1:
2: \input rbmacros.tex
3: \def\sS{{\cal S}}
4: \def \hhalf {{\textstyle {1 \over 2}}}
5:
6: \def\bone{{\bf 1}}
7: \def\df{{\mathop {\ =\ }\limits^{\rm{df}}}}
8:
9:
10:
11:
12:
13: \centerline{\bf THE SUPREMUM OF BROWNIAN LOCAL TIMES}
14: \cl{\bf ON H\"OLDER CURVES}
15: \footnote{$\empty$}{\rm Research partially supported by NSF grant DMS-9700721.}
16:
17: \vskip0.4truein
18: \centerline{{\bf Richard F.~Bass}
19: \qq and \qq
20: {\bf Krzysztof Burdzy}}
21:
22: \vskip1truein
23:
24: {\narrower
25: {\bf Abstract}. For $f: [0,1]\to \R$, we consider $L^f_t$,
26: the local time of space-time Brownian motion on the curve
27: $f$.
28: Let $\sS_\al$ be the class of all functions whose H\"older norm of order $\al$
29: is less than or equal to 1. We show that the supremum of $L^f_1$ over
30: $f$ in $\sS_\al$ is finite is $\al>\frac12$ and infinite if $\al<\frac12$.
31: }
32:
33: \vfill\eject
34:
35:
36: \subsec{1. Introduction}
37:
38: Let $W_t$ be one-dimensional Brownian motion and let $f:[0,1]\to \R$
39: be a H\"older continuous function. There are a number of equivalent ways to define the
40: local time of $W_t$ along the curve $f$. We will show the equivalence
41: below, but for now define $L^f_t$ as the limit in probability of
42: $$\frac{1}{2\eps}\int_0^t \bone_{(f(s)-\eps, f(s)+\eps)}(W_s)\, ds$$
43: as $\eps\to 0$. Let
44: $$\sS_\al=\{f: \sup_{0\leq t\leq 1}|f(t)|\leq 1,
45: |f(s)-f(t)|\leq |s-t|^\al\hbox{ if }s,t\leq 1\}.$$
46:
47: We were led to the results in this paper by the following question.
48:
49: \proclaim Question 1.1. Is $\sup_{f\in \sS_1} L^f_1$ finite or infinite?
50:
51: \bigskip
52:
53: Our interest in this problem arose when we were working on Bass and Burdzy (1999).
54: A positive answer to Question 1.1 at that time would have provided a proof
55: of uniqueness for a certain stochastic differential equation; we ended up
56: using different methods.
57:
58: However, probably the greatest interest in Question 1.1 has to do with
59: questions about metric entropy. The metric entropy of $\sS_1$ is known to
60: be of order
61: $1/\eps$; see, e.g., Clements (1963). That is, if one takes the cardinality of the smallest $\eps$-net
62: for $\sS_1$ (with respect to the supremum norm) and takes the logarithm,
63: the resulting number will be bounded above and below by positive constants
64: times $1/\eps$. It is known (see Ledoux and Talagrand (1991)) that this
65: is too large for standard chaining arguments to be used to prove finiteness
66: of $\sup_{f\in \sS_1} L^f_1$.
67: Nevertheless, the supremum in Question 1.1 is finite.
68:
69: It is a not uncommon belief among the probability community that metric
70: entropy estimates are almost always sharp: the supremum of a process is
71: finite if the metric entropy is small enough, and infinite otherwise.
72: That is not the case here.
73: Informally, our main result is
74: \bigskip
75: \noindent{\bf Theorem 1.2}. {\sl The supremum of $f\to L^f_1$
76: over $\sS_\al$ is finite if $\al >\frac12$ and infinite if $\al < \frac12$.
77: }
78: \ms
79: \ni See Theorems 3.6 and 3.8 for formal statements.
80:
81: The metric entropy of
82: $\sS_\al$ when $\al\in (\frac12, 1]$ is far beyond what chaining methods can handle.
83: Sometimes the method of majorizing measures provides a better result
84: than that of metric entropy.
85: We do not know if this is the case here.
86:
87: For previous work on local times for space-time curves, see Burdzy and
88: San Mart\'in (1995) and Davis (1998). For some results on local times
89: on Lipschitz curves for two-dimensional Brownian motion, see
90: Bass and Khoshnevisan (1992) and Marcus and Rosen (1996).
91:
92: In Section 2 we prove the equivalence of various definitions of $L^f_t$
93: as well as some lemmas of independent interest.
94: In Section
95: 3 we prove finiteness of the supremum over $\sS_\al$ when $\al>\frac12$ and
96: that this fails when $\al<\frac12$. We also show that
97: $(f,t)\to L^f_t$ is jointly continuous on $\sS_\al \times [0,1]$
98: when $\al > 1/2$.
99:
100: The letter $c$ with subscripts will denote finite positive constants
101: whose exact values are unimportant.
102: We renumber them in each proof.
103:
104: \bs
105: \ni{\bf Acknowledgments} We would like to thank F.~Gao, E.~Gin\'e,
106: J.~Kuelbs, T.~Lyons, and J.~Wellner for their interest and help.
107: We would like to express our special gratitude to R.~Adler and M.~Barlow
108: for long discussions of the problem and many instances of specific
109: advice.
110:
111:
112: \subsec{2. Preliminaries}
113:
114: We discuss three possible definitions of $L^f_t$.
115: \item{(i)} $L^f_t=\lim_{\eps\to 0} \frac{1}{2\eps}
116: \int_0^t \bone_{(f(s)-\eps, f(s)+ \eps)}(W_s) ds$;
117: \item{(ii)} $L^f_t$ is the continuous additive functional of space-time
118: Brownian motion associated to the potential $U^f(x,t)=\int_0^{1-t} p(s,x,f(t+s)) ds$,
119: where $p$ is the transition density for one-dimensional Brownian motion;
120: \item{(iii)} (for $f\in \sS_1$ only) $L^f_t$ is the local time in
121: the semimartingale sense at 0 of the process $W_t-f(t)$.
122:
123: One of the goals of this section is to show the equivalence of these
124: definitions. We begin with the
125: following lemma which will be used repeatedly throughout the paper.
126:
127: \proclaim Lemma 2.1. Suppose $A^1_t$ and $A^2_t$ are two nondecreasing
128: continuous processes with $A^1_0=A^2_0=0$. Let $B_t=A_t^1-A_t^2$.
129: Suppose
130: that for all $s\leq t$, and some right-continuous filtration $\{\F_t\}$,
131: $$\E[A^i_t-A^i_s\mid \F_s]\leq M, \qq \hbox{a.s.}\qq i=1,2,$$
132: and
133: for all $s\leq t$
134: $$\bigl|\,\E[B_t-B_s\mid \F_s]\,\bigr|\leq \gamma, \qq \hbox{a.s.}$$
135: There exist $c_1,c_2$ such that for all $\lam >0$,
136: $$\P(\sup_{s\leq t} |B_s|>\lam\sqrt{\gamma M})\leq c_1e^{-c_2\lam}.$$
137:
138:
139: \proof We have
140: $$(B_t-B_s)^2= 2\int_s^t (B_t-B_r) dB_r.$$
141: Using a Riemann sum approximation (cf.~Bass (1995), Exercise I.8.28) we obtain
142: $$\eqalign{\E&[(B_t-B_s)^2\mid \F_s]
143: =2\E\Big[\int_s^t (B_t-B_r) \, dB_r\mid \F_s\Big]\cr
144: &=2\E\Big[\int_s^t \E[B_t-B_r\mid \F_r] \, dB_r\mid \F_s\Big]\cr
145: &\leq 2\E\Big[\int_s^t \gamma (dA_r^1+dA_r^2)\mid \F_s\Big]\leq
146: 4\gamma M.\cr}$$
147: This inequality holds a.s.\ for each $s$. The left hand side is equal
148: to $$\E[B_t^2\mid \F_s]-2B_s\E[B_t\mid \F_s]+B_s^2$$ and hence is right
149: continuous. Therefore there is a null set outside of which
150: $$\E[(B_t-B_s)^2\mid F_s]\leq 4\gamma M$$ for all $s$.
151: In particular, if $T$ is a stopping time, by Jensen's inequality we
152: obtain
153: $$\E[|B_t-B_T|\, \mid \F_T]\leq (\E[ (B_t-B_T)^2\mid \F_T])^\half
154: \leq (4\gamma M)^\half.$$
155: Our result now follows by Bass (1995, Theorem I.6.11), and Chebyshev's inequality.
156: \qed
157:
158: Let $W_t$ be one-dimensional Brownian motion.
159: Define $$p(t,x,y)=(2\pi t)^{-\half} \exp(-|x-y|^2/2t),\eqno (2.1)$$
160: the transition
161: density of one dimensional Brownian motion.
162: In the rest of the paper, $\F_t$ will denote the (right-continuous)
163: filtration generated by $W_t$.
164:
165: For a measurable function $f:[0,1]\to \R$
166: set $\norm{f}=\sup_{t\leq 1}|f(t)|$.
167: Let $$D^f_t(\eps)=\frac{1}{2\eps}\int_0^t
168: \bone_{(f(s)-\eps, f(s)+\eps)}(W_s)\, ds.
169: $$
170:
171: \proclaim Proposition 2.2. For $f$ measurable on $[0,1]$,
172: there exists a nondecreasing continuous process $L^f_t$ such that
173: $\E \norm{D^f(\eps)-L^f}^2\to 0$ as $\eps\to 0$.
174:
175:
176: \proof
177: Let $\E^{(x,t)} $ denote the expectation corresponding
178: to the distribution of Brownian motion starting
179: from $x$ at time $t$, i.e., satisfying $W_t = x$.
180: For any $x$ and any $t\leq 1$,
181: $$\eqalignno{\E^{(x,t)} \frac{1}{2\eps}\int_0^{1-t}
182: \bone_{(f(t+s)-\eps,f(t+s)+\eps)}(W_{t+s})\, ds
183: &=\frac{1}{2\eps}\int_0^{1-t} \int_{f(t+s)-\eps}^{f(t+s)+\eps} p(s,x,y)\, dy\, ds
184: \cr&\leq c_1\int_0^{1-t} \frac{1}{\sqrt{ s}} \, ds\leq c_2\sqrt{1-t}\leq c_2.
185: &(2.2)\cr}$$
186: This implies that,
187: $$\E[ D^f_1(\eps)-D^f_t(\eps)\mid \F_t]=\E^{(W_t,t)}
188: \frac{1}{2\eps}\int_0^{1-t} \bone_{(f(t+s)-\eps,f(t+s)+\eps)}(W_{t+s})\, ds
189: \leq c_2.\eqno(2.3)$$
190: The supremum of
191: $$\frac{1}{2\eps} \int_{f(t+s)-\eps}^{f(t+s)+\eps} p(s,x,y)\, dy
192: $$
193: over $\eps>0$, $t\leq 1$ and $s\leq 1-t$ is bounded.
194: By the continuity of $p(s,x,y)$ in $y$ and the bounded convergence
195: theorem, as $\eps\to 0$,
196: $$\frac{1}{2\eps}\int_0^{1-t} \int_{f(t+s)-\eps}^{f(t+s)+\eps} p(s,x,y)\, dy\, ds
197: \to \int_0^{1-t} p(s,x,f(t+s))\, ds$$
198: uniformly over $x$ and $t$.
199: Calculations similar to those in (2.2) and (2.3) yield
200: the following estimate: for any $\eta>0$,
201: $$\left|\E[(D^f_1(\eps_1)-D^f_1(\eps_2))
202: -(D^f_t(\eps_1)-D^f_t(\eps_2))\mid \F_t]\right|
203: \leq \eta, \qq \hbox{a.s.},\eqno(2.4)$$
204: for all $t\leq 1$ provided $\eps_1$ and
205: $\eps_2$ are small enough.
206:
207: Because of (2.3) and (2.4), we can apply Lemma 2.1 with
208: $A^1_t = D^f_t(\eps_1)$ and $A^2_t = D^f_t(\eps_2)$.
209: The estimate in that lemma shows that,
210: in a sense, the supremum of the difference
211: between $ D^f_t(\eps_1)$ and $ D^f_t(\eps_2)$
212: is of order $\sqrt{\eta}$.
213: We see that $\E(\norm{D^f(\eps_1)-D^f(\eps_2)}^2)\to 0$
214: as $\eps_1, \eps_2\to 0$.
215: This implies that $\{D^f(\eps_n)\}$ is a Cauchy sequence, and therefore
216: $D^f(\eps_n)$ converges as $n\to \infty$, for any sequence
217: $\{\eps_n\}$ converging to $0$.
218: Denote the limit by $L^f_t$; it is routine to check that
219: the limit does not depend on the sequence $\{\eps_n\}$.
220: Since the convergence
221: is uniform over $t$ and $t\to D^f_t(\eps)$
222: is continuous for every $\eps$, then $L^f_t$ is continuous in $t$.
223: For a similar reason, $t\to L^f_t$ is nondecreasing.
224: \qed
225:
226: \ni{\bf Remark 2.3.}
227: A very similar proof shows that $L^f_t$ is the limit in
228: $L^2$ of $$\frac{1}{\eps} \int_0^t \bone_{[f(s), f(s)+\eps)}(W_s) ds.$$
229:
230: \ms
231: \noindent{\bf Remark 2.4}.
232: Let
233: $$U^f(x,t)=\int_0^{1-t} p(s,x,f(t+s)) \, ds.$$
234: A straightforward limit argument shows that
235: $$\E[L^f_1-L^f_t\mid \F_t]=\int_0^{1-t} p(s,W_t,f(t+s))\, ds.\eqno(2.5)$$
236: It follows that $U^f(W_t,t)$ is a potential
237: for the space-time Brownian motion $t\to (W_t,t)$.
238: Hence the function $U^f(x,t)$
239: is excessive with respect to space-time Brownian motion, and therefore
240: $L^f_t$ can also be viewed as the continuous additive functional
241: for the space-time Brownian motion $(W_t,t)$
242: whose potential is $U^f$.
243:
244:
245: \proclaim Corollary 2.5. Suppose $f_n\to f$ uniformly.
246: Then $ \norm{L^{f_n}-L^f}$ converges to 0 in
247: $L^2$.
248:
249: \proof {}From (2.5),
250: $$\E[L^f_1-L^f_u\mid \F_u]\leq c_1\int_0^{1-u} \frac{1}{\sqrt{ s}}\, ds
251: \leq c_2 \sqrt{1-u}\leq c_2$$
252: and
253: $$\eqalign{\bigl|\, \E[L^{f_n}_1-L^{f_n}_u\mid \F_u]&-\E[L^f_1-L^f_u\mid \F_u]\,\bigr|\cr
254: &=\Bigl|\int_0^{1-u} [p(s,W_u,f_n(u+s))-p(s,W_u,f(u+s))]\, ds\Bigr|\cr
255: &\leq \int_0^{1-u} |p(s,W_u,f_n(u+s))-p(s,W_u,f(u+s))|\, ds.\cr}$$
256: The right hand side tends to 0 by the assumption that
257: $f_n\to f$ uniformly, and the result now follows by Lemma 2.1,
258: using the same argument as at the end of the proof of Proposition 2.2.
259: \qed
260:
261: If $f$ is a Lipschitz function, then $W_t-f(t)$ is a semimartingale.
262: We can therefore define a local time for $W_t$ along the
263: curve $f$ by setting
264: $K^f_t$ to be the local time (in the semimartingale sense) at 0 of
265: $Y_t=W_t-f(t)$. That is,
266: $$K^f_t=|Y_t|-|Y_0|-\int_0^t \sgn(Y_s)\, dY_s.$$
267:
268: \proclaim Proposition 2.6. With probability one, $K^f_t=L^f_t$ for all $t$.
269:
270:
271: \proof By Revuz and Yor (1994) Corollary VI.1.9,
272: $$K^f_t=\lim_{\eps\to 0} \frac{1}{\eps}\int_0^t
273: \bone_{[0,\eps)}(Y_s) d\angel{Y}_s. \eqno (2.6)$$
274: Since $Y_t=W_t-f(t)$, then $\angel{Y}_t=\angel{W}_t=t$, and so by
275: Remark 2.3, $K^f_t=L^f_t$ a.s. Since both $K^f_t$ and $L^f_t$ are continuous
276: in $t$, the result follows.
277: \qed
278:
279: \subsec{3. The supremum of local times}
280:
281: Our first goal is to obtain an estimate on the number of rectangles
282: of size $(1/N)\times (2/\sqrt{N})$ that are hit by a Brownian path.
283: Fix any $a\in \R$ and $b\in (a, a+2/\sqrt N]$.
284: Let
285: $$I_j=\{ \exists t\in [(j-1)/N, j/N]: a\leq W_t\leq b)\},$$
286: and
287: $$A_k=\sum_{j=1}^k \bone_{I_j}.$$
288:
289: \proclaim Lemma 3.1. There exist $c_1$ and $ c_2$ such that for all $\lam>0$,
290: $$\P(A_k\geq \lam \sqrt k )\leq c_1e^{-c_2\lam}.$$
291:
292: \proof There is probability $c_3>0$ independent of $x$ such that
293: $$\P^x(\sup_{s\leq 1/N} |W_s-W_0|
294: <1/\sqrt N)>c_3.$$
295: So by the strong Markov property applied at the first
296: $t\in [(j-1)/N, j/N]$ such that $ a\leq W_t\leq b$,
297: $$c_3\P^x(I_j)\leq \P^x(W_{j/N}\in [a-(1/\sqrt N), a+(3/\sqrt N)]).$$
298: This and the standard bound
299: $$\P^x(W_t\in [c,d])=\int_c^d \frac{1}{\sqrt{2\pi t}}e^{-|y-x|^2/2t}dy\leq \frac{1}{\sqrt {2\pi t}}|d-c|,$$
300: imply that
301: $$\P^x(I_j)\leq c_4\frac{1}{\sqrt N}\frac{1}{\sqrt{j/N}}=\frac{c_4}{\sqrt j}.$$
302: Therefore
303: $$\E^x A_k=\sum_{j=1}^k \P(I_j)\leq c_5 \sqrt k.\eqno(3.1)$$
304: By the Markov property,
305: $$\E[A_k-A_i\mid \F_{i/n}]\leq 1+\E^{W(i/n)} A_k\leq c_6\sqrt k.\eqno(3.2)$$
306: Corollary I.6.12 of Bass (1995) can be applied to the sequence
307: $A_k/(c_7\sqrt{k})$, in view of (3.1) and (3.2). That
308: result say that $\E \exp (c_8 \sup_k A_k/(c_7\sqrt{k})) \leq 2$
309: for some $c_8 >0$. This easily implies our lemma.
310: \qed
311:
312:
313: Fix an integer $N>0$.
314: Let $R_{\ell m}= R_{\ell m}(N)$ be the rectangle defined by
315: $$R_{\ell m}= [\ell/N, (\ell+1)/N]\times [m/N^\al, (m+1)/N^\al],
316: \qq 0\leq \ell\leq N,
317: \quad -N^\al-1\leq m\leq N^\al.$$
318: Let $K$ be
319: such that $N/K$ is an integer and $\sqrt N < N/K\leq \sqrt N+1$.
320: Set
321: $$Q_{ik}=Q_{ik}(N) = [iK/N, (i+1)K/N]\times [k(K/N)^\al,(k+1)(K/N)^\al],$$
322: for $0\leq i\leq K$ and $ -(N/K)^\al-1 \leq k\leq (N/K)^\al$.
323: Note that $Q_{ik}(N) = R_{ik}(N/K) $ but it will
324: be convenient to use both notations.
325:
326: \proclaim Proposition 3.2.
327: Let $\al \in (1/2,1]$ and $\eps\in (0,1/16)$.
328: There exist $c_1, c_2$, and $c_3$ such that:
329: \item{(i)} there exists a set $D_N$ with $\P(D_N)\leq c_1N\exp(-c_2N^{\eps/2})$;
330: \item{(ii)} if $\omega\notin D_N$ and $f\in \sS_\al$,
331: then there are at most $c_3 N^{(3/4)+(\eps/2)}$
332: rectangles $R_{\ell m}$ in $[0,1]\times [-1,1]$ which contain
333: both a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
334:
335:
336: \proof
337: Let
338: $$I_{ikj}=\{ \exists t\in [iK/N+(j-1)/N, iK/N+j/N]:
339: k(K/N)^\al\leq W_t\leq (k+1)(K/N)^\al \},$$
340: $$A_{ik}=\sum_{j=1}^K \bone_{I_{ikj}},$$
341: and
342: $$C_{ik}=C_{ik}(N)=\{A_{ik}\geq K^{(1/2)+\eps}\}.$$
343: By Lemma 3.1 with $k=[K]$ and $\lam = K^\eps$, and the Markov property
344: applied at $kK/N$
345: we have $\P(C_{ik})\leq c_4\exp(-c_5 K^{\eps})$.
346:
347: There are at most $c_6N^{(1/2)+(\al/2)}$ rectangles $Q_{ik}$, so if
348: $D_N=\cup_{i,k} C_{ik}$, where
349: $0\leq i\leq K$ and $ -(N/K)^\al-1 \leq k\leq (N/K)^\al$,
350: then
351: $$\P(D_N)\leq c_7 N^{(1+\al)/2} \exp(-c_5 K^{\eps})
352: \leq c_7 N \exp(-c_8 N^{\eps/2}).$$
353:
354: Now suppose $\omega\notin D_N$. Let $f$ be
355: any function in $\sS_\al$. If $f$ intersects $Q_{ik}$ for some $i$ and
356: $k$, then $f$ might intersect $Q_{i,k-1}$ and $Q_{i,k+1}$.
357: But because $f\in\sS_\al$, it cannot
358: intersect $Q_{ir}$ for any $r$ such that $|r-k|>1$. Therefore
359: $f$ can intersect at most $3(K+1)$ of the $Q_{ik}$.
360:
361: Look at any one of the $Q_{ik}$ that $f$ intersects. Since $\omega\notin
362: D_N$, then there are at most $K^{(1/2)+\eps}$ integers $j$ that
363: are less than $K$ and for which the path of $W_t(\omega)$ intersects
364: $([iK/N+(j-1)/N,iK/N+j/N]\times[-1,1])\cap Q_{ik}$.
365: If $f$ intersects a rectangle $R_{\ell m}$,
366: then it can intersect a rectangle $R_{\ell r}$
367: only if $|r-m|\leq 1$, since $f\in\sS_\al$.
368: Therefore there are at most $3K^{(1/2)+\eps}$ rectangles
369: $R_{\ell m}$ contained
370: in $Q_{ik}$ which contain both a point of the graph of $f$
371: and a point of the graph of $W_t(\omega)$.
372:
373: Since there are at most $3(K+1)$ rectangles $Q_{ik}$ which contain
374: a point of the graph of $f$, there are therefore at most
375: $$3(K+1) 3K^{(1/2)+\eps}\leq c_9 N^{(3/4)+(\eps/2)}$$
376: rectangles $R_{\ell m}$ that
377: contain both a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
378: \qed
379:
380: \bs
381:
382: We can now iterate this to obtain a better estimate.
383:
384:
385: \proclaim Proposition 3.3. Fix $\al\in(1/2,1]$
386: and $\delta,\eta>0$. There exist $c_1$ and $N_0$
387: such that if $N\geq N_0$:
388: \item{(i)} there exists a set $E$ with $\P(E)\leq \eta$;
389: \item{(ii)} if $\omega\notin E$ and $f\in \sS_\al$,
390: then there are at most
391: $c_1 N^{(1/2)+\delta}$ rectangles $R_{\ell m}(N)$ contained in $[0,1]\times [-1,1]$
392: which contain
393: both a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
394:
395:
396: \proof
397: For any $\eps$, the quantity $c_1N\exp(-c_2N^{\eps/2})$ is summable.
398: First choose $\eps\in (0,\delta/4)$ and
399: then choose $N_1$ large so that, using Proposition 3.2 and its notation,
400: $$\sum_{N=N_1}^\infty \P(D_N) \leq \sum_{N=N_1}^\infty
401: c_1N\exp(-c_2N^{\eps/2}) <\eta.$$
402: Let $E=\cup_{N=N_1}^\infty D_N$.
403:
404: Fix $\omega\notin E$.
405: Suppose $N$ is large
406: enough so that $\sqrt N\geq 2N_1$.
407: Recall the definition of $K$ and note that $N/K$
408: differs from $\sqrt{N}$ by at most $1$.
409: Then by Proposition 3.2 applied with $N/K$,
410: there are at most $c_2(\sqrt N)^{(3/4)+\eps}$ rectangles
411: $R_{ik}(N/K)$
412: that contain both a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
413: Recall the definitions of the events $C_{ik}$ and $D_N$ from
414: Proposition 3.2 and its proof. Since we are assuming that
415: $\omega \notin E$, we also have $\omega \notin C_{ik}(N)$
416: for any $i,k$. This implies that
417: inside each rectangle $R_{ik}(N/K)$, there are at most
418: $c_3(\sqrt N)^{(1/2)+\eps}$ rectangles $R_{\ell m}(N)$ that contain both
419: a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
420: Thus there are at most
421: $$c_4(\sqrt N)^{(3/4)+\eps}(\sqrt N)^{(1/2)+\eps}=c_4N^{(5/8)+\eps}$$
422: rectangles $R_{\ell m}(N)$ that contain both
423: a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
424:
425: We continue
426: iterating: take $N$ large so that $N\geq (4N_1)^4$. There are
427: $c_4(\sqrt N)^{(5/8)+\eps}$ rectangles $R_{\ell m}(N/K)$
428: that contain both
429: a point of the graph of $f$ and a point of the graph of $W_t(\omega)$.
430: Each of these contains at most $c_5(\sqrt N)^{(1/2)+\eps}$
431: rectangles $R_{\ell m}(N)$ that contain both
432: a point of the graph of $f$ and a point of the graph of $W_t(\omega)$,
433: for a total of
434: $$c_6(\sqrt N)^{(5/8)+\eps}(\sqrt N)^{(1/2)+\eps}=c_6N^{(9/16)+\eps}$$
435: rectangles $R_{\ell m}(N)$.
436:
437: Continuing, if $N$ is large enough, we can get the exponent of
438: $N$ as close to $(1/2)+\eps$ as we like. In particular, by a finite
439: number of iterations, we can get the exponent less than $(1/2)+\delta$.
440: \qed
441:
442:
443: \bs
444:
445: Recall the definition of $p(t,x,y)$ in (2.1).
446:
447:
448: \proclaim Lemma 3.4. If $\norm{f-g}\leq \eps$, then for some constant
449: $c_1$ and all $\eps < \frac12$,
450: $$\int_0^1 |p(t,0,f(t))-p(t,0,g(t))|\, dt\leq c_1\eps\log (1/\eps).$$
451:
452: \proof For $t\leq \eps^2$, we use the estimate $p(t,0,x)\leq c_2 t^{-\half}$
453: and obtain
454: $$\int_0^{\eps^2} |p(t,0,f(t))-p(t,0,g(t))|\, dt
455: \leq 2c_2\int_0^{\eps^2} \frac{1}{\sqrt t}dt\leq c_3\eps.$$
456: For $t\geq \eps^2$, note that
457: $$\Bigl|\frac{\del p(t,0,x)}{\del x}\Bigr|
458: =c_4 t^{-\half} \frac{|x|}{t}e^{-x^2/2t}= c_4 t^{-1}
459: \frac{|x|}{\sqrt t} e^{-x^2/2t}\leq c_5 t^{-1},$$
460: since $|y|e^{-y^2/2}$ is bounded. We then obtain
461: $$\int_{\eps^2}^1 |p(t,0,f(t))-p(t,0,g(t))|\, dt
462: \leq \int_{\eps^2}^1 |f(t)-g(t)| c_5 t^{-1}dt
463: \leq c_5\eps \int_{\eps^2}^1 t^{-1}\,dt=c_6\eps\log(1/\eps).$$
464: Adding the two integrals proves the lemma.
465: \qed
466:
467:
468:
469: \proclaim Proposition 3.5. Let $f$ and $g$ be two functions
470: with
471: $$\sup_{(j-1)/N\leq t\leq j/N} |f(t)-g(t)|\leq \delta.$$
472: Then, for all $\lam >0$,
473: $$\P\big(|(L^f_{j/N}-L^f_{(j-1)/N}) - (L^g_{j/N}-L^g_{(j-1)/N})|
474: \geq \lam N^{-1/4}(\delta\log (1/\delta))^\half\big)\leq c_1 e^{-c_2\lam}.$$
475:
476: \proof Write $s$ for $(j-1)/N$ and $A^f_t=L^f_{s+t}-L^f_s$,
477: $A^g_t=L^g_{s+t}-L^g_s$. We have
478: for $s\leq r\leq t\leq s+(1/N)$,
479: $$\E[A^f_t-A^f_r\mid \F_r]=\E^{W_r} A^f_{t-r}
480: \leq \sup_z \E^z A^f_{1/N}.$$
481: But for any $z$,
482: $$\E^z A^f_{1/N}=\int_0^{1/N} p(t,z,f(t))\, dt
483: \leq \int_0^{1/N} \frac{1}{\sqrt t}dt\leq c_3 N^{-\half}.$$
484: We have a similar bound for $\E^z A^g_{1/N}$.
485: For the difference, we have
486: $$|\E[ (A^f_t-A^g_t)-(A^f_r -A^g_r)\mid \F_r]|
487: =|\E^{W_r} [A^f_{t-r}-A^g_{t-r}]|.$$
488: However, for any $z$,
489: $$\eqalign{|\E^z[[A^f_{t-r}-A^g_{t-r}]| &=\Bigl|\int_s^{s+t-r} [p(u,z,f(u))
490: -p(u, z,g(u))] du\Bigr|\cr
491: &\leq \int_0^1 |p(u, 0, \wt f(u))-p(u,0, \wt g(u))|\, du,\cr}$$
492: where we define $\wt f(u)=f(u)-z$ for all $u$ and we
493: define $\wt g(u)=g(u)-z$ if $s\leq u\leq s+(t-r)$ and
494: $\wt g(u)=\wt f(u)$ otherwise. So $\norm{\wt f(u)-\wt g(u)}\leq \delta$,
495: and by Lemma 3.4,
496: $$|\E^z[[A^f_{t-r}-A^g_{t-r}]| \leq c_4\delta\log(1/\delta).$$
497:
498: Our result now follows by Lemma 2.1.
499: \qed
500:
501:
502: \proclaim Theorem 3.6. For any $\al \in (1/2,1]$,
503: there exists $\wt L^f_t$ such that
504: \item{(i)} for each $f\in \sS_\al$, we have
505: $\wt L^f_t=L^f_t$ for all $t$, a.s.,
506: \item{(ii)} with probability one, $ f\to \wt L^f_1$ is a continuous
507: map on $\sS_\al$ with respect to the supremum norm, and
508: \item{(iii)} with probability one, $\sup_{f\in \sS_\al} \wt L^f_1 <\infty$.
509:
510: \proof
511:
512: \noindent{\it Step 1}. In this step, we will define
513: and analyze a countable dense family of functions in $\sS_\al$.
514:
515: Let $N=2^n$ and let $T_n$ denote the class of functions $f$ in $\sS_\al$
516: such that
517: on each interval $[(j-1)/N, j/N]$ the function $f$ is linear with
518: slope either $N^{1-\al}$ or $-N^{1-\al}$ and $f(j/N)$ is a multiple
519: of $1/N^\al$ for each $j$.
520: Note that the collection of all functions which are piecewise linear
521: with these slopes contains some functions which are not in $\sS_\al$-- such
522: functions do not belong to $T_n$.
523:
524: Consider any element $h$ of $\sS_\al$. Let $h^{(n)}$ denote
525: a function in $T_n$ which approximates $h$ in the following sense.
526: We will define $h^{(n)}$ inductively on intervals of the form
527: $[(j-1)/N, j/N]$.
528: First we take the initial value $h^{(n)}(0)$ to be the
529: closest integer multiple of $1/N^\al$ to $h(0)$ (we take the smaller
530: value in case of a tie).
531: The slope of $h^{(n)}$ is chosen to be positive on
532: $[0, 1/N]$ if and only if $h^{(n)}(0) \leq h(0)$. Once the function
533: $h^{(n)}$ has been defined on all intervals $[(j-1)/N, j/N]$,
534: $j = 1,2,\dots, k$, we choose the slope of $h^{(n)}$
535: on $[k/N, (k+1)/N]$ to be $N^{1-\al}$ if and only if
536: $h^{(n)}(k/N) \leq h(k/N)$. Strictly speaking, our definition
537: generates some functions with values in $[-1-1/N^\al, 1+ 1/N^\al]$
538: rather than in $[-1,1]$ and so $h^{(n)}$ might not belong to
539: $\sS_\al$. We leave it to the reader to check that
540: this does not affect our arguments.
541:
542: We will argue that $|h^{(n)}(t) - h(t)| \leq 2/N^\al$
543: for all $t$. This is true for $t=0$ by definition.
544: Suppose that $1/N^\al \leq |h^{(n)}(t) - h(t)| \leq 2/N^\al$ for some $t=j/N$.
545: Then the fact that both functions belong to $\sS_\al$
546: and our choice for the slope of $h^{(n)}$ easily
547: imply that the absolute value of the difference between the two functions
548: will not be greater at time $t= (j+1)/N$ than at time $t=j/N$.
549: An equally elementary argument shows that in the case
550: when $ |h^{(n)}(t) - h(t)| \leq 1/N^\al$, the distance
551: between the two functions may sometimes increase but will
552: never exceed $2/N^\al$. The induction thus proves the claim
553: for all times $t$ of the form $t=j/N$. An extension to all
554: other times $t$ is easy.
555:
556: Later in the proof we will need to consider the difference
557: between $h^{(n)}$ and $h^{(n+1)}$.
558: First let us restrict our attention to the interval
559: $[\ell/N, (\ell+1)/N]$. The estimates from the previous paragraph
560: show that $|h^{(n)}(t) - h^{(n+1)}(t)| \leq 4/N^\al$
561: on this interval.
562: Let
563: $$F_{h,\ell }=\{|(L^{h^{(n)}}_{(\ell+1)/N}-L^{h^{(n)}}_{\ell/N})
564: -(L^{h^{(n+1)}}_{(\ell+1)/N}-L^{h^{(n+1)}}_{\ell/N})|
565: \geq N^{-(1/4)-(\al/2)+\eps}\}.$$
566: By Proposition 3.5 with $\lam = N^{\eps}$, for any $h\in \sS_\al$, $\ell$
567: and $n$,
568: $$\P(F_{h,\ell })\leq c_1 \exp(-c_2N^{\eps}).$$
569:
570: There are only $N+1$ integers $\ell$ with $0 \leq \ell\leq N$.
571: For a fixed $\ell$, there are no more than $3N^\al$ possible values
572: of $h^{(n)}(\ell/N)$, and the same is true for $h^{(n)}((\ell+1)/N)$.
573: The analogous
574: upper bound for the number of possible values for each of
575: $h^{(n+1)}(\ell/N)$, $h^{(n+1)}((\ell+1/2)/N)$ and
576: $h^{(n+1)}((\ell+1)/N)$ is
577: $6N^\al$. Hence, if we let
578: $$G_N=\bigcup_{h\in \sS_\al} \bigcup_{0 \leq \ell\leq N} F_{h,\ell },$$
579: then
580: $$\P(G_N)\leq c_3 N^6 \exp(-c_2N^{\eps}).$$
581:
582: We will derive a similar estimate for $f^{(n)}$ and $h^{(n)}$,
583: where $f,h \in \sS_\al$. Let us assume that $\| f - h \| \leq 1/N^\al$.
584: Then $|f^{(n)}(t) - h^{(n)}(t)| \leq 5/N^\al$ for all $t$.
585: If we define
586: $$\wt F_{f,h,\ell }=\{|(L^{f^{(n)}}_{(\ell+1)/N}-L^{f^{(n)}}_{\ell/N})
587: -(L^{h^{(n)}}_{(\ell+1)/N}-L^{h^{(n)}}_{\ell/N})|
588: \geq N^{-(1/4)-(\al/2)+\eps}\}.$$
589: then
590: $$\P(\wt F_{f,h,\ell })\leq c_7 \exp(-c_8N^{\eps}).$$
591: Next we let
592: $$\wt G_N=\bigcup_{f,h\in \sS_\al}
593: \bigcup_{0 \leq \ell\leq N} \wt F_{f,h,\ell }.$$
594: Counting all possible paths $f^{(n)}$ and $h^{(n)}$
595: yields an estimate analogous to the one for $G_N$,
596: $$\P(\wt G_N)\leq c_9 N^5 \exp(-c_8 N^{\eps}).$$
597:
598: \medskip
599: \noindent{\it Step 2}.
600: In this step, we will prove uniform continuity of $f \to L^f_1$
601: on the set $T_\infty=\bigcup_{n=1}^\infty T_n$.
602:
603: Fix arbitrarily small $\eta,\beta>0$. Choose $\eps>0$
604: so small that
605: $(1/4)-(\al/2)+2\eps < 0$.
606: Recall the events $D_N$ from Proposition 3.2. Since
607: $\sum_N (\P(D_N)+\P(G_N)+\P(\wt G_N)) < \infty$,
608: we can take $N_0$ sufficiently large so that $\P(H)\leq \eta$, where
609: $H=\bigcup_{N=N_0}^\infty (D_N\cup G_N\cup \wt G_N)$.
610: Without loss of generality we may take $N_0$ to be an integer power of
611: $2$, say $N_0=2^{n_0}$.
612:
613: Fix an $\omega\notin H$. Consider any $f,h\in T_\infty$
614: with $\| f - h\| \leq 1/N_0^\al$.
615: Note that
616: $$|L^h_1-L^{h^{(n_0)}}_1|\leq \sum_{n=n_0}^\infty
617: |L^{h^{(n+1)}}_1-L^{h^{(n)}}_1|,\eqno(3.3)$$
618: and
619: $$|L^{h^{(n+1)}}_1-L^{h^{(n)}}_1|\leq \sum_{m=1}^{2^n}
620: |(L^{h^{(n+1)}}_{(m+1)/2^n}-L^{h^{(n+1)}}_{m/2^n}) -
621: (L^{h^{(n)}}_{(m+1)/2^n}-L^{h^{(n)}}_{m/2^n})|. \eqno (3.4)$$
622:
623: Consider $2^n=N\geq N_0$. Since $\omega \notin \bigcup _{N\geq N_0} D_N$,
624: Proposition 3.3 implies that
625: there are at most $c_1 N^{(1/2)+\eps}$ values of $m$
626: for which there is a rectangle $R_{mi}$ in which there is a point of the
627: graph of $h^{(n)}$ or of $h^{(n+1)}$ and a point of the graph of $W_t(\omega)$.
628: So there are no more than $c_1N^{(1/2)+\eps}$ summands
629: on the right hand side of (3.4) that are non-zero.
630:
631: For a value of $m$ for which the summand on the right
632: hand side is nonzero, it is at most $N^{-(1/4)-(\al/2)+\eps}$,
633: because $\omega\notin \bigcup _{N\geq N_0} G_N$.
634: Multiplying the number of nonzero summands by the the largest value
635: each summand can be, we obtain
636: $$\eqalignno{|L^{h^{(n+1)}}_1-L^{h^{(n)}}_1|&\leq c_1N^{(1/2)+\eps}N^{-(1/4)-(\al/2)+\eps}\cr
637: &=c_1 N^{(1/4)-(\al/2)+2\eps}= c_1 (2^n)^{(1/4)-(\al/2)+2\eps}.& (3.5)\cr}$$
638: We have assumed that $\eps$ is so small that
639: $(1/4)-(\al/2)+2\eps <0$, so the bound in (3.5) is summable in $n$.
640: We increase $n_0$, if necessary, so that
641: $\sum_{n\geq n_0} c_1 (2^n)^{(1/4)-(\al/2)+2\eps} \leq \beta/3$.
642: Then (3.3) implies that
643: $$|L^h_1-L^{h^{(n_0)}}_1|\leq \beta/3.$$
644: Similarly,
645: $$|L^f_1-L^{f^{(n_0)}}_1|\leq \beta/3.$$
646:
647: A similar reasoning will give us a bound for
648: $|L^{f^{(n_0)}}_1 - L^{h^{(n_0)}}_1|$.
649: We have
650: $$|L^{f^{(n_0)}}_1 - L^{h^{(n_0)}}_1|
651: \leq \sum_{\ell=1}^{2^n}
652: |(L^{f^{(n)}}_{(\ell+1)/N}-L^{f^{(n)}}_{\ell/N})
653: -(L^{h^{(n)}}_{(\ell+1)/N}-L^{h^{(n)}}_{\ell/N})|.$$
654: First, the number of non-zero summands is bounded
655: by $c_1N_0^{(1/2)+\eps}$, for the same reason as above.
656: We have assumed that $\| f - h\| \leq 1/N_0^\al$,
657: so, in view of the fact that $\omega \notin \bigcup _{N\geq N_0}\wt G_N$,
658: the size of a non-zero summand is bounded by
659: $N_0^{-(1/4)-(\al/2)+\eps}$.
660: Hence,
661: $$|L^{f^{(n_0)}}_1 - L^{h^{(n_0)}}_1|
662: \leq c_1N_0^{(1/2)+\eps}N_0^{-(1/4)-(\al/2)+\eps}
663: = c_1 (2^{n_0})^{(1/4)-(\al/2)+2\eps}\leq \beta/3.$$
664:
665: By the triangle inequality, with probability greater than $1-\eta$,
666: $$|L^f_1-L^h_1| \leq \beta$$
667: if $f,h\in T_\infty$ and $\| f - h\| \leq 1/N_0^\al \df \delta(\beta)$.
668: We now fix
669: an arbitrarily small $\eta_0>0$ and a sequence $\beta_k \to 0$,
670: and find $\delta(\beta_k)>0$ such that
671: with probability greater than $1-\eta_0/2^k$,
672: $$|L^f_1-L^h_1| \leq \beta_k,$$
673: if $f,h\in T_\infty$ and $\| f - h\| \leq \delta(\beta_k)$.
674: This implies that, with probability greater than $1-\eta_0$, the function
675: $f \to L^f_1$ is uniformly continuous on $T_\infty$.
676: Since $\eta_0$ is arbitrarily small, the uniform continuity
677: is in fact an almost sure property, although the modulus
678: of continuity may depend on $\omega$.
679:
680: For an arbitrary $f\in\sS_\al$, define
681: $\wt L^f=\lim_{n\to\infty} L^{f^{(n)}}_1$.
682: By Corollary 2.5, $L^f=\wt L^f$ a.s. Therefore
683: $\wt L^f$ is a version of $L^f$.
684:
685: Since the function $f\to L^f_1$ is uniformly continuous
686: on $T_\infty$, its extension to $\sS_\al$ is uniformly
687: continuous with the same (random) modulus of continuity.
688: The family $\sS_\al$ is
689: equicontinuous, hence a compact set with respect to $\norm{\cdot}$.
690: Therefore the supremum of $\wt L^f_1$ over $\sS_\al$ is finite, a.s.
691: \qed
692:
693:
694: \noindent{\bf Remark 3.7}. It is rather easy to see
695: that, with probability one, $f\to \wt L^f_t$ is actually jointly
696: continuous on $\sS\times [0,1]$.
697: To see this, note that in the proof of Proposition 3.5 we used Proposition 2.1,
698: so what we actually proved was that
699: $$\P\left(\sup_{(j-1)/n\leq t\leq j/n}
700: |(L^f_t-L^f_{(j-1)/n})-(L^g_t-L^g_{(j-1)/n})|
701: \geq \lam N^{-1/4}(\delta \log(1/\delta))^\half\right)
702: \leq e^{-c_1\lam}. $$
703: If we replace (3.4) by
704: $$\sup_t|L^{h^{(n+1)}}_t-L^{h^{(n)}}_t|\leq
705: \sum_{m=1}^{2^n}\sup_{m/2^n\leq t\leq (m+1)/2^n}
706: |(L^{h^{(n+1)}}_t-L^{h^{(n+1)}}_{m/2^n}) -
707: (L^{h^{(n)}}_t-L^{h^{(n)}}_{m/2^n})|,$$
708: then proceeding as in the proof of Theorem 3.6,
709: we obtain the joint continuity.
710: \bs
711:
712:
713:
714:
715: We will show that, in a sense, $\sup_{f\in \sS_{\al}}L^f_1=\infty$,
716: a.s., if $\al < \half$. This statement is quite intuitive
717: -- one would like to let $f(\omega)=W_t(\omega)$ so that $L^f_1(\omega)=\infty$ --
718: but we have not defined the local time simultaneously
719: for all $f\in \sS_{\al}$, and there is a difficulty with the
720: number of null sets. Theorem 3.6 suggests that
721: the question of joint existence is tied to the question
722: of the finiteness of the supremum, so we have
723: to express our result in a different way.
724:
725:
726: \proclaim Theorem 3.8. Suppose $\al<\half$.
727: Then there exists a countable family $F\subset \sS_\al$ such that
728: $\sup_{f\in F}L^f_1=\infty$ a.s.
729:
730:
731:
732: \proof
733: Let $\ell^x_t$ be the ordinary local time at $x$ for Brownian motion.
734: It is well known (see Karatzas and Shreve (1994)) that
735: there exists a version of this process which is jointly
736: continuous in $x$ and $t$.
737:
738: Suppose that a piecewise linear function $f$ is equal to $y$ on an interval
739: $[s,t]$. Then Proposition 2.2 and a similar well known
740: result for $\ell^y$ show that with probability one,
741: for all $u\in [s,t]$,
742: $$L^f_u - L^f_s = \ell^y_u - \ell^y_s.$$
743:
744: Fix $\al \in (0,1/2)$.
745: Let $F$ be the countable family of all functions $f$
746: defined on the interval $[0,1]$ such that for some
747: integers $n=n(f)$ and $m=m(f)$,
748: on each interval of the form
749: $[(j-1)/n, (j-\frac12)/n]$ the function $f$ is a
750: constant multiple of $2^{-m}$,
751: $f$ is linear on the intervals $[(j-\frac12)/n, j/n]$, and $f\in \sS_\al$.
752: Then, with probability one, for all $j$, all $f\in F$ and $n=n(f)$,
753: $$ L^{f((j-1)/n)}_{(j-(1/2))/n}- L^{f((j-1)/n)}_{(j-1))/n}
754: =\ell^{f((j-1)/n)}_{(j-(1/2))/n}- \ell^{f((j-1)/n)}_{(j-1))/n}. \eqno(3.6)$$
755: In the rest of the proof we assume that this assertion
756: and the joint continuity of $\ell^x_t$ hold for all $\omega$.
757:
758:
759: Let
760: $$T=\inf\{t: |W_t|\geq 1\hbox{ or } \exists r,s\leq t
761: \hbox{ such that } |W_r-W_s|\geq (\textstyle{\frac14} |r-s|)^\al\}.\eqno(3.7)$$
762: By the well-known results on the modulus
763: of continuity for Brownian motion, $T>0$ a.s.
764:
765: Let $\eps>0$. There exists $\delta$ such that $\P(T<\delta)<\eps$.
766: Fix $n$.
767: On the interval $[(j-1)/n,(j-\frac12)/n]$,
768: let $ f_1(t)= W((j-1)/n)$. On
769: the interval
770: $[(j-\frac12)/n,j/n]$ let $ f_1(t)$ be linear with
771: $ f_1(j/n)=W(j/n)$. Let $f_2(t)= f_1(t)$ for $t\leq \delta/2$ and
772: constant for $t\geq \delta/2$.
773:
774: It is quite easy to show that $f_2\in \sS_\al$ for each $\omega$ in
775: the set $\{T>\delta\}$ using the definition (3.6) of $T$.
776: By the Markov property, the random variables
777: $$X_j=\ell^{f_2((j-1)/n)}_{(j-(1/2))/n}- \ell^{f_2((j-1)/n)}_{(j-1))/n}$$
778: form an independent sequence, and by Brownian scaling,
779: $Y_j=\sqrt{2n} X_j$ has the same distribution as $\ell^0_1$. Let $c_1=\E \ell^0_1$.
780: By Chebyshev's inequality,
781: $$\P\Big(\Bigl| \sum_{j=1}^{[\delta n/2]}
782: (Y_j-c_1)\Bigl|\geq c_1\delta n/4\Big)
783: \leq \frac {[\delta n/2] {\rm Var}\, Y_1}{ (c_1\delta n/4)^2}
784: \leq \frac{c_2\E(\ell^0_1)^2}{\delta n}=\frac{c_3}{\delta n}.$$
785: Take $n$ large so that $c_3/(\delta n)<\eps$. Then there exists a set $A_n$
786: of probability at most $2\eps$ such that if $\omega\notin A_n$, then
787: $T(\omega)\geq \delta$ and
788: $$\sum_{j=1}^{[\delta n/2]} X_j\geq c_4\sqrt {\delta n}.$$
789:
790: We now choose $m$ large and
791: find $f_3\in F$ so that on each interval
792: $[(j-1)/n, (j-\frac12)/n]$ the function $f_3$ is a multiple of $2^{-m}$,
793: $f_3$ is linear on the intervals $[(j-\frac12)/n, j/n]$, and
794: $$\sum_{j=1}^{[\delta n/2]} \Big[ \ell^{f_3((j-1)/n)}_{(j-(1/2))/n}-
795: \ell^{f_3((j-1)/n)}_{(j-1))/n}\Big]\geq c_4\sqrt {\delta n}/2;$$
796: this is possible by the joint continuity of $\ell^x_t$.
797:
798:
799: By (3.6) we can replace $\ell$ by $L$
800: in the last formula, so
801: $$L^{f_3}_1 \geq \sum_{j=1}^{[\delta n/2]} \Big[ L^{f_3}_{(j-(1/2))/n}-
802: L^{f_3}_{(j-1))/n}\Big]\geq c_4\sqrt {\delta n}/2.$$
803: We conclude that
804: $$\sup_{f \in F} L^f_1 \geq c_4\sqrt {\delta n}/2,$$
805: with probability greater than or equal to $1-2\eps$.
806: Since $n$ and $\eps$ are arbitrary, the proposition is proved.
807: \qed
808:
809:
810: \centerline{\bf References}
811: \medskip
812:
813:
814: \item{1.}R.F. Bass (1995). {\sl Probabilistic Techniques in Analysis}.
815: Springer-Verlag, New York.
816:
817: \item{2.}R.F. Bass and D. Khoshnevisan (1992). Local times on curves and uniform
818: invariance principles. {\sl Probab. Theory
819: Related Fields \bf 92}, 465--492.
820:
821: \item{3.}R.F. Bass and K. Burdzy (1999). Stochastic bifurcation models. {\sl Ann. Probab. \bf 27}, 50--108.
822:
823: \item{4.}K. Burdzy and J. San Mart\'in (1995). Iterated law of iterated logarithm. {\sl Ann. Probab. \bf 23},
824: 1627--1643.
825:
826: \item{5.}G.F. Clements (1963). Entropies of several sets of real valued functions. {\sl Pacific J. Math. \bf 13}, 1085--1095.
827:
828: \item{6.}B. Davis, (1998). Distribution of Brownian local time on curves. {\sl Bull. London Math. Soc. \bf 30}, 182--184.
829:
830: \item{7.} I.~Karatzas and S.~Shreve (1994) {\it Brownian Motion and
831: Stochastic Calculus, Second Edition}. Springer-Verlag,
832: New York.
833:
834: \item{8.}M. Ledoux and M. Talagrand (1991).
835: {\sl Probability in Banach spaces. Iso\-per\-i\-me\-try and Processes.}
836: Springer-Verlag, Berlin.
837:
838: \item{9.}M.B. Marcus and J. Rosen (1996). Gaussian chaos and sample path properties of additive functionals of symmetric
839: Markov processes. {\sl Ann. Probab. \bf 24}, 1130--1177.
840:
841: \item{10.}D. Revuz and M. Yor (1994). {\sl Continuous Martingales
842: and Brownian Motion, 2nd ed.} Springer-Verlag, Berlin.
843:
844:
845:
846:
847: \vskip1truecm
848:
849: \parskip=0pt
850: \hbox{\vbox{
851: \obeylines
852: Richard F.~Bass
853: Department of Mathematics
854: University of Connecticut
855: Storrs, CT 06269
856: e-mail: bass@math.uconn.edu
857: {\ }
858: }
859: \hskip-8truecm
860: \vbox{
861: \obeylines
862: Krzysztof Burdzy
863: Department of Mathematics
864: University of Washington
865: Box 354350
866: Seattle, WA 98195-4350
867: e-mail: burdzy@math.washington.edu}
868: }
869:
870:
871: \bye
872: