Abstract:We extend the kernel-differentiation method (or likelihood-ratio method) for the linear response of random dynamical systems, after revisit its derivation in a microscopic view via transfer operators. First, for the linear response of physical measures, we extend the method to an ergodic version, which is sampled by one infinitely-long sample path; moreover, the magnitude of the integrand is bounded by a finite decorrelation step number. Second, when the noise and perturbation are along a given foliation, we show that the method is still valid for both finite and infinite time. We numerically demonstrate the ergodic version on a tent map, whose linear response of the physical measure does not exist; but we can add some noise and compute an approximate linear response. We also demonstrate on a chaotic neural network with 51 layers $\times$ 9 neurons, whose perturbation is along a given foliation; we show that adding foliated noise incurs a smaller error than noise in all directions. We give a rough error analysis of the method and show that it can be expensive for small-noise. We propose a potential future program unifying the three popular linear response methods.

What problem does this paper attempt to address?

This paper attempts to solve the problem of linear response of physical measures in stochastic dynamical systems. Specifically, it extends the kernel differentiation method (or likelihood ratio method) to calculate the linear response of stochastic dynamical systems and re - derives its formula from a microscopic perspective through the transfer operator. The main contributions of the paper include: 1. **Introducing the ergodic version of the kernel differentiation method**: - For the linear response of physical measures, the author proposes an ergodic version of the method, which can be sampled through an infinitely long sample path. - The size of the integral term of this method is limited by a finite number of decorrelation steps. 2. **Handling noise and perturbations along a given foliation**: - When the noise and perturbations are along a given foliation, it is proved that this method is still valid in both finite and infinite time cases. - Numerical experiments show that in some cases, adding low - dimensional noise along the foliation can lead to smaller errors. 3. **Application examples**: - The application of the ergodic version of the method is demonstrated on the tent map. Even though the physical measure of this system has no linear response, an approximate linear response can be calculated by adding noise. - Numerical verification is carried out on a chaotic neural network with 51 layers × 9 neurons. The results show that adding noise along the foliation produces smaller errors than adding noise in all directions. 4. **Theoretical analysis and future work**: - A rough error analysis of this method is given, pointing out that this method may be costly in the case of small noise. - A potential future research plan is proposed, aiming to unify three popular linear response methods. ### Formula summary - **Ergodic kernel differentiation formula** (Theorem 1): \[ \delta \int \Phi(x) h(x) dx = \lim_{W \to \infty} \lim_{L \to \infty} -\frac{1}{LW} \sum_{n = 1}^{L} \sum_{l = 1}^{W} (\Phi(X_{n + l}) - \Phi_{\text{avg}}) \frac{\delta f_\gamma X_l \cdot dp}{p(Y_{1 + l})} \] where $\Phi_{\text{avg}}=\int \Phi h$. - **Centered foliation kernel differentiation formula** (Theorem 2): \[ \delta \int \Phi(x_T) d\mu_T(x_T)=E\left[ (\Phi_T(x_T)-\Phi_{\text{avg},T}) \sum_{m = 1}^T \left( \frac{\delta \tilde{f}_m(z_m) \cdot d p_m(z_m, x_m)}{p_m(z_m, x_m)} \right) \right] \] These formulas provide new tools and methods for understanding and calculating the linear response of stochastic dynamical systems.

Ergodic and foliated kernel-differentiation method for linear responses of random systems

Equivalent Nonlinear System Method for Stochastically Excited and Dissipated Integrable Hamiltonian Systems

Stochastic averaging of n-dimensional non-linear dynamical systems subject to non-Gaussian wide-band random excitations

Determining Response Probability Density of Nonlinear Stochastic Systems by Modifying Analytical Result of Degenerated Systems.

A systematic path to non-Markovian dynamics II: Probabilistic response of nonlinear multidimensional systems to Gaussian colored noise excitation

A hypocoercivity related ergodicity method with rate of convergence for singularly distorted degenerate Kolmogorov equations and applications

Linear response in large deviations theory: A method to compute non-equilibrium distributions

Ergodicity of Langevin Dynamics and its Discretizations for Non-smooth Potentials

Some properties on the reversibility and the linear response theory of Langevin dynamics

Effective quenched linear response for random dynamical systems

Statistical stability and linear response for random hyperbolic dynamics

Topics in the Analysis and Computation of Stochastic Differential Equations

Non-local linear response in anomalous transport

Computing Nonequilibrium Responses with Score-shifted Stochastic Differential Equations

An ergodic averaging method to differentiate covariant Lyapunov vectors

A variational approach to nonlinear and interacting diffusions

Rearranged Stochastic Heat Equation: Ergodicity and Related Gradient Descent on ${\mathcal P}({\mathbb R})$

Function approximation by neural nets in the mean-field regime: Entropic regularization and controlled McKean-Vlasov dynamics

Non-Markovian dynamics: the governing equations of probability density functions for nonlinear dynamical systems with combined fractional Gaussian and Gaussian white noises

Ergodicity and mixing for stochastic partial differential equations

Data-driven approach for extracting the most probable exit trajectory of stochastic dynamical systems with non-Gaussian Lévy noise