Converse of the Girsanov theorem?

262 Views Asked by At

This question is regarding the Girsanov theorem. Define on some filtered probability space $(\Omega,F,({\scr F}_t)_t,P)$ $$Z_t=\exp\Bigg[\sum_{i=1}^{n}\int_0^t \lambda_s^{(i)}dB^{(i)_s}-\frac{1}{2}\int_0^t \|\lambda_s\|^2 ds\Bigg]. $$ where $\lambda_i$ is such that $$\int_0^t(\lambda_s^{i})^2ds<\infty\text{ a.s. }$$ and $B$ is an $n$-dimensional Brownian motion. Define an equivalent probability measure $Q$ such that $$\frac{dQ}{dP}=Z_T\text{ on }{\scr F}_T. $$ What the Girsanov theorem (as stated in Karatzas and Shreve) says is if $Z_t$ is a martingale ($E[Z_t]=1)$ then the process $\tilde{W}$ defined by $$\tilde{W}_t=W_t- \int_0^t\lambda_s^ids\text{ for all }i=1,...,n,$$ is a Brownian motion on $(\Omega,F_T,Q)$. My question is does any probability measure $Q$ equivalent to $P$ take the form $$\frac{dQ}{dP}=\exp\Bigg[\sum_{i=1}^{n}\int_0^T \lambda_s^{(i)}dB^{(i)_s}-\frac{1}{2}\int_0^T \|\lambda_s\|^2 ds\Bigg]?$$ Or ( the converse?) if we define a $\tilde{W}$ as above for a given process $\lambda,$ does there exist an equivalent probability measure such that $$\frac{dQ}{dP}=Z_T\text{ on }F_T$$ such that $\tilde{W} $is a BM w.r.t $Q?$

I have not assumed here that the filtration is the one generated by the BM but if it helps that can be assumed. Any ideas?

1

There are 1 best solutions below

0
On

In the following answer integrals are multivariate. Assume that martingale representation with respect to the Brownian motion holds, i.e., for any ${\scr F}_T$-measurable integrable random variable $X_T$ and the corresponding $P$-martingale $X_t=E_t[X_T]$ there is an $B$-integrable predictable process $\eta$ such that $$ X = E[X_T] + \int_0^\cdot \eta_s dB_s.$$

Now take $Z$ to be the unique $P$-martingale closed by $dQ/dP > 0$. From the martingale representation theorem one obtains $$ Z = 1 + \int_0^\cdot \eta_s dB_s > 0. $$ Set $\lambda = \eta/Z$ and perform change of variables to $\ln Z$ to conclude that $Z_T=dQ/dP$ is of the desired form. The 'converse' will only hold if $\lambda$ is well behaved so that ${\scr E}(\lambda\cdot B)$ stopped at $T$ is a martingale. There are plenty of $\lambda$-s that integrate $B$ but do not have this property.