Consider two independent linear Brownian motions $B'=(B'_t)_{t\geqslant0}$ and $B''=(B''_t)_{t\geqslant0}$, starting from $B'_0=B''_0=0$, and the process $X=(X_t)_{t\geqslant0}$ defined by $$X_t=\max\{B'_t,B''_t\}$$

What is known about the distribution of the process $X$?

The question is admittedly a little vague, hence we present a few remarks to help narrow it.

1. For each positive $t$, the PDF $f_t$ of $X_t$ is $$f_t(x)=2\varphi_t(x)\Phi_t(x)$$ where $\varphi_t$ and $\Phi_t$ are the centered normal PDF and CDF with variance $t$. Equivalently, $$f_t(x)=\frac2{\sqrt{t}}\varphi\left(\frac{x}{\sqrt{t}}\right)\Phi\left(\frac{x}{\sqrt{t}}\right)$$ where $\varphi$ and $\Phi$ are the standard normal PDF and CDF. In particular, $X$ is not a Brownian motion.

2. The process $X$ is a submartingale.

To show this in an elementary way, introduce the notations $B=(B',B'')$, and $\mathcal F^Y_t=\sigma(Y_s;s\leqslant t)$ for every time $t$ and every process $Y=(Y_t)_{t\geqslant0}$. Then, $X_t\geqslant B'_t$ almost surely hence, for every fixed $s<t$, $$E(X_t\mid \mathcal F^B_s)\geqslant E(B'_t\mid \mathcal F^B_s)=E(B'_t\mid \mathcal F^{B'}_s)=B'_s$$ By symmetry, $E(X_t\mid \mathcal F^B_s)\geqslant B''_s$ hence $E(X_t\mid \mathcal F^B_s)\geqslant X_s$. Finally, $\mathcal F^X_s\subseteq\mathcal F^B_s$ hence $$E(X_t\mid \mathcal F^X_s)=E(E(X_t\mid \mathcal F^B_s)\mid\mathcal F^X_s)\geqslant X_s$$ as desired.

3. The process $X$ is recurrent, in the sense that, for every $s$, almost surely, $$\sup_{t\geqslant s}X_t=+\infty\qquad\inf_{t\geqslant s}X_t=-\infty$$ Note that this implies that, for every nonnegative time $s$ and real number $x$, the sets of times $\{t\geqslant s\mid X_t=x\}$, $\{t\geqslant s\mid X_t\geqslant x\}$ and $\{t\geqslant s\mid X_t\leqslant x\}$ are all almost surely unbounded.

4. The process $X$ is (most probably) not Markov.

We did not write a full proof of this but the idea is that considering a (many-to-one) functional of a Markov process (these are often called hidden Markov models) usually destroys the Markov property. But one should beware that counterexamples exist, for example, $|B'|$ is Markov...

So, to begin with a precise question:

What would be a simple argument that $X$ is not a Markov process?

  • $\begingroup$ Is $f_t(x)=\varphi_t(x)\Phi_t(x)$ a short-hand notation for $f_t(x)=(2/t)\varphi(x/t)\Phi(x/t)$? $\endgroup$ – Alecos Papadopoulos Jun 11 '17 at 20:42
  • $\begingroup$ @AlecosPapadopoulos One knows that $\varphi_t$ the centered normal PDF with variance $t$ is such that $$\varphi_t(x)=\varphi(x/\sqrt{t})/\sqrt{t}$$ and that $\Phi_t$ the centered normal CDF with variance $t$ is such that $$\Phi_t(x)=\Phi(x/\sqrt{t})$$ hence $$f_t(x)=2\varphi_t(x)\Phi_t(x)=2\varphi(x/\sqrt{t})\Phi(x/\sqrt{t})/\sqrt{t}$$ $\endgroup$ – Did Jun 11 '17 at 21:02
  • $\begingroup$ Thanks, I forgot the square-root. $\endgroup$ – Alecos Papadopoulos Jun 11 '17 at 21:30
  • $\begingroup$ Not sure this qualifies as a simple argument, or how to make this precise, but looking at the history of $X_t$ we can determine if $B'$ ~ $B''$ (in probability) at the current time $t$ -- at times they are approximately equal, the process $X$ more rarely decreases. This probability informs us better of future states $X_{k+t}$ than knowledge of $X_t$ alone, violating the Markovian property. In particular if we recently observed non-decreasing behavior, we are more likely to continue observing this behavior (since $B'$ ~ $B''$). $\endgroup$ – Real Feb 28 '18 at 5:05
  • $\begingroup$ @Real Interesting argument (thanks!), which, as you say, would need a solid formalization. $\endgroup$ – Did Feb 28 '18 at 8:11

Some contributions.


Per clarification comments,

$$f_t(x)=2\phi_t(x)\Phi_t(x)=\frac 2 {\sqrt{t}}\phi(x/\sqrt{t})\Phi(x/\sqrt{t}) \tag{1}$$

This it the PDF of a Skew Normal distribution with location parameter $\xi=0 $, scale parameter $\sqrt{t}$ and shape (or "skew" or "slant") parameter $\alpha=1$.

Regarding moments, we have

$$\mathbb E(X_t) = \sqrt{\frac t{\pi}}, \;\;\; \text{Var}(X_t)= \left(1-\frac {1}{\pi}\right) \cdot t,\;\;\; \mathbb E(X_t^2) = t \tag{2}$$

while the skewness and kurtosis coefficients are the same for all $t$.

When $\alpha=1$, the CDF of the Skew Normal equals the square of the standard normal CDF,

$$P[X_t\leq x]=F_t(x) = [\Phi(x/\sqrt{t})]^2 \tag{3}$$

which for example gives us that $P[X_t\leq 0] = 1/4,\;\forall t$. So as the $\{X_t\}$ process travels along the index, although its mean and variance increase, the allocation of probability on the two sides of zero remains the same.

An interesting property of a Skew Normal r.v. with zero location parameter (not zero-mean) is its relation to the half-normal and chi-square distributions,

$$|X_t/\sqrt{t}| \sim HN(1),\;\;\; X_t^2/t \sim \chi^2_1 \tag{4}$$

..."as if" $X_t$ was a normal r.v. with zero-mean and variance $t$.


It is a known result that a Skew Normal random variable can be represented as the sum (or the difference, depending on the sign of the shape parameter) of a normal random variable and an independent half-normal r.v. For the specific parameter values of our case, let two i.i.d. normals $N_t, Z_t \sim N(0, \sigma^2=t)$. Then defining the random variable

$$Y_t = \frac 1{\sqrt{2}}\cdot (N_t + |Z_t|) \tag{5}$$

we have that $Y_t\sim X_t$ (this identity in law does not automatically extend to any joint distribution along the index of course).

For $s<t$ we have

$$P[X_s\leq x_s, X_t\leq x_t] = P[\max\{B'_s,B''_s\}\leq x_s,\max\{B'_t,B''_t\}\leq x_t]$$

$$=P[B'_s\leq x_s,B''_s\leq x_s,B'_t\leq x_t,B''_t\leq x_t]$$

$$=P[B'_s\leq x_s,B'_t\leq x_t, B''_s\leq x_s,B''_t\leq x_t]$$

$$P[B'_s\leq x_s,B'_t\leq x_t]\cdot P[B''_s\leq x_s,B''_t\leq x_t]$$

Denoting $\Phi_2(u_1,u_2;\rho)$ the CDF of the bivariate standard normal distribution with correlation coefficient $\rho$, we obtain

$$P[X_s\leq x_s, X_t\leq x_t] = F_{s,t}(x_s, x_t)=\left [\Phi_2\left(\frac {x_s}{\sqrt{s}},\frac {x_t}{\sqrt{t}};\sqrt{\frac st}\right)\right]^2 \tag{6}$$

One can see that the joint CDF leads to the marginal CDF $(3)$ as it should. The rather complicated joint PDF corresponding to $(6)$ appears not to be particularly useful. Still, having the CDF available, and it being a widely studied, tabulated and software-implemented function, permits us to calculate various threshold probabilities usually of interest like $P[X_s> 0, X_t> 0]$ for example.


Using Baye's Theorem together with $(3)$ and $(6)$ we obtain

$$P[X_t\leq x_t\mid X_s\leq x_s]=\frac {P[X_s\leq x_s, X_t\leq x_t]}{P[X_s\leq x_s]} = \left(\frac{\Phi_2\left(x_s/\sqrt{s},x_t/\sqrt{t};\sqrt{s/t}\right)}{\Phi(x_s/\sqrt{s})}\right)^2 \tag{7}$$

  • $\begingroup$ Thanks for the input. Finite marginals are indeed rather direct and the question is asking about some process properties of $X$. $\endgroup$ – Did Jul 6 '17 at 11:22

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.