MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

(I'm new to Poisson processes, so please edit if my terminology is incorrect.)

Edit: per comments, here is a (more) general version of the originally posted problem (which is now at the bottom, below the line); I hope it's more clear/helpful.

The context is approximating a distribution defined by the $\arg\min$ of a (non-Poisson) process in Theorem 1 of this paper.

Setup: I want to know the first two moments of $\Delta_m \stackrel{d}{=}\arg\min_u Z(u)$, where $u=(u_1,\ldots,u_d)'\in\mathbb R^d$. The $d$ elements correspond to regressors, where the first element is one (the constant term) and the other elements are regressors distributed according to measure $\mu$. Centering the non-constants at zero, $$Z(u)\equiv -(mu_1,0,0,\ldots,0)' + D + \sum_{k\ne0} \int_0^{u'X_k}\\!\left[1(\Gamma_k\le s)- 1(\Gamma_k<0)\right]\\,\mathrm ds,$$ where $m$ is a positive integer, index $k$ takes integer values, $X_k$ are iid with distribution $\mu$, $\Gamma_k=E_1+\cdots+E_k$ for $k>0$ with $E_i\stackrel{iid}{\sim}\rm{Exp}(1)$ independent of all $X_k$ also, $\Gamma_k=-(E_{-1}+\cdots+E_{-k})$ for $k<0$, and $D$ is a random (independent) vector with mean zero and known (if a bit complicated) distribution that I hope I can lay aside for now and add back later. The original paper says the points $\{(\Gamma_k,X_k')':k\ne0\}$ are points of a Poisson process with mean measure $$m(d\epsilon,dx)\equiv \lambda(d\epsilon)\mu(dx),$$ where $\lambda$ is Lebesgue measure.

With $d=2$, we can graph the points $(X_k,\Gamma_k)$ in two dimensions, using only the non-constant element of $X_k$. Picking $(u_1,u_2)'$ determines a line in this plane, $\Gamma=u_1+u_2X$. The first first-order condition (FOC) says that the number of points in the green region minus the number in the red is equal to $m$:

The second FOC is similar, but weighting each point by the $X_k$ value, and should equal zero. Or if you weight the points by $|X_k|$, the red and green for $X_k<0$ switch:

Note: these are not exact equalities, because the integral term in $Z(u)$ is discontinuous; the FOCs are more like "the smallest $u$ such that these are $\ge m$ and $\ge0$."

Question: what are the first two moments of $\Delta_m\stackrel{d}{=}\arg\min_u Z(u)$?

I think the mean is $(\Gamma_m,0)$, but can't prove it; I have no ideas for the variance.

Original post, for special case $d=1$.

Original setup: consider a Poisson process on $[0,\infty)$, with points $\{ \Gamma_k\}_{k=1}^\infty$, $\Gamma_k=E_1+\cdots+E_k$, $E_i\stackrel{iid}{\sim}\rm{Exponential}(1)$.

Consider $u=\inf\{t:1\le\sum_{k=1}^\infty \mathbb 1(\Gamma_k\le t)\}$, where $1(\cdot)$ is the indicator function (one if true, zero if false). In this special case, $u=\Gamma_1=E_1$, and thus $u\sim\rm{Exp}(1)$. We have a closed form solution for $u$, and the first two (central) moments are $\mathbb E(u)=1$ and $\rm{Var}(u)=1$.

Original question: can we derive $\mathbb E(u)=1$ and $\rm{Var}(u)=1$ without using the closed form of $u$? (In my general problem, there is no closed form.) So we know $u=\inf\{t:1\le\sum_{k=1}^\infty \mathbb 1(\Gamma_k\le t)\}$, but we can't just use $u=\Gamma_1$ to calculate moments.

Original notes/thoughts:

As a reminder, the mean measure of the considered Poisson process is $m(A)=\lambda(A)$, where $\lambda$ is Lebesgue measure (e.g., $\lambda([a,b])=b-a$; just the total length of set $A\subset [0,\infty)$). This means that the expected number of points in any interval is equal to the length of the interval.

Also, the probability of one event occurring in $[t,t+dt]$ is $dt$ since the rate is one here.

Also potentially helpful: define the random counting measure $\hat N(t)=\sum_{k=1}^\infty 1(\Gamma_k\le t)$. Then, $\mathbb E(\hat N(t))=\lambda([0,t])=t$, and $u=\inf\{t:1\le \hat N(t)\}$.

For the first moment, I noticed that in this case, $\inf\{t:1\le \mathbb E(\hat N(t))\}=\inf\{t:1\le t\}=1=\mathbb E(u)$. That is, solving for $u$ after plugging in the mean measure happens to yield the mean of $u$. This struck me as not true generally: $\mathbb E(u(X))\ne u(\mathbb E(X))$. But maybe some property of the Poisson process and/or $u$'s characterization means that $u$ is a ``linear'' function of the process, so this is generally true here?

For the second moment, I haven't made any progress. Tried thinking about the expectation $\mathbb E\left[(u-\mathbb E(u))^2\right]=\mathbb E[(u-1)^2]$, but don't know how I could get that without cheating. Also wondered if there was a ``variance of the process,'' if the plug-in approach for the first moment turns out to be valid. But basically stuck.

Last, recall either $\rm{Var}(u)$ or $\mathbb E(u^2)$ is sufficient, since $\rm{Var}(u)=\mathbb E(u^2)-(\mathbb E(u))^2$.

share|cite|improve this question
I think it would help if you made your question more concrete by telling us what you really want to calculate. For example, maybe you want to replace $1(\lambda_k\le t)$ with some other function? – Ori Gurel-Gurevich Sep 3 '11 at 3:30
It also would help if you can let us know the more general problem you are trying to solve, for which there might be a common underlying pattern as the Poisson case. For example, my first attempt at guessing the general problem is, for a non-exploding counting process N (increments are all 1), get the distributional information about its first jump time (or at least 1st and 2nd moments). Is this correct or not? – Poldavian Sep 3 '11 at 13:29
@Poldavian and Ori: thanks for the comments--I have added to my question accordingly. I hope the general problem (general $d$ instead of $d=1$ or $d=2$) is more clear now. Please (anyone) let me know of any further improvements I can make. – David M Kaplan Sep 7 '11 at 17:49

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Browse other questions tagged or ask your own question.