Convergence and regularity of probability laws by using an interpolation method (Q2412667)

The authors discuss the problem of the existence of a (regular) density of probability measures. After having established a general criterion, they apply this to several settings: (i) path dependent SDEs driven by a Brownian motion and with Dini (log-Hölder)-continuous coefficients; (ii) the stochastic heat equation driven by white noise and with Dini-continuous coefficients; (iii) piecewise deterministic Markov processes coming from a jump-type SDE having smooth coefficients satisfying certain growth and non-degeneracy conditions. The main result of the paper concerns the existence of regular densities. Roughly speaking, if a probability measure \(\mu\) can be approximated by measures \(\mu_n\) having densities \(\phi_n(x)\,dx\), criteria are given such that \(\mu\) inherits a (regular) density: for example, if \(\phi_n \in C^{q+2m}(\mathbb R^d)\) such that \[ d_k(\mu,\mu_n) \|\phi_n\|^\alpha_{q+2m,2m,p}\leq C,\quad \alpha > (q+2k+d/p_*)/2m \] then \(\mu\) has a density \(\phi(x)\,dx\) with \(\phi\in W^{q,p}\). Here, \(p\in (1,\infty)\), \(p_*\) is the conjugate index of \(p\), \(m,q,k\in \mathbb N\), \(W^{q,p}\) is the standard Sobolev space over \(\mathbb R^d\), \(\|\phi\|_{k,m,p}\) is a classical \(L^p\)-Sobolev norm of order \(k\in\mathbb N\) with a polynomial weight function \((1+|x|)^m\) and, finally, \(d_k\) is a distance between measures defined by \[ d_k(\mu,\mu_n) = \sup\left\{\left|\int f\,d\mu - \int f\,d\nu\right| : \sum_{0\leq |\alpha|\leq k}\|\partial^\alpha f\|_\infty\leq 1\right\} \] (so, \(d_0\) is the TV-distance and \(d_1\) is the Fortet-Mourier distance). The proof of the above result relies on a more general result cast using Orlicz and Sobolev-Orlicz spaces. Let \(\mathsf{e}\) be a suitably regular Young function (defining the Orlicz space \(L^{\mathsf{e}}\)) and denote by \(\beta(t) = \beta_{\mathsf{e}}(t) = t/\mathsf{e}^{-1}(t)\). Then \[ \pi_{q,k,m,\mathsf{e}}(\mu,(\mu_n)_n) = \sum_{n=0}^\infty 2^{n(q+k)}\beta(2^{nd})d_k(\mu,\mu_n) + \sum_{n=0}^\infty 2^{-2nm}\|\phi_n\|_{2m+q,2m,\mathsf{e}} \] (the latter is the Sobolev-Orlicz norm) and \[ \rho_{q,k,m,\mathsf{e}}(\mu) := \inf\pi_{q,k,m,\mathsf{e}}(\mu,(\mu_n)_n) \] with the inf extending over all sequences of absolutely continuous (w.r.t.\ Lebesgue measure) measures. \[ \mathcal S_{q,k,m,\mathsf{e}} := \left\{\mu : \rho_{q,k,m,\mathsf{e}}(\mu)<\infty\right\} \] Then, the abstract main result says that \(\mathcal S_{0,k,m,\mathsf{e}}\subset L^{\mathsf{e}}\) and the density \(\phi\) of \(\mu\) satisfies the inequality \(\|\phi\|_{L^{\mathsf{e}}}\leq C\rho_{0,k,m,\mathsf{e}}(\mu)\) with a universal constant \(C\). There are also higher-order analogues of this result. The key estimate, essentially an interpolation result, is the remark that one has \[ \|f\|_{q,\mathsf{e}}\leq C\rho_{q,k,m,\mathsf{e}}(\mu) \] for all measures of the form \(\mu(dx) = f(x)\,dx\) and \(f\in C^{2m+q}(\mathbb R^d)\) and with a universal constant \(C\).