Let $$\{B^{(i)}_t : 0 \leq t \leq T, i \geq 1\}$$ be a sequence of i.i.d. standard Brownian motions. For each $n,$ let $V^{(n)}_t, \,0 \leq t \leq T$ be a continuous bounded process adapted to the filtration $\mathcal{F}_t^n$ generated by the first $n$ Brownian motions.

Let $P_n$ denote the law of $(B^{(1)}, \dots, B^{(n)})$ on $C := C([0, T], \mathbb{R}^n).$

Since $V^{(n)}$ satisfies conditions for a change of measure, we can apply a Girsanov transform to yield a measure $Q_n$ with $$\frac{dQ_n}{dP_n} = \exp\Big( \sum_{i = 1}^n \int_0^\cdot V^{(n)}_sdB^{(i)}_s - \frac{1}{2}\int_0^\cdot [V_s^{(n)}]^2 ds\Big).$$

So $Q_n$ has the same law on $C$ as $(B^{(1)} + V^{(n)}, \dots, B^{(n)} + V^{(n)}),$ an $n$ dimensional Brownian motion with a drift of $V^{(n)}$ in each coordinate.

If we define two triangular arrays whose row for each $n$ are $$R^P_n : =(B^{(1)}, \dots, B^{(n)}),$$ and $$R^Q_n := (B^{(1)} + \int_0^\cdot V^{(n)}, \dots, B^{(n)} + \int_0^\cdot V^{(n)}),$$

we immediately see the laws of the rows $R^P_n$ and $R^Q_n$ are equivalent for each finite $n,$ and in fact under the finite triangular array $R_1^P \times \cdots \times R_n^P \equiv R_1^Q \times \cdots \times R_n^Q$ via the canonical product measure.

We may embed each $R^P_n$ to induce a measure on $C^\infty = C([0, T], \mathbb{R}^{\mathbb{N}})$ and can place a standard metric on this to make it Polish. Similarly with $R^Q_n.$ Now that each $R^P_n$ has a measure on the same space, consider its filtration $\mathcal{F}_n^P$ and tail $\sigma$-field $\mathcal{F}_\infty^P.$ (Similar for $R^Q_n$). Notice that because $V^{(n)}$ is adapted, we know $\mathcal{F}_n^Q \subset \mathcal{F}^P_n$ so for our purposes here we restrict to $\mathcal{F}^P_n.$

By the Girsanov transform, we see that null sets in $\mathcal{F}_n^P$ under $R^P$ are also null sets under $R^Q.$

Let us further assume that $V^{(n)}$ converges a.s. and in $L_p$ to some bounded process $V.$

Can we assert the null sets in $\mathcal{F}_\infty$ under $R^P$ are also null under $R^Q?$

In essence, we wish to extend the $0-1$ law of a triangular array of i.i.d. Brownian motions to that of a triangular array of BMs with an adapted row-wise convergent drift.

I apologize in advance for any errors or typos! Any references would be much appreciated!


No, the measures are not necessarily equivalent on the tail $\sigma$-field.

For a concrete example, take $V_t^{(n)} = t$, a deterministic continuous process of bounded variation. Then trivially the $V_t^{(n)}$ converge to $V_t=t$ in whatever sense you want. Consider the tail random variable $X = \limsup_{n \to \infty} \frac{1}{n} \sum_{k=1}^n B^{(k)}_1$. Under $P$, each $B^{(k)}_1$ has an $N(0,1)$ distribution, and they are iid, so by the strong law of large numbers we have $P(X=0)=1$, where we note that $\{X=0\} \in \mathcal{F}_\infty$. But under $Q$, each $B^{(k)}_1$ has the law of a standard Brownian motion plus $V_1=1$, so it is $N(1,1)$. Thus we have $Q(X=1)=1$ and $Q(X=0)=0$. As such, $P,Q$ are mutually singular on $\mathcal{F}_\infty$.

You need some sort of summability criterion on the $V^{(n)}$; this is some corollary of the Cameron-Martin theorem for Gaussian measures. They should not only converge, but converge to 0, and reasonably fast. I would have to work it out, but I think you might want something like $\sum_n \mathbb{E}\left[\sup_{[0,T]} |V_t^{(n)}|^2\right] < \infty$. Bogachev's book Gaussian Measures is probably the ultimate resource here, though there may be lighter-weight sources also.

  • $\begingroup$ Thanks, Nate. In your example, the slln still holds for the sequence with drift. Do you know if under these conditions a slln holds for $n^{-1}\sum_{i=1}^nl(B^{(i)} + V^{(n)})$ where $l$ is a functional on $C([0, T], \mathbb{R})?$ $\endgroup$ – user253775 Feb 22 '17 at 3:21
  • $\begingroup$ I mean $n^{-1}\sum_{i=1}^nl(B^{(i)} + \int_0^\cdot V^{(n)}).$ $\endgroup$ – user253775 Feb 22 '17 at 4:48
  • $\begingroup$ If it's a linear functional, and if SLLN holds for $V$, then it should. $\endgroup$ – Nate Eldredge Feb 22 '17 at 5:05

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.