I'm trying to understand this proof of the following specifing integration by parts.


Let $\Omega=Point_{\mathbb{R}}$ the set of point distributions in $\mathbb{R}^3$ (i.e an element $w \in \Omega$ is a locally finite subset of $\mathbb{R}^3$). We equip this space with a canonic tribe $\mathcal{A}$ and a probability $\mathbb{P}$.

Let $\phi : \Omega \rightarrow \mathbb{R}^3$ a measurable function. For $w \in \Omega$, we define the realization of $\phi$ the mapping $R_\omega[\phi] : \mathbb{R}^3 \rightarrow \mathbb{R}^3$ :

$$R_\omega[\phi](y):=\phi(y+w)=\phi(\tau_yw), \quad y \in\mathbb{R}^3$$ where $\tau_y :\Omega \rightarrow \Omega$ is the shift application (which is measurable in this case).

Let $\varphi : \Omega \rightarrow\mathbb{R^3}$ a smooth function and $w \in \Omega$, we define the stochastic gradient by $$\bar{\nabla}(\varphi)(\omega):=\nabla (t \mapsto \varphi(t+w))(0).$$ where $\nabla$ is the regular spatial gradient. Using this definition, we define the partial derivative $\bar{\partial}_i$.

the integration by parts

Let $u,v$ smooths functions from $\Omega$ to $\mathbb{R}^3$. According to the text I'm reading, we have the following integration by parts : $$\mathbb{E} \left[ \bar{\partial}_iu v\right]=-\mathbb{E} \left[ u \bar{\partial}_i v\right]$$ The proof starts with the following equality, that I'm strugling to understand : \begin{aligned} \mathbb{E} \left[\bar{\partial}_iu v\right]= \mathbb{E} \left[\int_{K_1} \partial_i R_\omega[u](y) \ R_\omega[v](y) \ \mathrm{d}y\right] \quad \quad (\star) \end{aligned} with $K_1=\left[-\frac{1}{2},\frac{1}{2}\right]^3$. Then, we can use the regular integration by parts formula for the $K_1$ integral, which gives us :

$$- \mathbb{E} \left[\int_{K_1} R_\omega[u](y) \ \partial_i R_\omega[v](y) \ \mathrm{d}y\right]+ \underbrace{\mathbb{E} \left[\int_{ \partial K_1} n_i R_\omega[u](y) \ R_\omega[v](y) \ \mathrm{d}y \right]}_{:=A}$$ and then it is said that $A=0$. The rest of the proof follows easily using $(\star)$ again.

My questions

  1. Where does the first equality $(\star)$ comes from ?
  2. Why do we have $A=0$ ? In the usual case it is because the test functions have a their support include in a compact but we do not have such hypothesis here.

Any helps or hint are welcomed !


Actually it is necessary to add one hypothesis so that the result holds.

We need to assume that the application shift $\tau_y : \Omega \rightarrow \Omega$ is $\mathbb{P}$ preserving for any vector $y \in \mathbb{R}^3$. Using this stationarity hypothesis, we see that :

$$\mathbb{E}\left[ \bar{\partial}_i v u \right]= \mathbb{E}\left[ \bar{\partial}_i (v \circ \tau_y) \ (u \circ \tau_y) \right], \quad \forall y \in \mathbb{R}^3. $$

Integrating this equality over $y$ on the domain $K_1$ which has a Lebesgue measure equals to 1, we get by Fubini's theorem the equality $(\star)$.

A similar argument on the boundary integral leaves us with $A=\mathbb{E}[uv] \int_{\partial K_1} n_i$, and it is well known that the latter integral worths $0$, hence $(\star \star)$.

| cite | improve this answer | |

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.