2

To preface: Is anyone aware of information on a distribution similar to the Inverse Gaussian but for the case where the time-dependent process's position follows a Student's T distribution? If so, are you aware if this function is available in any programming languages, especially in its CDF form?

Consider a stochastic process involving arrival (or hitting) times for entities undergoing Brownian motion with drift. If the drift rate is known but the underlying variance is variable, what is a good way to model the arrival time distribution? I did some research on this site and in a few books but did not find any resources. I made an attempt to model the Brownian motion with variable variance as a hierarchical distribution as follows. (The work below will be similar to the derivation of the Inverse Gaussian distribution, which is covered in many places, but the approach I followed is from here.)

Brownian motion with drift, where position at time $t$ is $X_t$, with drift rate $\mu$, infinitesimal variance $\sigma^2$, and $W_t\sim\mathcal{N}(0,t)$, can be modeled as follows.

$$X_t=\mu t+\sigma W_t$$

For the distribution of $X_t$, I chose to use a gamma distribution to model the precision $(\tau^2=1/\sigma^2)$ owing to the natural conjugacy relationship (additionally it appears to be a decent model for the true distribution I am working with). For parameters $\alpha$ and $\beta$, we know that after marginalization we will obtain a t-distribution with $\nu=2\alpha$ (with $\nu\ge3$ for finite variance); we will then set $\beta$ based off a variance estimator (I skip this step here).

$$X_t\mid\tau^2\sim\mathcal{N}(\mu t,t/\tau^2), \;\tau^2\sim\mathcal{G}(\alpha,\beta)$$

We marginalize the conditional distribution to find the density for $X$ (and define $Y=\tau^2$).

\begin{align} p(X=x)&=\int_0^\infty p(X=x\mid Y=y)p(Y=y)dy\\ &=\int_0^\infty\left(\frac{\sqrt{y}}{\sqrt{2\pi t}}e^{-(x-\mu t)^2 y/2t}\right) \left(\frac{\beta^{\nu/2}}{\Gamma(\nu/2)}y^{\nu/2-1}e^{-\beta y}\right)dy\\ &=\frac{\beta^{\nu/2}}{\Gamma(\nu/2)\sqrt{2\pi t}}\int_0^\infty y^{(\nu/2+1/2)-1}e^{-((x-\mu t)^2+2t\beta) y/2t}dy\\ &=\frac{\Gamma((\nu+1)/2)\beta^{\nu/2}}{\Gamma(\nu/2)\sqrt{2\pi t}}\left(\frac{2t}{(x-\mu t)^2+2t\beta}\right)^{\nu/2+1/2} \end{align}

(The integration for the last line results from the kernel of the gamma distribution.) This is a t-distribution which we aim to invert to obtain the distribution for arrival time. We begin with the derivative of the distribution function and substitute $T=\sqrt{(x-\mu t)^2/\beta t}$ along with $dT=1/\sqrt{\beta t}dx$.

\begin{align} dP&=\frac{\Gamma(\nu/2+1/2)\beta^{\nu/2}}{\Gamma(\nu/2)\sqrt{2\pi t}}\left(\frac{2t}{(x-\mu t)^2+2t\beta}\right)^{\nu/2+1/2}dx\\ &=\frac{\Gamma(\nu/2+1/2)}{\Gamma(\nu/2)\sqrt{2\pi}}\left(\frac{2}{T^2+2}\right)^{\nu/2+1/2}dT\\ &=\frac{\Gamma(\nu/2+1/2)}{\Gamma(\nu/2)\sqrt{2\pi}}\left(\frac{2}{T^2+2}\right)^{\nu/2+1/2}\frac{dT}{dt}dt\\ &=\frac{\Gamma(\nu/2+1/2)}{\Gamma(\nu/2)\sqrt{2\pi}}\left(\frac{2}{\frac{(x-\mu t)^2}{\beta t}+2}\right)^{\nu/2+1/2}\left(\frac{1}{2}\sqrt{\frac{\beta t}{(x-\mu t)^2}}\frac{-2\mu\beta t(x-\mu t)-\beta(x-\mu t)^2}{\beta^2t^2}\right)dt\\ &=\frac{\Gamma(\nu/2+1/2)}{2\Gamma(\nu/2)\sqrt{2\pi}}\left(\frac{2\beta t}{(x-\mu t)^2+2\beta t}\right)^{\nu/2+1/2}\left(\frac{x+\mu t}{\sqrt{\beta t^3}}\right)dt \end{align}

This resulting distribution for arrival time (which I will refer to as the Inverse T or IT distribution for brevity) is the equivalent to an Inverse Gaussian (IG) distribution but if the Brownian motion's distribution at each time interval was a t-distribution rather than normal. I plotted the density function for the IT below in comparison with the IG using similar parameters for each (same position $x$, drift rate $\mu$, and overall distribution variance (functions of $\sigma^2$ for IG and $\beta$ for IT); for IT, degrees of freedom $\nu=3$). The IT distribution has a few desired properties (fat tails, right skew), and one particularly surprising property (expected value appears to increase slightly with increasing standard deviation, a property not present in the IG).

To recap, I am hoping someone would be familiar with this distribution and point me toward some useful resources. (Also, it is entirely possible I made a mistake in my math that could use correction.)

enter image description here

  • What do you mean by plotting the two pdfs "using the same parameters for each"? – Ryan Shen Aug 09 '24 at 03:48
  • 1
    @RyanShen Apologies for imprecise language here. Drift rate $(\mu)$ for the two processes was set to the same value. Infinitesimal variance $(\sigma^2)$ was set so that both distributions had the same measured variance. The mixed distribution also has degrees of freedom, which in this case was set to $\nu=3$. – NotQuiteHuman Aug 09 '24 at 13:50
  • @RyanShen The second sentence should be: Infinitesimal variance for the IG $(\sigma^2)$ and the gamma rate parameter $(\beta)$ were set so that both distributions had the same measured variance. – NotQuiteHuman Aug 09 '24 at 14:18
  • Thanks. Maybe an edit to the post is helpful for others. – Ryan Shen Aug 09 '24 at 14:22
  • If the interest lies in hitting time, do we really need to go through the intermediate step of $t$-distribution? Directly mixing the InvGau distribution with inverse-gamma seems more straightforward? – Ryan Shen Aug 09 '24 at 14:24
  • @RyanShen Thanks for the edit suggestion, I added a note on parameter selection for the plot. As for your latter comment, I believe one can directly mix the IG dist, but this is probably undesirable for two reasons: 1) the parameterization of the IG dist is not the same as the Gaussian, so a gamma mixture would apply to a function of $\lambda$ and $\mu$ rather than the simpler case for Gaussian; and 2) Brownian motion is the intuitive physical process that results in a normal distribution for $x(t)$, rather than the less intuitive $t(x)$. – NotQuiteHuman Aug 09 '24 at 20:33

1 Answers1

1

Whitmore (1986, SJS) has derived this (and a more general version of this) distribution in Normal-Gamma Mixtures of Inverse Gaussian Distributions. It also referenced the special case to Wasan, M. T. (1969). First passage time distribution of Brownian motion with positive drift (inverse Gaussian distribution). Queen's Papers in Pure and Applied Mathematics, 19, Queen's University, Kingston, Ontario.

Zack Fisher
  • 2,481