Is there a known correlation structure among the maximums of a Brownian motion on disjoint intervals ?

Let $(W_t)_{t\geq 0}$ be a one-dimensional standard Brownian motion, and take the partition $0=t_0< t_2 <...< t_i, t_j <...t_N=T$ and define the maximum of the Brownian motion on the interval $[t_i,t_{i+1})$ as $Y_{t_i,t_{i+1}}:=\sup_{t_i\leq t < t_{i+1}}W_t.$

My question is to find the function $f(t_i,t_{i+1},t_j,t_{j+1})$ characterizing the correlation between $Y_{t_i,t_{i+1}}$ and $Y_{t_j,t_{j+1}}$, for all $i,j$. That is, $E[Y_{t_i, t_{i+1}}Y_{t_j, t_{j+1}}]=f(t_{i},t_{i+1},t_{j},t_{j+1})$, $i \neq j$.

  • $\begingroup$ Hi, as $Y_{t_i,t_{i+1}}=|B_{t_{i+1}−t_i}|+Y_{t_i}$ (equality in law) the expectation of the product shoud be tractable, shouldn't it ? $\endgroup$
    – The Bridge
    Mar 8, 2011 at 18:05

1 Answer 1


You ask for the correlation function, defined for every $0\le s\le t\le u\le v$ by the formula $$ C(s,t;u,v)=E(Y_{s,t}Y_{u,v})-E(Y_{s,t})E(Y_{u,v}). $$ One first computes $E(Y_{s,t})$. For every $t\ge0$, one introduces $$ M_t=\max\{B_s;0\le s\le t\}, $$ where $(B_t)$ is another standard Brownian motion, independent from $(W_t)$. For every $x\ge0$, Désiré André's reflection principle yields $$ P(M_t\ge x)=2P(B_t\ge x)=P(|B_t|\ge x). $$ One gets $E(M_t)=E(|B_t|)=\sqrt{2t/\pi}$. Since $Y_{s,t}$ is distributed like $W_s+M_{t-s}$, one deduces that $$ E(Y_{s,t})=\sqrt{2(t-s)/\pi}. $$ To compute $E(Y_{s,t}Y_{u,v})$, one can use the decompositions $$Y_{s,t}=W_s+M_{t-s},\qquad Y_{u,v}=W_s+B_{t-s}+Z, $$ where $Z$ is independent on everything else. (And $Z=V_{u-t}+N_{v-u}$ where $V_{u-t}$ and $N_{v-u}$ are independent and independent from everything else, $V_{u-t}$ is distributed like $W_{u-t}$ and $N_{v-u}$ is distributed like $M_{v-u}$, but one will not need this.)

The fact that $W_s$ is centered and the independence properties given above yield $$ C(s,t;u,v)=E(W_s^2)+E(B_{t-s}M_{t-s})=s+E(B_{t-s}M_{t-s}). $$ The computation of $E(B_tM_t)$ is standard. One can use once again André's reflexion principle, which says that $$ P(M_t\ge x,B_t\in\mathrm{d}y)=g_t(\max(2x-y,y))\mathrm{d}y, $$ for every $x\ge0$, where $g_t$ is the centered Gaussian density of variance $t$. Now, $$ E(B_tM_t)=\int_0^{+\infty}\mathrm{d}x\int_{\mathbb{R}}yP(M_t\ge x,B_t\in\mathrm{d}y). $$ A (carefully executed) interversion of the order of integration yields $$ E(B_tM_t)=\int_0^{+\infty}y^2g_t(y)\mathrm{d}y=\frac12E(B_t^2)=\frac12t, $$ and finally, $$ \mathrm{Cov}(Y_{s,t},Y_{u,v})=\frac12(s+t). $$ Special cases are $$ \mathrm{Cov}(Y_{s,t},B_{u})=\frac12(s+t),\quad \mathrm{Cov}(B_{t},Y_{u,v})=t, $$ and the same method yields $$ E((Y_{s,t})^2)=4t-3s. $$

  • 1
    $\begingroup$ Nice. $ $ $ $ $ $ $\endgroup$ Mar 19, 2011 at 4:04
  • $\begingroup$ This looks good !!! $\endgroup$
    – kakuritsu
    Mar 21, 2011 at 13:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.