Integration by parts formula and shift Harnack inequality for stochastic equations (Q2450246): Difference between revisions

Consider a diffusion process on a Riemannian manifold \(\mathcal{M}\) with semigroup \(P_{t}\). Of importance is the so-called Bismut formula given by \[ \nabla_{v}P_{t}f(x)=\operatorname{E}[ f(X_{t}) M_{t}| X_{0}=x ], \] where \(\nabla_{v}\) is the directional derivative along the vector \(v\) in a tangent space of \(\mathcal{M}\). Here, \(M_{t}\), which does not depend on \(f\), is given by \(\int_{0}^{t}V_{s}\,dB_{s}\), where \(B_{s}\) is a Brownian motion on a tangent space and \(V_{s}\) is defined as a multiplicative Feynmann-Kac functional determined by the Ricci curvature of \(\mathcal{M}\). It is known that from the Bismut formula one can obtain a Harnack inequality for \(|P_{t}f|^{p}(x)\). Also, one can obtain a formula for \(\nabla_{v}\log p_{t}(\cdot,y)(x)\), i.e. taking the derivative with respect the first argument of \(p_{t}\), where \(p_{t}\) is the density of \(P_{t}\). In this paper, an integration by parts formula given by \[ P_{t}(\nabla_{v}f)(x)=\operatorname{E}[f(X_{t})N_{t}|X_{0}=x] \] is studied, where \(N_{t}\) turns out to be more elaborate than \(M_{t}\) in the previous display. The author proposes a different coupling argument to derive the previous formula. In this coupling argument, two processes start at the same point, but the difference of them reaches a fixed value at a given time. Using the previous formula, the author derives an inequality, called the shift Harnak inequality, for \(|P_{t}f|^{p}(x)\), which is different than the one it is obtained using the Bismut formula. A formula for \(\nabla_{v}\log p_{t}(x,\cdot)(y)\), i.e., taking the derivative with respect the second argument of \(p_{t}\), is also obtained. The author also illustrates the ideas with examples in finite and infinite dimensional stochastic models, such as matrix-valued and functional-valued stochastic differential equations, as well as partial stochastic differential equations.