).
1484:
1485: \item The Fourier coefficients of $\mu$ form a {\bf positive definite}
1486: sequence.
1487: \eitem
1488: }
1489: \begin{thmproof}
1490: The first condition follows from \thmpart{1} of Theorem \ref{stat.four}.
1491: Notice that, if more than one $\dG-$translate of $\chi$ lies inside
1492: $\dual{\sU}$, then all of them will produce the same equation, by
1493: \thmpart{2} of Theorem \ref{stat.four} (since $\mu_\sU$ is locally
1494: $\dG-$invariant).
1495:
1496: The second condition is just the Bochner-Herglotz theorem to
1497: guarantee that the measure $\mu$ is {\bf nonnegative}
1498: \cite{Katznelson}. This forces $\mu$ to be a probability measure,
1499: because now $\mu[\gA^\sM] \ = \ \inn{\mu, \chr{} } \ = \ \inn{\mu_\sU,
1500: \chr{} } \ = \ \mu_\sU[\gA^\sU] \ = \ 1.$ (since $\mu_\sU$
1501: itself is a probability measure).
1502: \end{thmproof}
1503:
1504:
1505:
1506:
1507:
1508:
1509:
1510: \section{Extension on Finite Modules \label{finite} }
1511:
1512: Suppose that $\sM$ is a {\em finite} $\dG-$module, $\sU \subset \sM$, and
1513: and $\mu_\sU \in \invMeas{\dG}{\gA^\sU}$. We will show that if
1514: $\mu_\sU$ is ``sufficiently close'' to a product measure, then
1515: it is extendible. More generally, we will show:
1516:
1517: \Theorem{\label{thm.fin}}
1518: {
1519: Let $\nu_\sU \in \invMeas{\dG}{\gA^\sU}$ be an extendible measure,
1520: with an invariant extension $\nu$ such that $\supp{\nu} = \gA^\sM$.
1521:
1522: There exists an $\eps > 0$ so that, if $\mu_\sU \in \invMeas{\dG}{\gA^\sU}$
1523: is any measure with $\norm{\mu_\sU - \nu_\sU}{v a r} \ < \ \eps$,
1524: then $\mu_\sU$ is also extendible. This $\eps$ is of the form:
1525:
1526: \[ \eps \ \ = \ \ \frac{1}{ H(\sM) } \cdot \min_{\ba \in \gA^\sM} \nu[\ba]
1527: \]
1528:
1529: ($\min_{\ba \in \gA^\sM} \nu[\ba] >0$ by hypothesis that $\supp{\nu} = \gA^\sM$), where
1530: $H(\sM)$ is a number determined by the $\dG-$module structure
1531: of $\sM$, and which satisfies the following bounds:
1532:
1533: \bitem
1534: \item (A) \ \ \ $H(\sM) \ \leq \ \card{\widehat{\gA^\sM}}$.
1535:
1536: \item (B) \ \ \ $H(\sM) \ \leq \ \card{\dG/\dH} \cdot \card{\widehat{\gA^\sU}}$.
1537: \eitem
1538:
1539: where $\dH$ is the {\bf stabiliser} of $\sM$ in $\dG$:
1540:
1541: \[ \dH \ := \ \set{\bh \in \dG}{\forall m \in \sM, \ \bh \cdot m = m } \]
1542:
1543: }
1544: \begin{thmproof}
1545: Define $\del_\sU := \mu_\sU - \nu_\sU$. Thus, $\del_\sU$ is a real-valued measure.
1546: Since $\mu_\sU$ and $\nu_\sU$ are locally $\dG-$invariant, $\del_\sU$ is also\footnote{
1547: Cylinder subsets of $\gA^\sU$ can have negative $\del_\sU-$measures, but these
1548: measures are still preserved under any shift which leaves the cylinder
1549: set inside $\gA^\sU$.}.
1550:
1551: Next we will define $\del$, a real-valued, $\dG$-invariant measure
1552: upon $\gA^\sM$, in terms of its Fourier coefficients. For every $\chi \in
1553: \dual{\sM}$,
1554:
1555: \bitem
1556: \item If there is some $\kappa \in \dual{\sU}$ and $\bg$ in $\dG$ so that
1557: $\chi \ = \ \kappa.\bg$, then let
1558: $\hdel(\chi) \ \ := \ \ \widehat{\del_\sU}(\kappa)$.
1559:
1560: \item Otherwise, let $\hdel_\chi \ \ := \ \ 0$.
1561: \eitem
1562:
1563: By \thmpart{2} of Theorem \ref{stat.four}, the definition of $\hdel_\chi$
1564: is independent of the choice of $\kappa$ and $\bg$, if more than one
1565: choice is available. By \thmpart{1} of the same theorem,
1566: the measure $\del$ is $\dG-$invariant.
1567:
1568: \Claim{$\del$ is a real-valued measure.}
1569: \begin{claimproof}
1570: Since $\del_\sU$ is a real-valued measure, we know that,
1571: for every $\chi \in \dual{\sU}, \ \ \widehat{\del_\sU} \lb( \barchi \rb)
1572: \ \ = \ \
1573: \overline{\widehat{\del_\sU}( \chi)}$.
1574: It follows that,
1575: for every $\chi \in \dual{\sM}, \ \ \widehat{\del} \lb( \barchi \rb)
1576: \ \ = \ \
1577: \overline{\widehat{\del}( \chi)}$,
1578: and from this, we conclude that $\del$ is also a real-valued measure.
1579: \end{claimproof}
1580:
1581: \claim{There is a number $H(\sM)$, determined by the $\dG-$module structure
1582: of $\sM$, and satisfying inequalities (A) and (B), so that
1583: $\norm{\del}{v a r} \ \ \leq \ \ H(\sM) \cdot \norm{\del_\sU}{v a r}$.
1584: }
1585: \begin{claimproof}
1586: From elementary harmonic analysis \cite{Katznelson} , we know that:
1587:
1588: \bitem
1589: \item $\norm{\widehat{\del_\sU}}{\oo} \ < \ \norm{\del_\sU}{v a r}$.
1590:
1591: \item $\norm{\del}{v a r} \ < \ \norm{\widehat{\del}}{1}$.
1592: \eitem
1593:
1594: Hence, it suffices to show that $ \norm{\widehat{\del}}{1} \ < \
1595: H(\sM) \cdot \norm{\widehat{\del_\sU}}{\oo}$, where $H(\sM)$ is the
1596: aforementioned number. To see inequality (A), notice that
1597:
1598: \[ \norm{\widehat{\del}}{1} \ \leq \ \card{\sM} \cdot \norm{\widehat{\del}}{\oo}
1599: \ = \ \card{\sM} \cdot \norm{\widehat{\del_\sU}}{\oo} \]
1600:
1601: where the second equality follows immediately from the definition of
1602: $\hat{\del}$.
1603:
1604: Now for inequality (B). For any $\chi \in
1605: \widehat{\gA^\sU}$, let $\dG.\chi := \set{
1606: \bg.\chi}{\bg \in \dG}$ be the {\bf orbit} of $\chi$ under the action of
1607: $\dG$. Then:
1608:
1609: \beq
1610: \norm{\hat{\del}}{1}
1611: & = &
1612: \sum_{\chi \in \widehat{\gA^\sM}} \lb| \hat{\del} (\chi) \rb| \\
1613: & = &
1614: \sum_{\chi \in \widehat{\gA^\sU}} \ \ \ \sum_{\xi \in \ \dG.\chi} \lb| \hat{\del} (\xi) \rb| \\
1615: & = &
1616: \sum_{\chi \in \widehat{\gA^\sU}} \ \ \ \sum_{\xi \in \ \dG.\chi} \lb| \widehat{\del_\sU} (\chi) \rb| \\
1617: & = &
1618: \sum_{\chi \in \widehat{\gA^\sU}} \card{\dG.\chi} \cdot \lb| \widehat{\del_\sU} (\chi) \rb| \\
1619: \eeq
1620:
1621: But for any $\chi \in
1622: \widehat{\gA^\sU}, \ \ \card{\dG.\chi} \ < \ \card{\dG / \dH}$.
1623: So this expression is less than
1624:
1625: \beq
1626: \sum_{\chi \in \widehat{\gA^\sU}}
1627: \card{\dG/\dH} \cdot \lb| \widehat{\del_\sU} (\chi) \rb|
1628: & = &
1629: \card{\dG/\dH} \cdot \norm{ \widehat{\del_\sU}}{1} \\
1630: & \leq &
1631: \card{\dG/\dH} \cdot \card{\widehat{\gA^\sU}} \cdot \norm{ \widehat{\del_\sU}}{\oo}
1632: \eeq
1633: \end{claimproof}
1634:
1635: Recall that $\nu$ is some invariant extension of $\nu_\sU$. Define:
1636:
1637: \[ \mu \ \ := \ \ \nu + \del \]
1638:
1639: \claim{$\mu$ is a nonnegative, $\dG-$invariant probability measure.}
1640: \begin{claimproof}
1641: $\mu$ is a sum of two real-valued, $\dG-$invariant
1642: measures, and thus is also a real-valued, $\dG-$invariant measure.
1643:
1644: Also, $\norm{\nu - \mu}{v a r} \ = \ \norm{\del}{v a r} \ <$ \
1645: $H(\sM) \cdot \norm{\del_\sU}{v a r} \ = \ H(\sM) \cdot
1646: \norm{\nu_\sU - \mu_\sU}{v a r}$. Thus,
1647:
1648: \beq
1649: \lefteqn{ \lb(\maketall\norm{\nu_\sU - \mu_\sU}{v a r} <
1650: \eps \ := \ \frac{1}{ H(\sM) } \cdot
1651: \min_{\ba \in \gA^\sM} \ \ \nu[\ba] \rb)} \\
1652: & \hspace{5em}\implies &
1653: \statement{ For every $\ba \ \in \gA^\sM, \ \ \mu[\ba] > 0$.} \\
1654: \eeq
1655:
1656: It remains to show that $\mu[\gA^\sM] = 1$, or, equivalently,
1657: that $\inn{\mu, \chr{}} \ = \ 1$. Since $\inn{\nu, \chr{}} \ = \ 1$,
1658: this is equivalent to showing that $\inn{\del, \chr{}} \ = \ 0$.
1659: But $\inn{\del, \chr{}} \ = \ \inn{\del_\sU, \chr{}}$, and
1660: $\inn{\del_\sU, \chr{}} \ = \ \inn{\nu_\sU, \chr{}} - \inn{\mu_\sU, \chr{}} \ = \ 0$.
1661: \end{claimproof}
1662:
1663: Finally, we want to show that $\mu$ is an extension of $\mu_\sU$.
1664: But
1665: \[ \marg{\sU}{\mu} \ = \ \marg{\sU}{\nu} + \marg{\sU}{\del}
1666: \ = \ \nu_\sU + \del_\sU \ = \ \mu_\sU. \]
1667: \end{thmproof}
1668:
1669: If $\rho$ is a probability measure on $\gA$, let $\rho^\sU$ be the
1670: corresponding product measure on $\gA^\sU$.
1671:
1672: \Corollary{}
1673: {
1674: Let $\sM$ and $H(\sM)$ be as in the previous theorem.
1675: Let $\rho$ be a probability measure on $\gA$ with full support, and
1676: let
1677: \[ \eps \ \ := \ \ \frac{1}{H(\sM)}
1678: \lb( \min_{a \in \gA} \rho(a)\rb)^\card{\sM} \]
1679: Let \ $\sU \subset \sM$. If
1680: $\mu \in \invMeas{\dG}{\gA^\sU}$, and $\norm{\mu - \rho^\sU}{v a r}
1681: \ < \ \eps$, then $\mu$ is extendible.
1682: }
1683: \begin{thmproof}
1684: $\rho^\sU$ extends to the $\dG-$invariant probability measure
1685: $\rho^\sM$ on $\sM$, and
1686: $\displaystyle \min_{\ba \in \gA^\sM} \rho^\sM[\ba] \ \ = \ \
1687: \lb( \min_{a \in \gA} \rho(a)\rb)^\card{\sM}$.
1688: \end{thmproof}
1689:
1690: \section{Envelopes: Reduction to Smaller Modules
1691: \label{envelope}}
1692:
1693:
1694: Suppose that $\sM$ and $\tlsM$ are $\dG-$modules, and that $\phi:\sM \into
1695: \tlsM$ is a $\dG-$module {\bf homomorphism} ----that is, for
1696: all $m \in \sM$ and $\bg \in \dG, \ \phi(\bg.m) \ = \ \bg.\phi(m)$.
1697:
1698: If $\tlba := \mtrx{\tla_\tlm}{\tlm \in \tlsM}{} \ \in \gA^\tlsM$,
1699: then define the element $\ba := \mtrx{a_m}{m \in \sM}{} \ \in \gA^\sM$,
1700: by the formula:
1701:
1702: \beqn
1703: \label{pullback.conf}
1704: \forall m \in \sM, \ \ a_m \ := \ \tla_{\phi(m)}
1705: \eeqn
1706:
1707: This determines a function $\gA^\phi:\gA^\tlsM \into \gA^\sM$,
1708: where $\gA^\phi(\tlba) \ := \ \ba $.
1709:
1710:
1711: If $\tlmu$ is a $\dG-$invariant measure on $\gA^\tlsM$, we define the
1712: {\bf pullback} of $\tlmu$ through $\phi$ to be the measure:
1713: $\pb{\phi} \tlmu \ := \ (\gA^\phi)^* \mu$. It is easily verified
1714: that $\pb{\phi} \tlmu$ is a $\dG-$invariant measure on $\gA^\sM$.
1715:
1716:
1717: Given a $\dG-$module $\sM$, a subset $\sU
1718: \subset \sM$ and a locally $\dG-$invariant
1719: measure $\mu_\sU$ on $\gA^\sU$, we want to find a smaller $\dG-$module
1720: $\tlsM$, a subset $\tlsU \subset \tlsM$, and a locally $\dG-$invariant
1721: measure $\tlmu_\tlsU$ on $\gA^\tlsU$, such that, if we can extend $\tlmu_\tlsU$
1722: to a $\dG-$invariant measure $\tlmu$ on $\gA^\tlsM$, then
1723: $\mu := \pb{\phi} \tlmu$ is an extension of $\mu_\sU$
1724:
1725: \Definition{Envelope}
1726: {
1727: Let $\sM$ be a ~$\dG-$module, and $\sU \subset \sM$.
1728:
1729: An {\bf envelope} for $\sU$ is a~ $\dG-$module $\tlsM$, along with
1730: a~ $\dG-$module homomorphism $\phi:\sM \into \tlsM$, such that
1731:
1732: \bitem
1733: \item (E1) When restricted to $\sU$, the function $\phi$ is injective.
1734:
1735: \item (E2) If $\sV \subset \sU$, then for any $\tlbg \in \dG$ such that
1736: $\tlbg.\phi(\sV) \ \subset \ \phi(\sU)$, we can find some element $\bg \in \dG$
1737: so that:
1738:
1739: \blist
1740: \item $\bg.\sV \subset \sU$,
1741: \item For all $v \in \sV, \ \ \ \phi(\bg.v) \ = \ \tlbg . \phi(v).$
1742: (Thus, $\phi(\bg.\sV) \ = \ \tlbg.\phi(\sV)$.)
1743: \elist
1744:
1745: \eitem
1746: }
1747:
1748: \subparagraph{Example:}{\em Envelopes in a Lattice}
1749:
1750: Suppose $\dG \ = \ \sM \ = \ \Zahl^D$, and let $\sU \subset \Zahl^D$ be
1751: finite, and small enough that it fits into a box of dimensions
1752: $N_1 \x N_2 \x \ldots \x N_D$. We will indicate the action of $\Latt$ on
1753: itself with the ``+'' symbol.
1754:
1755: Consider the $\Zahl^D-$module:
1756:
1757: \[ \tlsM \ \ := \ \ \frac{\Zahl}{2 N_1 \Zahl} \ \x \
1758: \frac{\Zahl}{2 N_D \Zahl} \ \x \ \ldots \x
1759: \frac{\Zahl}{2 N_D \Zahl} \]
1760:
1761: and let $ \phi: \sM \into \tlsM$ be the $\Zahl^D-$module homomorphism:
1762:
1763: \[ \phi(n_1,\ldots,n_D) \ := \
1764: \lb( n_1 + \frac{\Zahl}{2 N_1 \Zahl}, \ \
1765: n_2 + \frac{\Zahl}{2 N_2 \Zahl}, \ \ \ldots , \ \
1766: n_D + \frac{\Zahl}{2 N_D \Zahl} \rb) \]
1767:
1768: Then $(\tlsM, \phi)$ is an envelope for $\sU$.
1769:
1770: \subparagraph{Remark:} In this example, the module
1771:
1772:
1773:
1774:
1775:
1776:
1777:
1778:
1779: \[ \tlsM \ \ := \ \ \frac{\Zahl}{ N_1 \Zahl} \ \x \
1780: \frac{\Zahl}{ N_2 \Zahl} \ \x \ \ldots \x
1781: \frac{\Zahl}{ N_D \Zahl} \]
1782:
1783: with the quotient map $\phi:\sM \into \tlsM$ would {\em
1784: not} necessarily have worked as an envelope for $\sU$. To see this,
1785: suppose that
1786:
1787: \[ \sU \ \ := \ \ \CC{1..N_1} \x \{1\} \x \{1\} \x \ldots \x \{1\} \]
1788:
1789: and let $ \sV \ \ := \ \ \{ \bv_1, \bv_2 \},$ where $\bv_1 \ := \
1790: (1,1,\ldots,1)$, while $\bv_2 \ := \ (2,1,1,\ldots,1)$. Let $\tlbg \ :=
1791: \ (N_1\!-\!1,\ 0,\ 0,\ \ldots,\ 0) \ \in \ \dG$. Then note that
1792:
1793: \[ \tlbg + \phi(\bv_1)
1794: \ \ = \ \
1795: \phi( \tlbg + \bv_1 )
1796: \ \ = \ \
1797: \phi(N_1,1,1,\ldots,1)
1798: \ \ = \ \
1799: \phi(\bv_3), \]
1800:
1801: where $\bv_3 := (N_1,1,1,\ldots,1)$, while
1802:
1803: \[ \tlbg + \phi(\bv_2)
1804: \ \ = \ \
1805: \phi( \tlbg + \bv_2 )
1806: \ \ = \ \
1807: \phi(N_1+1,1,1,\ldots,1)
1808: \ \ = \ \
1809: \phi(1,1,1,\ldots,1)
1810: \ \ = \ \
1811: \phi(\bv_1) \]
1812:
1813: Now, there is no element $\bg \in \dG$ so that
1814: $\bg + \sV \ \ = \ \ \lb\{ \bv_1, \bv_3 \rb\}. $
1815: Thus, although $\tlbg + \phi(\sV) \subset \phi(\sU)$, we {\em cannot}
1816: find some $\bg \in \dG$ so that $\bg + \sV \subset \sU$ and $\phi(\bg + \sV) \ = \ \tlbg + \phi(\sV)$.
1817:
1818: \pause
1819:
1820: \Proposition{\label{thm.env}}
1821: {
1822: Let $\sM$ be a~ $\dG-$module, and $\sU \subset \sM$. Let $\phi:M \into \tlsM$
1823: be an envelope for $\sU$, and $\tlsU := \phi(\sU)$.
1824:
1825: \blist
1826: \item For any probability measure $\mu_\sU$ on $\gA^\sU$, there is a unique
1827: probability measure $\tlmu_\tlsU$ on $\gA^\tlsU$ so that
1828: $ \mu_\sU \ \ = \ \ \pb{\phi} \tlmu_\tlsU.$
1829:
1830: \item If $\mu_\sU$ is locally $\dG-$invariant, then so is $\tlmu_\tlsU$.
1831:
1832: \item If $\tlmu$ is an extension of $\tlmu_\tlsU$ to a~ $\dG-$invariant
1833: probability measure on $\gA^\tlsM$, then $\nu \ := \ \pb{\phi} \tlmu$ is an
1834: extension of $\mu_\sU$ to a~ $\dG-$invariant probability measure on
1835: $\gA^\sM$,
1836: \elist
1837: }
1838:
1839: \begin{thmproof}
1840:
1841: \proofof{\thmpart{1}} By hypothesis, $\phi_| : \sU \into \tlsU$ is injective. Let
1842: $\psi: \tlsU \into \sU$ be the inverse map, and define $\tlmu_\tlsU
1843: := \pb{\psi} \mu_\sU$. Thus,
1844: $\mu_\sU \ = \ \pb{\phi} \tlmu_\tlsU$. Since $\phi_{|\sU}$ is injective,
1845: the measure $\tlmu_\tlsU$ is the unique one satisfying this equation.
1846:
1847: \proofof{\thmpart{2}}
1848: Let $\tlsV \subset \tlsU$, and $\tlbc \in \gA^\tlsV$.
1849: Suppose $\tlbg \in \dG$ is
1850: such that $\tlbg.\tlsV \subset \tlsU$ as well. We want to show:
1851:
1852: \[ \tlmu_\tlsU \lb[ \shift{\tlbg}\,\tlbc \rb] \ \ = \ \ \tlmu_\tlsU \lb[ \tlbc \rb] \]
1853:
1854: Let $\sV \ := \ \psi (\tlsV) \ \subset \ \sU$, and let $\bc \ := \ \gA^\phi(\tlbc)$, where $ \gA^\phi:\gA^\tlsV \into \gA^\sV$ is as defined by
1855: equation (\ref{pullback.conf}) near the
1856: beginning of \S\ref{envelope}. Thus, if $\tlbc \ = \ \seq{\tlc_v}{v
1857: \in \tlsV}$, then
1858: $\bc \ = \ \seq{c_v}{v \in \sV}$, where, for all
1859: $v \in \sV, \ \ c_v := \ \tlc_{\phi(v)}$.
1860:
1861: Let $\bC$ be the cylinder set in $\gA^\sU$ associated to $\bc$
1862: (and likewise, $\tlbC$ for $\tlbc$). Thus, $\tlbC \ = \
1863: \gA^\psi(\bC).$ Since $\tlsM$ is an envelope, there is a $\bg \in \dG$
1864: satisfying condition (E2). By (E2)(1), $\shift{\bg}\,\bC$ is also a cylinder
1865: set in $\gA^\sU$, and since $\mu_\sU$ is locally $\dG-$invariant,
1866: $\mu_\sU \lb[ \shift{\bg}\,\bC \rb] \ = \ \mu_\sU[\bC]$.
1867:
1868: \Claim { $\gA^\psi(\shift{\bg}\,\bC) \ = \ \shift{\tlbg}\,\tlbC$}
1869:
1870: \begin{claimproof}
1871: Let $\tlba := \mtrx{\tla_\tlu}{\tlu \in \tlsU}{} \ \in \gA^\tlsU$,
1872: and suppose that $\tlba = \gA^\psi(\ba)$, where
1873: $\ba := \mtrx{a_u}{u \in \sU}{} \ \in \gA^\sU$. Then
1874: $\statement{$\tlba \in \gA^\psi(\shift{\bg}\,\bC)$}
1875: \iff$
1876: $\statement{$\ba \in \shift{\bg}\,\bC$}
1877: \iff$
1878: $\statement{$\forall v \in \sV, \ a_{\lb(\bg.v\rb)} \ = \ c_v$}
1879: \iff_{\!\!\!(1)}$ \\
1880: $ \statement{$\forall v \in \sV, \ \
1881: \tla_{\lb(\tlbg. \phi(v)\rb)} \ = \ \tlc_{\phi(v)}$ }
1882: \iff$ $\statement{$\tlba \in \shift{\tlbg}\,\tlbC $}.$
1883:
1884: (1) Because, for all $v \in \sV$, \
1885: $ \tla_{\lb(\tlbg. \phi(v)\rb)} \ = \ \tla_{\phi(\bg.v)}
1886: \ = \ a_{\bg.v}$, \ \ and $ c_v = \tlc_{\phi(v)}$.
1887: \end{claimproof}
1888:
1889: Thus, $\tlmu_\tlsU \lb[ \shift{\tlbg}\,\tlbC \rb]
1890: \ = \
1891: \tlmu_\tlsU \lb[ \gA^\psi(\shift{\bg}\,\bC) \rb]
1892: \ \ = \ \
1893: \mu_\sU \lb[ \shift{\bg}\,\bC \rb]
1894: \ = \
1895: \mu_\sU \lb[ \bC \rb]
1896: \ = \
1897: \tlmu_\tlsU \lb[ \tlbC \rb]. $
1898:
1899:
1900: \proofof{\thmpart{3}} This is straightforward.
1901: \end{thmproof}
1902:
1903:
1904:
1905:
1906:
1907: \section{Embedding of Locally Stationary Measures \label{embed}}
1908:
1909: Suppose that $(X,\sX,\nu)$ is a probability space, and $T$ is
1910: a $\nu$-preserving action of $\Latt$ upon $X$. Let
1911: $\sP:X \into \gA$ be a measurable function (ie. a
1912: $\gA$-labelled, {\bf measurable partition} of $X$), and
1913: let $\sP^\Latt:X \into \gA^\Latt$ be the map
1914: $x \mapsto \mtrx{\sP\lb(T^\bn(x) \rb)}{\bn \in \Latt}{}$.
1915: The projection of $\mu$ through $\sP^\Latt$
1916: is then a stationary probability measure on $\gA^\Latt$,
1917: called the {\bf stochastic process induced by $\sP$ and $T$}.
1918: Call this measure $\eta$.
1919:
1920: Suppose that $\sU \subset \Latt$, and $\mu_\sU \in
1921: \statMeas{\gA^\sU}$. The map $\sP$ is an {\bf embedding} of $\mu_\sU$
1922: in the system $(X,\sX,\nu; \ T)$ if $\marg{\sU}{\eta} = \mu_\sU$.
1923: When can $\mu_\sU$ be thus embedded?
1924:
1925: \Theorem{\label{thm.embed}}
1926: {
1927: Suppose that $\sU \subset \Latt$ is finite, and that $\mu_\sU$ lies
1928: in the {\bf interior} of $\extMeas{\gA^\sU}$.
1929: Suppose that $(X,\sX,\nu; \ T)$ is {\bf ergodic}. Then $\mu_\sU$
1930: can be embedded in $(X,\sX,\nu; \ T)$.
1931: }
1932: \begin{thmproof}
1933: We will first show how to construct an ``approximate'' embedding for
1934: $\mu_\sU$. The approximation method involves a certain degree of error,
1935: which can be exactly characterized and then compensated for.
1936:
1937: Suppose $U \in \Natur$, so that $\sU \subset \sB(U)$.
1938: Let $\mu \in \statMeas{\gA^\Latt}$ be an extension of $\mu_\sU$.
1939: Then for any $N > 0, \ \ \mu_{\sB(N)} := \marg{\sB(N)} \mu$ is a
1940: locally stationary probability measure on $\gA^{\sB(N)}$. Also,
1941: if $\sU_0 \subset \sB(N)$ is any translation of $\sU$, then
1942: $\marg{\sU_0}{\mu_{\sB(N)}} \ = \ \mu_{\sU_0}$, where $\mu_{\sU_0}$ is
1943: the obvious ``translation'' of $\mu_\sU$ to the domain $\sU_0$.
1944:
1945: The {\bf Rokhlin Tower Lemma} for $\Latt$-actions says that,
1946: for any $\eps > 0$ and $N \in \Natur$, there is a subset $R \in \sX$
1947: so that the disjoint union:
1948:
1949: \[ \Disj_{\bn \in \sB(N+U)} T^\bn (R) \]
1950:
1951: has measure greater than $1-\eps$.
1952:
1953: Let $x \in X$ be a generic point for $R$, and suppose we look at the
1954: ``name'' of $x$ with respect to the partition $\{R, \ X \setminus R\}$:
1955: for all $\bn \in \Latt$, colour the point $\bn$ ``black'' if $T^\bn x \in
1956: R$, and ``white'' otherwise. Let $\sR \subset \Latt$ be the set of
1957: ``black'' points. The Rokhlin Tower condition is equivalent to saying
1958: that the union:
1959:
1960: \[ \Disj_{\br \in \sR} \lb( \maketall \sB(N+U)+\br \rb) \]
1961:
1962: is disjoint, and has \Cesaro density greater than $1 - \eps$ in $\Latt$.
1963:
1964: To define a measurable function $\sP:X \into \gA$, we will provide a
1965: scheme to determine its value at every point in the $\Latt-$orbit
1966: of $x$, in terms of the $\{R, \ X \setminus R\}-$name of $x$ (this is
1967: sometimes called ``colouring the name of $x$''). The scheme well-defines
1968: the values of $\sP$ on the orbit of every generic point in $X$ ---thus,
1969: it defines $\sP$ almost everywhere on $X$.
1970:
1971: Defining the value of $\sP$ on the $\Latt-$orbit of $x$ is equivalent to
1972: defining a function $\bp: \Latt \into \gA$ ---in other words, a
1973: configuration. Do this as follows: Let $\phi:\sR \into \gA^{\sB(N)}$
1974: be some function so that, for each $\ba \in \gA^{\sB(N)}$, the
1975: \Cesaro density of the subset $\phi^{-1}(\ba)$ inside $\sR$ is equal to
1976: $\mu_{\sB(N)} [\ba]$ (since the set $\sR$ itself has a well-defined
1977: \Cesaro density, such a function can always be constructed).
1978: For each $\bu \in \sR$, let $\bp_{\sB(N)+\bu} = \phi(\bu)$. This
1979: immediately defines $\bp$ on ``most'' of $\Latt$. Now, fix some $\ga \in
1980: \gA$, and label {\em all} remaining points in $\Latt$ with the symbol $\ga$.
1981:
1982: The function $\sP$ induces a stationary probability measure $\eta$
1983: on $\gA^\Latt$. \ \ $\eta_\sU := \marg{\sU} \eta$ is ``close'' to
1984: $\mu_\sU$, but slightly ``enriched'' in words that contain big blocks
1985: of the ``$\ga$'' symbol, while impoverished in words that don't. If
1986: we fix $\eps > 0$ and $N \in \Natur$, then
1987: $ \eta_\sU \ = \ F_{\eps,N} [\mu_\sU]$,
1988: where $F_{\eps,N}:\extMeas{\gA^\sU} \into \extMeas{\gA^\sU}$ is an
1989: {\bf affine function}.
1990:
1991: So, if we want to actually produce the measure $\mu_\sU$ as an {\em
1992: outcome} of this procedure, we must find some $\nu_\sU \in
1993: \extMeas{\gA^\sU}$, so that $\mu_\sU = F_{\eps,N} [\nu_\sU]$. In other
1994: words, in order to use this construction to build an embedding of
1995: $\mu_\sU$ within $X$, we must find some $N$ and $\eps$ so that $\mu_\sU
1996: \in I_{N,\eps} := F_{N,\eps}\lb( \extMeas{\gA^\sU} \rb)$.
1997:
1998: \claim{For any $\del > 0$, there exist $\eps$ and $N$ so that
1999: $\Lbsg[I_{\eps,N}] \ \geq \
2000: (1-\del)\cdot\Lbsg \lb[\extMeas{\gA^\sU}\rb]$,
2001: where $\Lbsg$ is the Lebesgue measure.}
2002: \begin{claimproof}
2003: $F_{\eps,N}$ is affine, and thus, differentiable with a constant
2004: derivative, $D_{\eps,N}$. For any $\delta_1 > 0$, we can find a small
2005: enough $\eps$ and large enough $N$ that, for every $\mu_\sU \in
2006: \extMeas{\gA^\sU}$, \ $\norm{ F_{\eps,N} [\mu_\sU] \ - \ \mu_\sU}{v a
2007: r} < \del_1$. Thus, for any $\del_2 > 0$, we can make $\del_1$ small enough so
2008: that
2009: $\norm{D_{\eps,N} - \Id{}}{\oo} < \del_2$ (where $\norm{\cdot}{\oo}$ is
2010: the operator norm). Thus, for any $\del$, we can in turn make $\del_2$
2011: small enough that the determinant of $D_{\eps,N}$ is within $\del$ of $1$.
2012: Thus, for large enough $N$ and small enough $\eps, \ \ F_{\eps,N}:
2013: \extMeas{\gA^\sU} \into I_{\eps,N}$ is a diffeomorphism, and, if $\Lbsg$
2014: is the Lebesgue measure, then $\Lbsg[I_{\eps,N}] \ \geq \ (1-\del)\cdot
2015: \Lbsg \lb[\extMeas{\gA^\sU}\rb]$.
2016: \end{claimproof}
2017:
2018: \claim{For any $\mu$ in the interior of $\extMeas{\gA^\sU}$ there exist $\eps$
2019: and $N$ so that $\mu \in I_{\eps,N}$.}
2020: \begin{claimproof}
2021: Identify $\Meas{\gA^\sU; \ \Real}$ with $\Real^{\gA^\sU}$, endowed with an
2022: inner product. $I_{\eps,N}$ is convex, so if $\mu \in \extMeas{\gA^\sU}
2023: \setminus I_{\eps,N}$, then there is some unit vector $\bv \in
2024: \Real^{\gA^\sU}$, so that $I_{\eps,N} \subset
2025: \set{ \bw \in \Real^{\gA^\sU}}{ \inn{ \bw - \mu, \ \bv} < 0 }$.
2026: Fix $\mu$, and regard $m_\bv$ as a function of $\bv$.
2027: The set \\ $\set{ \bw \in \Real^{\gA^\sU}}{ \inn{ \bw - \mu, \ \bv} \geq 0 }
2028: \intsct \extMeas{\gA^\sU}$ has nontrivial Lebesgue measure
2029: $m_\bv \cdot \Lbsg \lb[\extMeas{\gA^\sU}\rb]$, for some $m_\bv > 0$.
2030: Since the unit sphere in $\Real^{\gA^\sU}$ is compact, there is some
2031: $M > 0$ so that $m_\bv \geq M$ for all $\bv$ in the sphere.
2032:
2033: Let $\del < M$, and, by Claim $1$, find $\eps$ and $N$ so that
2034: $\Lbsg[I_{\eps,N}] \geq (1-\del)\cdot
2035: \Lbsg \lb[\extMeas{\gA^\sU}\rb]$. Then we have $M \cdot \Lbsg \lb[\extMeas{\gA^\sU}\rb]$ \ $> \ \del \cdot \Lbsg \lb[\extMeas{\gA^\sU}\rb]$ \ $> \
2036: \Lbsg \lb[\extMeas{\gA^\sU} \setminus I_{\eps,N} \rb]$ \\ $\geq \
2037: \Lbsg\lb[ \set{ \bw \in \Real^{\gA^\sU}}{ \inn{ \bw - \mu, \ \bv} > 0 }
2038: \intsct \extMeas{\gA^\sU} \rb]$ \ $> \ M \cdot \Lbsg \lb[\extMeas{\gA^\sU}\rb]$,
2039: a contradiction.
2040: \end{claimproof}
2041:
2042: We conclude that any point $\mu$ in the interior of $\extMeas{\gA^\sU}$
2043: is in $I_{\eps,N}$ for some $\eps$ and $N$, and thus, can be
2044: ``embedded'' in the system $(X,\sX,\mu; \ T)$ via the aforementioned
2045: construction.
2046: \end{thmproof}
2047:
2048:
2049:
2050:
2051:
2052:
2053:
2054:
2055:
2056:
2057:
2058:
2059:
2060:
2061:
2062:
2063:
2064:
2065:
2066:
2067:
2068:
2069:
2070:
2071:
2072:
2073:
2074:
2075:
2076:
2077:
2078:
2079:
2080:
2081:
2082:
2083:
2084:
2085:
2086:
2087:
2088:
2089:
2090:
2091:
2092:
2093:
2094:
2095:
2096:
2097:
2098:
2099:
2100:
2101:
2102:
2103:
2104:
2105:
2106:
2107:
2108:
2109:
2110:
2111:
2112:
2113:
2114:
2115:
2116:
2117: \section{(quasi)Periodic, Ergodic, and Mixing Extensions
2118: \label{sect.per}}
2119:
2120: \subsection{Periodic Probability Measures
2121: \label{per.prob.meas}}
2122:
2123: If $\dP \subset \Natur^D$, then a configuration
2124: $\ba \in \gA^\Latt$ is called {\bf $\dP$-periodic} if, for all $n \in \Latt$
2125: and $p \in \dP$,
2126: \ \ \ $a_{n+p} = a_{n}$. If $\inn{\dP}$ is the sublattice generated
2127: by $\dP$, and $\tlsM := \Latt/\inn{\dP}$, with
2128: $\Latt$ acting upon $\tlsM$ by translation, then $\tlsM$ is
2129: $\Latt-$module. The quotient map $\phi:\Latt \into \tlsM$ is a
2130: homomorphism of $\Latt-$modules. Configuration $\ba$ is
2131: $\dP$-periodic if and only if $\ba \ = \ \gA^\phi \tlba$, for some
2132: word $\tlba \in \gA^\tlsM$ (in the notation of Section
2133: \ref{envelope}).
2134:
2135:
2136: In general, if $\sM$ is a $\dG-$module, $\tlsM$ is another $\dG-$module,
2137: and $\phi:\sM \into \tlsM$ is a $\dG-$module homomorphism, then
2138: we will say that an element $\ba \in \gA^\sM$ is {\bf $\tlsM$-periodic}
2139: if $\ba \ = \ \gA^\phi [\tlba]$, for some $\tlba \in \gA^\tlsM$.
2140:
2141: If $\mu$ is a $\dG-$invariant measure on $\gA^\sM$, then
2142: $\mu$ is {\bf $\tlsM$-periodic} if the elements of the space
2143: $\lb(\gA^\sM, \mu \rb)$ are $\mu-$almost surely $\tlsM-$periodic.
2144: This is the case if and only if there is a $\dG-$invariant
2145: measure $\tlmu$ on $\gA^\tlsM$, such that
2146: $\mu \ = \ \pb{\phi} [\tlmu]$.
2147:
2148: \subsection{Periodic Extensions}
2149:
2150: Suppose that $\sU \subset \sM$, and $\mu_\sU$ is a locally $\dG-$invariant
2151: measure upon $\gA^\sU$. Can we extend $\mu_\sU$ to a periodic measure
2152: on $\gA^\sM$?
2153:
2154: \Theorem{\label{ext.per}}
2155: {
2156: Suppose that $\tlsM$ is a {\em finite} $\dG-$module, a quotient of
2157: $\sM$ via the map $\phi:\sM \into \tlsM$, and an {\bf envelope} for $\sU$.
2158: Let $H(\tlsM)$ be the constant described in Theorem \ref{thm.fin}
2159:
2160: Let $\tlnu$ be a $\dG-$invariant measure on $\gA^\tlsM$, with full
2161: support, and let
2162:
2163: \[ \eps \ := \
2164: \frac{1}{H(\sM)} \cdot \min_{\tlba \in \gA^\tlsU} \tlnu[\tlba] \]
2165:
2166: Let $\nu \ = \ \lb(\gA^\phi\rb)^* \tlnu$, and let $\nu_\sU := \marg{\sU} \nu$.
2167: If $\mu_\sU$ is any locally $\dG-$invariant measure on $\gA^\sU$ so that
2168: $ \norm{\mu_\sU - \nu_\sU}{v a r} \ < \ \eps$,
2169: then $\mu_\sU$ can be extended to a $\dG-$invariant, $\tlsM-$periodic
2170: probability measure on $\gA^\sM$.
2171: }
2172: \begin{thmproof}
2173: Let $\tlsU := \phi(\sU) \subset \tlsM$. By \thmpart{1} of Theorem
2174: \ref{thm.env}, the measure $\tlmu_\tlsU := \pb{(\phi^{-1})} \mu_\sU$
2175: is a locally $\dG-$invariant measure on $\gA^\tlsU$. Further, if
2176: $\tlnu_\tlsU := \marg{\tlsU} \tlnu$, then $\norm{\tlmu_\tlsU -
2177: \tlnu_\tlsU}{v a r} < \eps$. Since $\tlsM$ is finite, we can apply
2178: Theorem \ref{thm.fin}, and extend $\tlmu_\tlsU$ to a $\dG-$invariant measure,
2179: $\tlmu$, on all of $\gA^\tlsM$.
2180:
2181: Now, define $\mu := \ \pb{\phi} [\tlmu]$. Then $\mu$ is
2182: a $\tlsM-$periodic, $\dG-$invariant measure by construction, and also,
2183: $\marg{\sU} \mu \ = \ \mu_\sU$.
2184: \end{thmproof}
2185:
2186: \Corollary{}
2187: {
2188: The set $\extMeas{\gA^\sU}$ has nontrivial {\bf interior} in the
2189: space $\Meas{\gA^\sU; \ \Real}$, and the set of $\tlsM$-periodically
2190: extendible measures has nontrivial interior within $\extMeas{\gA^\sU}$.
2191: }
2192: \begin{thmproof}
2193: Let $\rho$ be any probability measure on $\gA$ with full support,
2194: and let $\mu_\sU := \rho^\sU$ be the product measure on $\gA^\sU$. In
2195: the notation of Theorem \ref{ext.per}, $\rho^\tlsM$ is a
2196: $\dG$-invariant extension of $\rho^\tlsU$, with full support, and
2197: induces a $\tlsM$-periodic extension of $\mu_\sU$ to $\gA^\sM$. By
2198: Theorem \ref{ext.per}, all measures in an open ball around $\mu_\sU$
2199: also have $\tlsM$-periodic extensions.
2200: \end{thmproof}
2201:
2202:
2203: \Corollary{}
2204: {
2205: Suppose $\sU \subset \Latt$ is finite, and fits inside a box of size
2206: $Q_1 \x Q_2 \x \ldots \x Q_D$. Suppose that $\bP := (P_1,\ldots,P_D)$, where
2207: $P_1 \geq 2 Q_1, \ P_2 \geq 2 Q_2, \ \ldots , P_2 \geq 2 Q_2$, and
2208: let $\nu$ be a $\bP-$periodic, stationary probability measure on $\gA^\Latt$.
2209: Let $\nu_\sU := \marg{\sU} \nu$.
2210:
2211: There is an $\eps > 0$ (a function of $\bP$ and $\nu$), so that,
2212: if $\mu_\sU$ is any locally stationary probability measure on $\gA^\sU$
2213: within $\eps$ of $\nu_\sU$ in total variation norm, then $\mu_\sU$
2214: has a $\bP-$periodic extension.
2215: }
2216: \qed
2217:
2218: For any $\bP := (P_1,\ldots,P_D)$, let
2219: $\invMeas{\bP}{\gA^\Latt}$ denote the set of $\bP-$periodic, stationary
2220: processes.
2221:
2222: If $\sU \subset \Latt$, then let $\invMeas{\bP}{\gA^\sU}$ denote the
2223: set of {\bf $\bP$-periodically-extendible measures}: those elements of
2224: $\statMeas{\gA^\sU}$ having an extension that is $\bP-$periodic.
2225: The following facts are not difficult to verify:
2226:
2227: \bitem
2228: \item $\invMeas{\bP}{\gA^\sU}$ is a closed, convex set.
2229:
2230: \item If $\mu \in \invMeas{\bP}{\gA^\sU}$ and
2231: $\nu \in \invMeas{\bQ}{\gA^\sU}$, then any convex combination of $\mu$
2232: and $\nu$ is inside $\invMeas{\bR}{\gA^\sU}$, where, for each $d \in \CC{1...D}, \ \
2233: R_d$ is the {\em lowest common multiple} of $P_d$ and $Q_d$.
2234: \eitem
2235:
2236: Let $\invMeas{per}{\gA^\sU}$ be the set of all locally stationary
2237: measures possessing a periodic extension of any periodicity.
2238: It follows that $\invMeas{per}{\gA^\sU}$ is also a convex set.
2239:
2240: \subsection{Essentially Aperiodic measures
2241: \label{essentially.aperiodic}}
2242:
2243: Not every extendible measure has a periodic extension. This
2244: follows from the existence of {\bf essentially aperiodic tile systems}
2245: ---that is, sets of tiles which can tile the plane, but {\em only}
2246: in an aperiodic fashion. In \cite{RMR}, Raphael Robinson exhibits
2247: a collection of six ``notched'' square tiles, which, along with their
2248: 4 rotations, will tile the plane, but {\em only} in an aperiodic fashion.
2249: We can code these six tiles as six $3 \x 3$ matrices in the alphabet
2250: $\gA := \{ 0, \fa,\fA,\fb,\fB,\fc,\fC \}$
2251:
2252: \[ \begin{array}{ccc}
2253: \frame{$\begin{array}{ccc}
2254: \fA & \fC & \fA \\
2255: \fB & 0 & \fd \\
2256: \fA & \fB & \fA \\
2257: \end{array}$}
2258: &
2259: \frame{$\begin{array}{ccc}
2260: \fa & \fc & \fa\\
2261: \fc & 0 & \fc \\
2262: \fa & \fC & \fa \\
2263: \end{array}$}
2264: &
2265: \frame{$\begin{array}{ccc}
2266: \fa & \fb & \fa\\
2267: \fc & 0 & \fc \\
2268: \fa & \fB & \fa \\
2269: \end{array}$}
2270: \\
2271: \frame{$\begin{array}{ccc}
2272: \fa & \fC & \fa\\
2273: \fB & 0 & \fC \\
2274: \fa & \fB & \fa \\
2275: \end{array}$}
2276: &
2277: \frame{$\begin{array}{ccc}
2278: \fa & \fb & \fa\\
2279: \fc & 0 & \fc \\
2280: \fa & \fb & \fa \\
2281: \end{array}$}
2282: &
2283: \frame{$\begin{array}{ccc}
2284: \fa & \fb & \fa\\
2285: \fb & 0 & \fb \\
2286: \fa & \fB & \fa \\
2287: \end{array}$}
2288: \\
2289: \end{array}\]
2290:
2291: Each tile has a ``0'' symbol in its center, surrounded by four ``corners''
2292: and four ``edges''. The tiles must be put together so that these corners
2293: and edges ``match'' according to the following {\bf mapping rules:}
2294:
2295: \bitem
2296: \item ``$\fb$'' edges must be matched to ``$\fB$'' edges.
2297: \item ``$\fc$'' edges must be matched to ``$\fC$'' edges.
2298: \item Where four tiles meet, exactly three corners must be
2299: of type ``$\fa$'', and one of type ``$\fA$''.
2300: \eitem
2301:
2302: These matching rules can be encoded as a subshift of finite type on
2303: the alphabet $\gA$, defined by some subset $\gR \subset \gA^\sU$, where
2304: $\sU := \CC{1..3}^2$. Any configuration in $\inn{\gR}$
2305: corresponds to some Robinson tiling. Now let $\mu$ be a stationary
2306: probability measure on $\inn{\gR}$, and let $\mu_\sU := \marg{\sU} \mu$.
2307: Then $\mu_\sU$ is a locally stationary measure, and $\supp{\mu_\sU} \ = \ \gR$.
2308:
2309: We claim that $\mu_\sU$ is ``essentially aperiodic''. To see this,
2310: suppose that $\nu$ was any extension of $\mu_\sU$. Then $\supp{\nu} \
2311: \subset \inn{\gR}$, and thus, almost every configuration in the
2312: probability space $(\gA^{\Zahl^2},\nu)$ is aperiodic.
2313:
2314: \subsection{Essentially Periodic Measures}
2315:
2316: At the opposite extreme are {\bf essentially periodic} measures:
2317: locally stationary measures which {\em only} have periodic extensions.
2318:
2319: For example, let $\gA := \{0,1\}$ and $\sU \ := \ \CC{1...9} \x \{0,1\}$, and
2320: let $\gB \subset \gA^\sU$ be the set:
2321:
2322: \[ \lb\{
2323: \begin{array}{c}
2324: \frame{${\scriptsize \begin{array}{ccccccccc}
2325: 0&0&0&0&0&0&0&0&0\\
2326: 1&1&1&1&1&1&1&1&0\\
2327: \end{array}}$}, \\
2328: \vdots\\
2329: \frame{${\scriptsize \begin{array}{ccccccccc}
2330: 0&0&0&0&0&1&0&0&0\\
2331: 0&0&0&0&0&0&1&1&0\\
2332: \end{array}}$},\\
2333: \frame{${\scriptsize \begin{array}{ccccccccc}
2334: 0&0&0&0&0&0&1&1&0\\
2335: 0&0&0&0&0&0&1&0&0\\
2336: \end{array}}$},\\
2337: \frame{${\scriptsize \begin{array}{ccccccccc}
2338: 0&0&0&0&0&0&1&0&0\\
2339: 0&0&0&0&0&0&0&1&0\\
2340: \end{array}}$},\\
2341:
2342: \frame{${\scriptsize \begin{array}{ccccccccc}
2343: 0&0&0&0&0&0&0&1&0\\
2344: 0&0&0&0&0&0&0&0&0\\
2345: \end{array}}$} \\
2346: \end{array}
2347: \rb\} \]
2348:
2349: In other words, all blocks are of the form
2350: \frame{${\scriptsize \begin{array}{cc}w_1 & 0 \\ w_0 & 0\end{array}}$}
2351: where $w_0$ and $w_1$ are successive $8-$bit binary numbers.
2352: Let $\gW \subset \gA^\sU$ be the set containing all elements of $\gB$ and
2353: all their horizontal cyclic permutations. $\gB$ defines a subshift
2354: of finite type, which contains only the orbit of
2355: a single, periodic configuration, having horizontal periodicity
2356: $9$, and vertical periodicity $256$. Call this configuration $\ba$
2357:
2358: If $\mu_\sU$ is the measure on $\gA^\sU$ assigning equal mass
2359: to each of the $2304$ elements of $\gB$, then $\mu_\sU$ has only one
2360: stationary extension: the measure $\mu$ which assigns equal mass
2361: to each of the $2304$ distinct translates of $\ba$. Thus,
2362: $\mu_\sU$ is {\bf essentially periodic}, with period $256 \x 9$.
2363:
2364: Note that the periodicity $256 \x 9$ is much larger than $2 \x 9$, which
2365: was the size of the initial domain $\sU$. Indeed, as this argument
2366: makes clear, the periodicity of essentially periodic measure can be made
2367: to grow exponentially with the size of the initial domain.
2368:
2369: \subsection{Ergodic Extensions \label{sect.erg.extend}}
2370:
2371: A stationary probability measure $\mu$ on $\gA^\Latt$ is called
2372: {\bf ergodic} if any measurable subset $\bU \subset \gA^\Latt$ which
2373: is invariant under all shifts has $\mu-$measure either zero or one.
2374: The set of ergodic measures on $\gA^\Latt$, which we denote by
2375: ``$\ergMeas{\gA^\Latt}$'', is exactly the set of {\em extremal
2376: points} of $\statMeas{\gA^\Latt}$ (see \cite{Furstenberg} or \cite{Walters})
2377: Hence, every stationary measure can be approximated arbitrarily
2378: well as a convex combination of ergodic measures.
2379:
2380: If $\sU \subset \Latt$, and $\mu_\sU \in \statMeas{\gA^\sU}$, then
2381: we say $\mu$ is {\bf ergodically extendible} if it can be extended
2382: to an ergodic measure on $\gA^\Latt$. The set of ergodically
2383: extendible measures will be written as ``$\ergMeas{\gA^\sU}$''.
2384: Since the map $\pr{\sU}^*:
2385: \Meas{\gA^\Latt} \into \Meas{\gA^\sU}$ is linear, any extremal
2386: point of $\extMeas{\gA^\sU}$ has a $\pr{\sU}^*-$preimage which is
2387: extremal in $\statMeas{\gA^\Latt}$. As a consequence, every
2388: extremal point of $\extMeas{\gA^\sU}$ is in $\ergMeas{\gA^\sU}$.
2389: Hence, every extendible measure on $\gA^\sU$ can be approximated arbitrarily
2390: well as a convex combination of ergodically extendible measures.
2391:
2392: \breath
2393:
2394: We will see in Section \ref{mix.extend} that, in fact, ``almost all''
2395: extendible measures are ergodically extendible. However, not every
2396: extendible measure is. To see this, suppose that $\sU \subset \Latt$
2397: is some finite domain, let $\gA$ and $\gB$ be two {\em disjoint}
2398: alphabets, and suppose that $\mu_\sU \in \statMeas{\gA^\sU}$ and
2399: $\nu_\sU \in \statMeas{\gB^\sU}$ are two extendible probability
2400: measures. Let $\eta_\sU := \frac{1}{2} \mu_\sU + \frac{1}{2}
2401: \nu_\sU$. Then $\eta_\sU$ is also extendible, and any extension of
2402: $\eta_\sU$ is of the form $\eta := \frac{1}{2} \mu + \frac{1}{2} \nu$,
2403: where $\mu$ and $\nu$ extend $\mu_\sU$ and $\nu_\sU$, respectively.
2404: $\eta$ can never be ergodic: $\ \gA^\Latt$ and $\gB^\Latt$ are
2405: disjoint, shift-invariant subsets of $\lb(\gA \disj \gB\rb)^\Latt$,
2406: each having $\eta-$measure $\frac{1}{2}$.
2407:
2408:
2409: \Proposition{}
2410: {
2411: Let $\sU \subset \Latt$ be finite.
2412: \blist
2413: \item Every ergodically extendible measure on $\gA^\sU$ is a limit point of
2414: periodically extendible measures.
2415: \item $\invMeas{per}{\gA^\sU}$ is a dense, convex subset of
2416: $\extMeas{\gA^\sU}$.
2417: \item
2418:
2419: $\invMeas{per}{\gA^\sU}$ contains the entire {\bf interior} of
2420: $\extMeas{\gA^\sU}$.
2421: \elist
2422: }
2423: \begin{thmproof} \thmpart{2} follows immediately from \thmpart{1}, and
2424: the fact that $\invMeas{per}{\gA^\sU}$ is convex, and the fact that
2425: $\extMeas{\gA^\sU}$ is the convex closure of $\ergMeas{\gA^\sU}$.
2426:
2427: \proofof{\thmpart{3}} This follows from \thmpart{2}, and the fact that, if $C$
2428: a dense, convex subset of a $D$-dimensional convex set $K$, then $C$
2429: contains $\interior{K}$. To see this, let $x \in \interior{K}$, and
2430: let $B$ be an open ball around $x$ inside of $\interior{K}$. Let $S$
2431: be the boundary of $B$, and let $s_1,\ldots,s_D$ be $D$ equidistant
2432: points in $S$, so that their convex closure, $co\{s_1,\ldots,s_D\}$ is
2433: a regular $D$-simplex containing the centre-point $x$.
2434:
2435: Since $C$ is dense in $K$, \ $C \intsct B$ is dense in $B$.
2436: Thus, find elements $c_1,\ldots,c_D \in C$ so that, for
2437: all $d \in \CC{1..D}$, \ $c_d$ is ``very close'' to $s_d$. Then
2438: $co\{c_1,\ldots,c_D\} \subset C$ is a $D$-simplex ``very close'' to
2439: $co\{s_1,\ldots,s_D\}$, and therefor contains $x$.
2440:
2441: \proofof{\thmpart{1}} Let $\mu_\sU \in
2442: \ergMeas{\gA^\sU}$, and let $\mu$ be an ergodic extension of
2443: $\mu_\sU$. Let $\ba \in \gA^\Latt$ be a {\em generic configuration}
2444: for $\mu$: in other words, for any finite subset $\sV \subset \Latt$
2445: and configuration $\bb \in \gA^\sV$,
2446:
2447: \[ \mu[\bb] \ \ = \ \ \lim_{N \goto \oo} \Freq{\bb}{\ba}{\sB(N)} \]
2448:
2449: where $\sB(N) := \CO{0...N}^D$ is the $D-$dimensional cube of
2450: side length $N$, and
2451:
2452: \beq
2453: \Freq{\bb}{\ba}{\sB(N)} & := &
2454: \frac{\btxt{ \# of times ``$\bb$'' appears inside $\ba_{\sB(N)}$ \etxt}}
2455: {N^D} \\
2456: & = &
2457: \frac{1}{N^D} \sum_{\bn \in \sB(N)} \chr{}\{\ba_{\sV+\bn} = \bb\}
2458: \eeq
2459:
2460: Such generic configurations exist, by the Birkhoff Ergodic Theorem.
2461:
2462: In particular, for any $\eps > 0$, we can find a large enough $N$ so that,
2463: for {\em all} $\bb \in \gA^\sU$,
2464:
2465: \[ \lb| \maketall
2466: \mu[\bb] - \Freq{\bb}{\ba}{\sB(N)} \rb| \ < \ \frac{\eps}{2} \]
2467:
2468: Suppose that all of $\sU$ fits inside a cube of side length $U$.
2469: Assume that $N$ is so large that the $U-$thick boundary of $\sB(N)$ is
2470: ``relatively small'':
2471:
2472: \[ \frac{\card{\sB(N)} - \card{\sB(N-U)}}{\card{\sB(N)}} \ < \
2473: \frac{\eps}{2} \]
2474:
2475: Now, identify $\sB(N)$ with $\sN := (\Zahl/N) \dirsum \ldots \dirsum (\Zahl/N)$,
2476: and treat $\ba_{\sB(N)}$ as an element of $\gA^\sN$. Then this configuration,
2477: along with its $N^D$ periodic translations on $\gA^\sN$, defines a
2478: stationary measure on $\gA^\sN$, which, in turn, defines an $\sN-$periodic,
2479: stationary measure on $\gA^\Latt$. Call this measure $\nu$, and then
2480: let $\nu_\sU := \marg{\sU} \nu$. It is straightforward to verify that
2481:
2482: \[ \norm{ \nu_\sU \ - \ \mu_\sU } \ < \ \eps \]
2483:
2484: and of course, by construction, $\nu_\sU \in \invMeas{per}{\gA^\sU}$.
2485: \end{thmproof}
2486:
2487: \subsection{Mixing, Weak Mixing, and Quasiperiodicity
2488: \label{mix.extend}}
2489:
2490: A stationary probability measure $\mu$ on $\gA^\Latt$ is called
2491: {\bf weakly mixing} if the stochastic process $(\gA^\Latt \x \gA^\Latt,\
2492: \mu \tensor \mu)$ is ergodic. $\mu$ is called {\bf mixing} if,
2493: for any measurable $A,B \subset \gA^\Latt$ of nonzero measure,
2494: any any sequence $\seq{\bn_k}{k \in \Natur} \subset \Latt$ tending
2495: to infinity, $\lim_{k \goto \oo} \mu\lb[ A \intsct \shift{\bn_k} B\rb]
2496: = \mu[A] \cdot \mu[B].$ A function $\phi \in \bL^2(\gA^\Latt, \mu)$
2497: is an {\bf eigenfunction} of the system $(\gA^\Latt, \mu)$
2498: if there is a group homomorphism $\chi:\Latt \into \Torus{1}$
2499: such that, for all $\bn \in \Latt$, \ \ $\shift{\bn} (\phi)
2500: = \chi(\bn) \cdot \phi$. The system is called {\bf quasiperiodic}
2501: if $\bL^2(\gA^\Latt, \mu)$ has an orthonormal basis of eigenfunctions.
2502:
2503: All of these concepts can be defined for any measure-preserving
2504: $\Latt-$action on a probability space $(X,\sX,\mu)$. Mixing
2505: implies weak mixing implies ergodicity, but weak mixing and
2506: quasiperiodicity are mutually exclusive.
2507: Furthermore, all of these properties are {\bf inheritable}
2508: through morphisms. If $(X,\sX,\mu; \ T)$ and $(\hX,\hsX,\hmu; \ \hT)$ are
2509: two measure-preserving $\Latt-$actions, then a {\bf morphism} between
2510: the systems is a measure-preserving surjection $\Psi:X \into \hX$ so
2511: that, for all $\bn \in \Latt, \ \Psi \circ T^{\bn} = \hT^{\bn} \circ
2512: \Psi$. If $\Psi$ is such a morphism, and $(X,\sX,\mu; \ T)$ is
2513: ergodic (respectively: weakly mixing, mixing, or quasiperiodic),
2514: then so is $(\hX,\hsX,\hmu; \ \hT)$.
2515:
2516: In particular, let $F: X \into \gA$ be a measurable function, so
2517: that $F$ and $T$ together {\bf induce} a stationary stochastic process
2518: on $\gA^\Latt$, having measure $\hmu$ (see Section \ref{embed}). If
2519: $\hX := \supp{\hmu} \subset \gA^\Latt$ and $\hT:=
2520: \shift{}$, then the map $F^\Latt:X \into \gA^\Latt$ is a morphism.
2521: Thus, if $(X,\sX,\mu; \ T)$ possesses any of the aforementioned
2522: inheritable properties, so does the process $(\gA^\Latt, \hmu)$.
2523:
2524: \Theorem{}
2525: {
2526: Suppose that $\sU \subset \Latt$ is finite, and that $\mu_\sU$ is
2527: in the {\bf interior} of $\extMeas{\gA^\sU}$. Then
2528: $\mu_\sU$ can be extended to a stationary process $\mu$ which
2529: is any of: ergodic, mixing, weakly mixing, or quasiperiodic.
2530: }
2531: \begin{thmproof}
2532: The argument is the same in all four cases. First, find a
2533: system $(X,\sX,\nu; \ T)$ which is ergodic, and which also has the
2534: property in question (for the first three, this is trivial; for
2535: the fourth, it is sufficient to know that ergodic, quasiperiodic
2536: systems exist). Next, use Theorem \ref{thm.embed} to
2537: {\bf embed} $\mu_\sU$ within the desired process. Let $\mu \in
2538: \statMeas{\gA^\Latt}$ be the stochastic process generated by
2539: this embedding. Then $\mu$ itself has the desired property.
2540: \end{thmproof}
2541:
2542: The same argument works for any other ``inheritable'' property of
2543: dynamical systems. The interpretion: knowledge of the local marginal
2544: $\mu_\sU$ tells you basically nothing about the asymptotic dynamical
2545: properties of the process $\mu$.
2546:
2547:
2548:
2549:
2550:
2551:
2552:
2553:
2554:
2555:
2556: \section{Decidability Questions \label{decide}}
2557:
2558: In section \ref{finite.type}, we showed:
2559:
2560: \bquote
2561: {\em It is formally undecidable whether, for a given subset $\gW
2562: \subset \gA^\sU$, the set $\extMeas{\gW}$ is nonempty.}
2563: \equote
2564:
2565: This raises the question of whether the Extension Problem itself is
2566: formally decidable.
2567:
2568: Let $\RReal$ be the set of all {\bf recursively computable} ({\bf r.c}) real
2569: numbers: that is, real numbers whose decimal expansion can be
2570: generated by some Turing Machine \cite{HopcroftUllman}. $\RReal$ is a
2571: countable field, containing all rational and real-algebraic numbers.
2572: Let $\RMeas{\gA^\sM; \ \Real}$ be the set of {\bf r.c.}, real-valued
2573: measures: those such that, if $\sV \subset \sM$ is finite, and $\ba
2574: \in
2575: \gA^\sV$, then the measure of $\ba$ is an element of $\RReal$.
2576: (Of course, some ``exotic'' measurable subsets of $\gA^\Latt$ may
2577: have non {\bf r.c.} measures).
2578: $\RMeas{\gA^\sM; \ \Real}$ is a vector space over the field $\RReal$.
2579:
2580: Let $\RinvMeas{\dG}{\gA^\sM}$ be the set of $\dG-$invariant probability
2581: measures, etc. Clearly, when we ask about the ``formal decidability'' of
2582: the Extension Problem, what we are {\em really} referring to is the
2583: Extension Problem for {\bf r.c.} measures:
2584:
2585: \bquote
2586: {\em If $\sU \subset \sM$, and $\mu_\sU \in \RinvMeas{\dG}{\gA^\sU}$,
2587: is $\mu$ extendible to a $\dG-$invariant measure on $\gA^\sU$?}
2588: \equote
2589:
2590: Note that we do not require the extension itself to be {\bf r.c}.
2591: If a recursive decision procedure explicitly {\em constructs} an extension,
2592: then this extension will be {\bf r.c.} by nature. However, it is
2593: conceivable that some recursive decision procedure might exist which
2594: demonstrates the existence of an extension by ``nonconstructive''
2595: means. It is conceivable that, although we can recursively
2596: decide that $\mu_\sU$ is extendible, no {\bf r.c} extension exists.
2597:
2598: A subset $\bS \subset \RinvMeas{\dG}{\gA^\sU}$ is called
2599: {\bf recursively decidable} ({\bf r.d}) if there is a Turing machine $\dM$,
2600: so that, when given any $\mu \in \RinvMeas{\dG}{\gA^\sU}$ as input,
2601: $\dM$ halts after some {\em finite} number of steps, and outputs
2602: either ``yes'' or ``no'', depending upon whether or not $\mu$ is
2603: an element of $\bS$.
2604:
2605: A subset $\bS \subset \RinvMeas{\dG}{\gA^\sU}$ is called {\bf
2606: recursively enumerable} ({\bf r.e}) if there is a Turing machine
2607: $\dM$, so that, when given any integer $n \in \Natur$ as input, $\dM$
2608: halts after a finite number of steps, and produces as output some
2609: measure $F_\dM[n] \in
2610: \bS$, and so that the function $F_\dM:\Natur \into \bS$ instantiated by
2611: $\dM$ is {\em surjective.} In other words, $\dM$ provides a mechanism to
2612: systematically ``list'' all elements of $\bS$.
2613:
2614: Equivalently, $\bS \subset \RinvMeas{\dG}{\gA^\sU}$ is
2615: {\bf recursively enumerable} if there is a Turing machine $\dM$,
2616: so that, when given any $\mu \in \RinvMeas{\dG}{\gA^\sU}$ as input,
2617: $\dM$ halts after some finite number of steps {\em unless}
2618: $\mu$ is {\em not} in $\bS$, in which case $\dM$ never halts.
2619:
2620: The following facts are easy to verify: Any {\bf r.d}
2621: set is {\bf r.e.}, but the converse is not true. However,
2622: if both $\bS$ and its complement are {\bf r.e.}, then $\bS$ is
2623: {\bf r.d}. Finally, although a countable union of
2624: {\bf r.d} sets is not necessarily itself {\bf r.d}, it is still
2625: {\bf r.e.} \cite{HopcroftUllman}.
2626:
2627: \Theorem{}
2628: {
2629: Let $\sU \subset \Latt$ be a finite subset. Then
2630:
2631: \blist
2632: \item For any $\bP \in \Natur^D, \ \ \RinvMeas{\bP}{\gA^\sU}$ is
2633: {\bf r.d}.
2634:
2635: \item $\RinvMeas{per}{\gA^\sU}$ is {\bf r.e.}.
2636:
2637: \item $\RstatMeas{\gA^\sU} \setminus \RextMeas{\gA^\sU}$
2638: is {\bf r.e.}.
2639: \elist
2640: }
2641: \begin{thmproof}
2642:
2643: \proofof{\thmpart{1}}
2644: If $\mu_\sU \in \RstatMeas{\gA^\sU}$, we want to know whether
2645: the set
2646: $S := \set{ \mu \in \invMeas{\bP}{\gA^\Latt}}{\marg{\sU} \mu = \mu_\sU}$
2647: is nonempty.
2648:
2649: Suppose $\bP := (P_1,\ldots,P_D)$. Let $\tlsM := (\Zahl/P_1) \dirsum \ldots
2650: \dirsum (\Zahl/P_D)$, and suppose that $\sU$ maps bijectively into the
2651: subset $\tlsU \subset \sM$ via the quotient map from $\Latt \into \tlsM$.
2652: Let $\tlmu_\tlsU \in \statMeas{\gA^\tlsU}$ be the projected image of $\mu_\sU$.
2653:
2654: The vector space of $\bP-$periodic, signed measures on $\gA^\Latt$ is
2655: linearly isomorphic to the finite dimensional vector space
2656: $\Meas{\gA^\tlsM; \ \Real}$. The image of $S$ under this isomorphism
2657: is the affine set
2658: \[\tlS := \set{\mu \in \Meas{\gA^\tlsM; \ \Real}}{\mu \
2659: \btxt{a stationary probability measure, and \etxt} \marg{\tlsU} \mu =
2660: \tlmu_\tlsU}.\]
2661: $\tlS$ is the solution set of a finite system of linear
2662: equations and linear inequalities in $\mu$:
2663:
2664: \bitem
2665: \item $\mu\lb[\gA^\tlsM\rb] = 1$.
2666: \item For all $n \in \Latt, \ \ \shift{n}_* \mu = \mu$.
2667: \item $\marg{\tlsU} \mu = \tlmu_\tlsU$.
2668: \item For all $\ba \in \gA^\sM, \ \ \mu[\ba] \geq 0$.
2669: \eitem
2670:
2671: Thus, it is {\bf r.d} whether $\tlS$ is nonempty, and thus,
2672: whether $\mu_\sU$ has a $\bP-$periodic extension.
2673:
2674: \proofof{\thmpart{2}}
2675: $\RinvMeas{per}{\gA^\sU}$ is a countable union of recursively
2676: decidable sets, and thus, {\bf r.e.}.
2677:
2678: \proofof{\thmpart{3}}
2679: Suppose $\mu \in \Meas{\gA^\Latt; \ \Cplx}$ has Fourier transform
2680: $\mtrx{\hmu_\chi}{\chi \in \dual{\Latt}}{}$, let $\sV \subset \Latt$
2681: be a finite subset, and let $\ba \in \gA^\sV$. It is easy
2682: to verify:
2683:
2684: \[ \mu[\ba] \ = \ \sum_{\chi \in \dual{\sV}}
2685: \hmu_\chi \cdot \overline{\chi(\ba)} \]
2686:
2687: Thus, if $\mu_\sU \in \statMeas{\gA^\sU}$, then by Theorems
2688: \ref{stat.four} and \ref{harm.ext}, $\mu$ is an extension of $\mu_\sU$ if and only if:
2689:
2690: \bitem
2691: \item For all $\chi \in \dual{\sU}, \ \ \hmu_\chi \ = \
2692: \inn{\mu_\sU, \chi}$
2693:
2694: \item For all $\bn \in \Latt$, and all $\chi \in \dual{\Latt}$, if
2695: $\xi := \chi \circ \shift{\bn}$ then $\hmu_\chi \ = \ \hmu_\xi$.
2696:
2697: \item For all finite $\sV \subset \Latt$ and $\ba \in \gA^\sV, \
2698: \ \ \sum_{\chi \in \dual{\sV}} \hmu_\chi \cdot \overline{\chi(\ba)} \ >\ 0$.
2699: \eitem
2700:
2701: Thus, an extension for $\mu_\sU$ is equivalent to a set of Fourier
2702: coefficients satisfying a countable collection of linear equations and
2703: inequalities.
2704:
2705: For all $N \in \Natur$, let $\sB(N) := \CC{0...N}^D$, and let $\Xi_N :=
2706: \dual{\sB(N)}$. If $N$ is large enough that
2707: $\sU \subset \sB(N)$, then we can start by trying to define all the
2708: Fourier Coefficients in the set $\set{\mu_\chi}{\chi \in \Xi_{N}}$. The
2709: three sets of linear constraints listed above now become a finite
2710: system of linear equations and inequalities ---if the solution set is
2711: nonempty, call it $S_N$.
2712:
2713: \Claim{Suppose that, for all $N \in \Natur$, the set $S_N$ is nonempty.
2714: Then $\mu_\sU$ is extendible.}
2715: \begin{claimproof}
2716: $S_N$ is a compact subset of the
2717: finite dimensional vector space $\Cplx^{\Xi_N}$. Furthermore, if
2718: $S_{N+1}$ is also nonempty, then any vector in $S_{N+1}$, when projected to
2719: $\Cplx^{\Xi_N}$, determines an element in $S_N$. Call this projection map
2720: $\pr{N}$.
2721:
2722: Fix $N$, and, for all $M > N$, let $\tlS^M_N := \pr{N} \circ \pr{N+1}
2723: \circ \ldots \circ \pr{M-1} (S_M)$, a nonempty compact subset of
2724: $S_N$. Also, $\tlS^{M+1}_N \supset \tlS^{M+2}_N
2725: \supset \tlS^{M+3}_N \supset \ldots$. \ Thus, $\tlS_N := \intsct_{M > N}
2726: \tlS^M_N$ is a nonempty compact subset. Further,
2727: $\pr{N} \lb(\tlS_{N+1}\rb) = \tlS_N$. Thus, any element of $\tlS_N$ can be
2728: ``extended'' to an element of $\tlS_{N+1}$, which can then be
2729: ``extended'' to $\tlS_{N+1}$, etc.
2730:
2731: Pick any element $\hmu_N \in \tlS_N$, and inductively extend it in
2732: this fashion, producing $\hmu_M \in \tlS_M$, for every $M > N$.
2733: Once this is done, the collection of vectors $\seq{ \hmu_M }{M > N}$
2734: defines a single element $\hmu \in \Cplx^{\dual{\Latt}}$.
2735: \ $\hmu$ is the Fourier transform of some measure $\mu$, and by construction,
2736: $\mu$ is a stationary probability measure, and an extension of $\mu_\sU$.
2737: \end{claimproof}
2738:
2739: Hence, if $\mu_\sU$ is {\em not} extendible, then, by contradiction,
2740: there must be some $N \in \Natur$ so that $S_N$ is empty.
2741: Since $S_N$ is the solution set of a finite system of linear equations
2742: and inequalities, it is {\bf r.d} whether $S_N$ is
2743: empty.
2744:
2745: Hence, by successively checking the nonemptiness of $S_N$ for
2746: each $N \in \Natur$, we have a recursive procedure which will
2747: halt {\em if} $\mu_\sU$ is {\em not} extendible, and tell us so.
2748: (If $\mu_\sU$ {\em is} extendible, however, the procedure will never
2749: halt). Thus we can recursively enumerate the elements of
2750: $\RstatMeas{\gA^\sU} \setminus \RextMeas{\gA^\sU}$.
2751: \end{thmproof}
2752:
2753:
2754: \Theorem{}
2755: {
2756: Let $\sU \subset \Latt$ be finite. The set
2757: $\RextMeas{\gA^\sU}$ is {\em not} {\bf r.e.}.
2758: }
2759: \begin{thmproof}
2760: Recall that, if $\gT \subset \gA^\sU$, then $\inn{\gT}$ is the
2761: associated {\bf subshift of finite type} (see Section \ref{finite.type}).
2762: Let $\bN := \set{\gT}{\inn{\gT} \ \btxt{is not trivial\etxt}}$,
2763: and let $\bT := \set{\gT}{\inn{\gT} \ \btxt{is trivial\etxt}}$.
2764: Recall that $\bN$ is not {\bf r.d} (see \cite{RMR}, \cite{BerDomino},
2765: or \cite{KitS}).
2766:
2767: \claim{Suppose $\gT \subset \gA^\sU$. If $\gT \in \bT$, then there
2768: is some $N \in \Natur$ so that no configuration in $\gA^{\sB(N)}$ is
2769: $\gT-$admissable}
2770: \begin{claimproof}
2771: Suppose that, for every $N \in \Natur$, there was a configuration
2772: $\ba^{[N]} \in \gA^{\sB(N)}$ that was $\gT-$admissable ---that is:
2773: for all $n \in \sB(N)$, if $n+\sU \subset \sB(N)$, then $\ba^{[N]}_{n+\sU} \in
2774: \gT$. Extend $\ba^{[N]}$ to an element of $\gA^\Latt$ by filling all the
2775: remaining entries in some arbitrary fashion ---call the extended
2776: configuration $\bb^{[N]}$
2777:
2778: Since $\gA^\Latt$ is compact, the sequence $\mtrx{\bb^{[N]}}{N \in \Natur}{}$
2779: has a convergent subsequence ---call it $\mtrx{\bb^{[N_k]}}{k \in \Natur}{}$
2780: ----which converges to some limit $\bb \in \gA^\Latt$.
2781:
2782: For any $M \in \Natur$, there is some $K \in \Natur$ so that, for all $k > K,
2783: \ \ \bb^{[N_k]}_{\sB(M)} = \bb_{\sB(M)}$. Hence, the central ``$\sB(M)-$block''
2784: of $\bb$ is $\gT-$admissable. This is true for every $M$;
2785: we conclude that $\bb$ is $\gT-$admissable. Thus, the set $\inn{\gT}$ is
2786: nonempty, since it contains $\bb$.
2787: \end{claimproof}
2788:
2789: \claim{The set $\bT$ is {\bf r.e.}.}
2790: \begin{claimproof}
2791: Fix $\gT \subset \gA^\sU$.
2792: For any finite $N$, it is {\bf r.d} whether
2793: or not $\gA^{\sB(N)}$ contains a $\gT-$admissable configuration (there are
2794: only a finite number of cases to check).
2795: Suppose we perform this procedure for every $N \in \Natur$. By
2796: Claim $1$, if $\gT \in \bT$, then we will eventually find an $N$
2797: where no $\gT-$admissable configuration exists. Thus, we have
2798: a procedure which will halt if $\gT \in \bT$, and tell us so.
2799: \end{claimproof}
2800:
2801: As a consequence, since $\bN$ is {\em not} {\bf r.d},
2802: we conclude that $\bN$ is not even {\bf r.e.}.
2803:
2804: \claim{Suppose that $\RextMeas{\gA^\sU}$ was {\bf r.e.}.
2805: Then $\bN$ is also {\bf r.e.}.}
2806: \begin{claimproof}
2807: Clearly, $\bN = \set{ \gT \subset \gA^\sU}{ \btxt{for some \etxt} \ \mu \in
2808: \RextMeas{\gA^\sU}, \ \ \supp{\mu} = \gT}$. Hence, any recursive
2809: procedure for enumerating the elements of $\RextMeas{\gA^\sU}$
2810: would also provide a means for enumerating the elements of $\bN$.
2811: \end{claimproof}
2812:
2813: By contradiction, $\RextMeas{\gA^\sU}$ cannot be {\bf r.e.}.
2814: \end{thmproof}
2815:
2816:
2817:
2818:
2819:
2820:
2821:
2822:
2823:
2824:
2825:
2826:
2827:
2828:
2829:
2830: \section{Conclusion
2831: \label{conclusion}}
2832:
2833: Although $\extMeas{\gA^\sU}$ itself is not recursively denumerable,
2834: both its complement and topological interior are (Section
2835: \ref{decide}). As yet, however, no efficient procedure exists for
2836: determining when a locally stationary measure is extendible. So far
2837: the only substantive result in this direction is Theorem
2838: \ref{ext.per}, which says, loosely, that if $\mu_\sU$ is
2839: ``sufficiently close'' to a periodically extendible measure with full
2840: support, then $\mu_\sU$ itself is periodically extendible.
2841:
2842: The existence of mixing, ergodic, etc. extensions is
2843: well-characterized in Section \ref{mix.extend}. However, as yet, no
2844: useful work has been done characterizing the {\bf entropy} of these
2845: extensions. In particular, we might ask: given that $\mu_\sU$ is
2846: extendible, what do the {\bf maximal-entropy} extensions of $\mu_\sU$
2847: look like? Is the maximal-entropy extension {\em unique}? Does it
2848: possess some kind of ``Markov'' property, analogous to the Markov
2849: Extension in $\Zahl$? Perhaps it is some kind of {\bf Markov Random
2850: Field} \cite{Rozanov}. Indeed, in general, what would a ``Markov
2851: extension'' of a locally stationary measure look like, if anything?
2852: In the nonprobabilistic, purely symbolic setting, the construction
2853: analogous to a Markov extension is a $\Zahl^D$-subshift of finite
2854: type, but these are still poorly understood. Even topological Markov
2855: shifts ---the simplest subshifts of finite type ---do not generalize
2856: easily to higher dimensions
2857: \cite{MarkleyPaul1}. The maximal entropy measures for such subshifts
2858: have been studied in \cite{MarkleyPaul2}; \ perhaps similar techniques
2859: can be applied to maximal-entropy extensions of probability measures.
2860:
2861:
2862: \paragraph*{Acknowledgements:}
2863:
2864: I would like to thank Andres del Junco, Jeremy Quastel, and Reem
2865: Yassawi for their advice and suggestions.
2866:
2867: \bibliographystyle{plain}
2868: \bibliography{../../bibliography}
2869:
2870:
2871: \hrulefill
2872:
2873: \begin{verbatim}
2874: Marcus Pivato
2875: Department of Mathematics,
2876: University of Houston
2877:
2878: Email: pivato@math.toronto.edu
2879: \end{verbatim}
2880: \end{document}
2881:
2882:
2883: