I am trying to find out the the Radon-Nikodym derivative of $μ_1$ w.r.t $μ_2$ where $μ_1$ is a Gaussian measure on $R^n$ with mean $a$, and standard deviation $σ$, and $μ_2$ is also a Gaussian measure, with mean $b$ and standard deviation $σ$.

Solution Attempt

Let $n\in N$ and let $B_0(R^n)$ denote the completion of the Borel $\sigma$-algebra on $R^n$. Let $\lambda^n:B_0(R^n)\rightarrow[0,R^+]$ denote the usual $n$-dimensional Lebesgue measure.

Based on the definition of Gaussian measure and it's properties we know the following two equivalent relations $\lambda^{n} \ll \mu_1^{n} \ll \lambda^{n}$ and $\lambda^{n} \ll \mu_2^{n} \ll \lambda^{n}$ where $\ll$ stands for absolute continuity of measures; Hence we have absolute continuity relations $\mu_1^{n} \ll \lambda^{n}$ and $\lambda^{n} \ll \mu_2^{n}$.

Also since $\lambda^{n} \ll \mu_1^{n} \ll \lambda^{n}$ we can use the Radon-Nikodym property $\dfrac {d\lambda } {d\mu_1}=\left( \dfrac {d{\mu_1}} {d\lambda }\right) ^{-1}$

Since both Lebesgue and Gaussian measures are $\sigma$-finite and based on the above relations we should be apply the Radon-Nikodym Chain Rule here. $$\dfrac {d\mu_2 } {d\mu_1} = \dfrac {d\mu_2 } {d\lambda}\dfrac {d\lambda } {d\mu_1} = \dfrac {d\mu_2 } {d\lambda}\left( \dfrac {d{\mu_1}} {d\lambda }\right) ^{-1}$$

The Gaussian measures $μ_1:B_0(R^n)\rightarrow[0,R^+]$ and $μ_2:B_0(R^n)\rightarrow[0,R^+]$ are defined as

${\mu_1}(E) = (2\pi{\sigma}^2)^\frac{-n}{2}\int _{E}exp\left( -\dfrac {1} {2\sigma ^{2}}\left\| x-a\right\| _{R_{n}}^{2}\right)d\lambda ^{n}\left( x\right) $

and ${\mu_2}(E) = (2\pi{\sigma}^2)^\frac{-n}{2}\int _{E}exp\left( -\dfrac {1} {2\sigma ^{2}}\left\| x-b\right\| _{R_{n}}^{2}\right)d\lambda ^{n}\left( x\right) $ respectively.

I am unsure though how to proceed further here, any help would be much appreciated.

Thanks in advance.


$$\frac{\mathrm d\mu_1}{\mathrm d\mu_2}(x)=\exp\left(-\frac1{2\sigma^2}(\|x-a\|^2-\|x-b\|^2)\right)=\exp\left(\frac{\langle a-b,x-\frac12(a+b)\rangle}{\sigma^2}\right) $$

  • 1
    $\begingroup$ Hi Mate Thanks for your answer. I was also hoping, if you would be kind enough to give me a clue how to evaluate the same derivative if we did not know the standard deviations but had a covariance matrix instead. $\endgroup$ Oct 3 '12 at 23:05

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.