Abstract:We investigate the sharp error estimate of the density under the relative entropy (or Kullback-Leibler divergence) for the traditional Euler-Maruyama discretization of stochastic differential equations (SDEs) with multiplicative noise. The foundation of the proof is the estimates of the derivatives for the logarithmic numerical density. The key technique is to adopt the Malliavin calculus to get the expressions of the derivatives of the logarithmic Green's function and to obtain an estimate for the inverse Malliavin matrix. The estimate of relative entropy then naturally gives sharp error bounds under total variation distance and Wasserstein distances. Compared to the usual weak error estimate for SDEs, such estimate can give an error bound for a family of test functions instead of one test function.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is about the error estimation of numerical solutions to stochastic differential equations (SDEs), especially when these equations have multiplicative noise. Specifically, the authors focus on the accurate error estimation in relative entropy (Kullback - Leibler divergence, abbreviated as KL divergence) of the density function when using the traditional Euler - Maruyama discretization method to numerically simulate such SDEs. The main contribution of the paper is that by introducing Malliavin calculus techniques to estimate the derivative of the log - density, more accurate error bounds are obtained than existing methods. These bounds are applicable not only to the total variation distance but also to the Wasserstein distance.
### Background and Motivation of the Paper
- **Research Background**: Stochastic differential equations (SDEs) are widely used in many fields, such as open systems in physics, stock price modeling of financial derivatives, and the training process of neural networks. Numerical discretization and simulation of SDEs are very important in practical applications, but existing error estimation methods are usually only applicable to smooth test functions and are not accurate enough for statistics of non - smooth function families.
- **Research Motivation**: Traditional weak error estimation methods rely on fixed test functions, which have limitations when dealing with non - smooth function families. For example, the Wasserstein - 1 distance needs to consider a series of Lipschitz continuous test functions. Therefore, a new method is needed to estimate the error between the numerical density and the density of time - continuous SDEs, especially under some metrics with high enough intensity.
### Main Research Contents
- **Problem Description**: The paper studies the accurate error estimation in relative entropy of the density function of the Euler - Maruyama method for SDEs with multiplicative noise. Relative entropy is a measure widely used in statistics and data science and can provide the difference between two probability distributions.
- **Key Technologies**: The authors adopt Malliavin calculus techniques to derive the derivative expression of the log - Green function and estimate the inverse Malliavin matrix. These techniques enable them to obtain the upper - bound estimate of the log - density gradient, which is a key step in the analysis.
- **Main Results**:
- The \(O(h^2)\) error bound in relative entropy of the Euler - Maruyama method is established.
- Through Pinsker's inequality and the transportation inequality, the \(O(h)\) error bounds in total variation distance and Wasserstein distance are further obtained.
### Paper Structure
- **Section 2**: Introduces the basic concepts and notations of relative entropy and Malliavin calculus.
- **Section 3**: Presents the main result, that is, the error bound of relative entropy, and gives the key estimation assumptions for the proof.
- **Section 4**: Discusses in detail the estimation of the numerical density, especially how to use Malliavin calculus to derive the gradient expression of the log - Green function and cancel the \(x\) and \(y\) derivatives in the main terms through the estimation of the inverse Malliavin matrix.
### Conclusion
By introducing Malliavin calculus techniques, the paper successfully solves the problem of accurate error estimation in relative entropy of numerical solutions to SDEs with multiplicative noise, providing a more accurate theoretical basis for numerical simulation. This result not only improves the existing error estimation in Wasserstein distance but also provides new ideas and tools for research in other related fields.