How to give a high-probability uniform estimation of a potential having access to noisy pointwise estimates of the associated vector field?

25 Views Asked by At

As in the title, our goal is to estimate uniformly and with high probability (and up to a constant) a potential having access to noisy pointwise estimates of the associated vector field (i.e., the gradient of the potential). Here how I tried to formalize the problem.

Suppose $d \in \mathbb{N}$ and let $\mathcal{B}_1^d$ be the unitary closed ball of $\mathbb{R}^d$ centered in the origin.

Let $\mathcal{F}$ be the class of all $C^1$ functions $F : \Omega \to \mathbb{R}$, where $\Omega$ is any open set of $\mathbb{R}^d$ containing $[-1,1]^d$, and such that $\forall x\in[-1,1]^d, \nabla F(x) \in \mathcal{B}_1^d$.

Suppose that, for each $F \in \mathcal{F}$, we have the following interaction protocol:

For each $t=1,2,\dots$

  1. We select a point $X_t \in [-1,1]^d$ (a selection based on past observations and possibly some randomization, but not on the knowledge of $F$)
  2. We observe a $\mathcal{B}_1^d$-valued random variable $Y_t^F$ such that $\mathbb{E}[Y_t^F \mid X_1,Y_1^F,\dots,X_{t-1},Y_{t-1}^F,X_t] = \mathbb{E}[Y_t^F \mid X_t] =\nabla F(X_t)$.

Basically, we query a point in the domain and we see a noisy (and bounded) reconstruction in that point of the vector field associated to the gradient of the potential $F$.

Regardless which potential $F \in \mathcal{F}$ we are interacting with, our goal is to give a uniform reconstruction of $F-F(0)$ with high-probability using the previous interaction protocol, i.e. we want to find a strategy to select the points $X_1,X_2,\dots$ and a family of estimators $(\Phi_{x,t})_{x \in (-1,1)^d,t \in \mathbb{N}}$, where $\forall x \in [-1,1]^d, \forall t \in \mathbb{N}, \Phi_{x,t} : \big([-1,1]^d\times\mathcal{B}_1^d\big)^t \to \mathbb{R}$, such that $$\forall \varepsilon >0, \forall \delta \in (0,1), \exists T \in \mathbb{N}, \forall t \ge T, \forall F \in \mathcal{F}, \\ \mathbb{P}\bigg[ \sup_{x \in [0,1]^d}\Big|\Phi_{x,t}(X_1,Y_1^F,\dots,X_t,Y_t^F) - \big(F(x)-F(0)\big)\Big| \ge \varepsilon \bigg] \le \delta.$$ Notice that estimating $F-F(0)$ instead of $F$ is just a trick to get rid of the constant we never see.

It is interesting to start with the case $d = 1$, where I'm quite confident that the problem is solvable via Monte Carlo integration in the following way. Define $$\Phi_{x,t}(x_1,y_1,\dots,x_t,y_t) = \frac{1}{t} \sum_{s=1}^t \big(y_t \cdot \mathbb{I}\{\min(x,0)\le x_t \le \max(x,0)\} \cdot \operatorname{sgn}(x)\big)$$ and select the random variables $X_1,X_2, \dots$ just as family of independent $[-1,1]$-valued uniform random variables.

But what about the general case? What are sensible estimators and strategies to solve the problem? And what are the best achievable decaying rate depending on $t, \varepsilon$ and on the dimension $d$ for the quantity $$\mathbb{P}\bigg[ \sup_{x \in [0,1]^d}\Big|\Phi_{x,t}(X_1,Y_1^F,\dots,X_t,Y_t^F) - \big(F(x)-F(0)\big)\Big| \ge \varepsilon \bigg] ?$$