The aim of this paper is to carry out convergence analysis and algorithm implementation of a novel sample-wise backpropagation method for training a class of stochastic neural networks (SNNs). The preliminary discussion on such an SNN framework was first introduced in [Archibald et al., Discrete Contin. Dyn. Syst. Ser. S, 15 (2022), pp. 2807–2835]. The structure of the SNN is formulated as a discretization of a stochastic differential equation (SDE). A stochastic optimal control framework is introduced to model the training procedure, and a sample-wise approximation scheme for the adjoint backward SDE is applied to improve the efficiency of the stochastic optimal control solver, which is equivalent to the backpropagation for training the SNN. The convergence analysis is derived by introducing a novel joint conditional expectation for the gradient process. Under the convexity assumption, our result indicates that the number of SNN training steps should be proportional to the square of the number of layers in the convex optimization case. In the implementation of the sample-based SNN algorithm with the benchmark MNIST dataset, we adopt the convolution neural network (CNN) architecture and demonstrate that our sample-based SNN algorithm is more robust than the conventional CNN.

We study quasi-Monte Carlo (QMC) integration over the multi-dimensional unit cube in several weighted function spaces with different smoothness classes. We consider approximating the integral of a function by the median of several integral estimates under independent and random choices of the underlying QMC point sets (either linearly scrambled digital nets or infinite-precision polynomial lattice point sets). Even though our approach does not require any information on the smoothness and weights of a target function space as an input, we can prove a probabilistic upper bound on the worst-case error for the respective weighted function space, where the failure probability converges to 0 exponentially fast as the number of estimates increases. Our obtained rates of convergence are nearly optimal for function spaces with finite smoothness, and we can attain a dimension-independent super-polynomial convergence for a class of infinitely differentiable functions. This implies that our median-based QMC rule is universal in the sense that it does not need to be adjusted to the smoothness and the weights of the function spaces and yet exhibits the nearly optimal rate of convergence. Numerical experiments support our theoretical results.

.Finding stationary states of Landau free energy functionals has to solve a nonconvex infinite-dimensional optimization problem. In this paper, we develop a Bregman distance based optimization method for minimizing a class of Landau energy functionals and focus on its convergence analysis in the function space. We first analyze the regularity of the stationary states and show the weakly sequential convergence results of the proposed method. Furthermore, under the Łojasiewicz–Simon property, we prove a strongly sequential convergent property and establish the local convergence rate in an appropriate Hilbert space. In particular, we analyze the Łojasiewicz exponent of three well-known Landau models, the Landau–Brazovskii, Lifshitz–Petrich, and Ohta–Kawasaki free energy functionals. Finally, numerical results support our theoretical analysis.KeywordsLandau free energy functionalBregman iterationconvergence rateŁojasiewicz exponentMSC codes65K0590C26

We propose two novel unbiased estimators of the integral ´[0,1] s f (u)du for a function f , which depend on a smoothness parameter r ∈ N. The first estimator integrates exactly the polynomials of degrees p < r and achieves the optimal error n −1/2−r/s (where n is the number of evaluations of f ) when f is r times continuously differentiable. The second estimator is also optimal in term of convergence rate and has the advantage to be computationally cheaper, but it is restricted to functions that vanish on the boundary of [0, 1] s . The construction of the two estimators relies on a combination of cubic stratification and control variates based on numerical derivatives. We provide numerical evidence that they show good performance even for moderate values of n.

The logarithmic Schrödinger equation (LogSE) has a logarithmic nonlinearity f (u) = u ln |u| 2 that is not differentiable at u = 0. Compared with its counterpart with a regular nonlinear term, it possesses richer and unusual dynamics, though the low regularity of the nonlinearity brings about significant challenges in both analysis and computation. Among very limited numerical studies, the semi-implicit regularized method via regularising f (u) as u ε ln(ε+|u ε |) 2 to overcome the blowup of ln |u| 2 at u = 0 has been investigated recently in literature. With the understanding of f (0) = 0, we analyze the non-regularized first-order Implicit-Explicit (IMEX) scheme for the LogSE. We introduce some new tools for the error analysis that include the characterization of the Hölder continuity of the logarithmic term, and a nonlinear Grönwall's inequality. We provide ample numerical results to demonstrate the expected convergence. We position this work as the first one to study the direct linearized scheme for the LogSE as far as we can tell.

.We provide the convergence analysis for a \(\operatorname{sinc}\)-Galerkin method to solve the fractional Dirichlet problem. This can be understood as a follow-up of [H. Antil, P. Dondl, and L. Striet, SIAM J. Sci. Comput., 43 (2021), pp. A2897–A2922], where the authors presented a \(\operatorname{sinc}\)-function based method to solve fractional PDEs. While the original method was formulated as a collocation method, we show that the same method can be interpreted as a nonconforming Galerkin method, giving access to abstract error estimates. Optimal order of convergence is shown without any unrealistic regularity assumptions on the solution.Keywordsfractional LaplacianDirichlet problemsinc-Galerkin methodconvergence rateMSC codes35R1165N3065N35

The perfectly matched layer (PML) is a very popular tool in the truncation of wave scattering in unbounded domains. In [S. N. Chandler-Wilde and P. Monk, Appl. Numer. Math., 59 (2009), pp. 2131–2154], the author proposed a conjecture that for scattering problems with rough surfaces, the PML converges exponentially with respect to the PML parameter in any compact subset. In the author’s previous paper [R. Zhang, SIAM J. Numer. Math., 60 (2022), pp. 804–823], this result has been proved for periodic surfaces in two-dimensional spaces, when the wave number is not a half-integer. In this paper, we prove that the method has a high-order convergence rate in the three-dimensional biperiodic surface scattering problems. We extend the two-dimensional results and prove that the exponential convergence still holds when the wave number is smaller than 0.5. For larger wave numbers, although exponential convergence is no longer proved, we are able to prove a higher-order convergence for the PML method.

We present a novel framework based on semi-bounded spatial operators for analyzing and discretizing initial boundary value problems on moving and deforming domains. This development extends an existing framework for well-posed problems and energy stable discretizations from stationary domains to the general case including arbitrary mesh motion. In particular, we show that an energy estimate derived in the physical coordinate system is equivalent to a semi-bounded property with respect to a stationary reference domain. The continuous analysis leading up to this result is based on a skew-symmetric splitting of the material time derivative, and thus relies on the property of integration-by-parts. Following this, a mimetic energy stable arbitrary Lagrangian-Eulerian framework for semi-discretization is formulated, based on approximating the material time derivative in a way consistent with discrete summation-by-parts. Thanks to the semi-bounded property, a method-of-lines approach using standard explicit or implicit time integration schemes can be applied to march the system forward in time. The same type of stability arguments applies as for the corresponding stationary domain problem, without regards to additional properties such as discrete geometric conservation. As an additional bonus we demonstrate that discrete geometric conservation, in the sense of exact free-stream preservation, can still be achieved in an automatic way with the new framework. However, we stress that this is not necessary for stability.

The discrete gradient structure and the positive definiteness of discrete fractional integrals or derivatives are fundamental to the numerical stability in long-time simulation of nonlinear integro-differential models. We build up a discrete gradient structure for a class of secondorder variable-step approximations of fractional Riemann-Liouville integral and fractional Caputo derivative. Then certain variational energy dissipation laws at discrete levels of the corresponding variable-step Crank-Nicolson type methods are established for time-fractional Allen-Cahn and time-fractional Klein-Gordon type models. They are shown to be asymptotically compatible with the associated energy laws of the classical Allen-Cahn and Klein-Gordon equations in the associated fractional order limits. Numerical examples together with an adaptive time-stepping procedure are provided to demonstrate the effectiveness of our second-order methods.

The purpose of this paper is to analyze a mixed method for the linear elasticity eigenvalue problem, which approximates numerically the stress, displacement, and rotation, by piecewise (<i>k</i> + 1), <i>k</i>, and (<i>k</i> + 1) th degree polynomial functions (<i>k</i> ≥ 1), respectively. The numerical eigenfunction of stress is symmetric. By the discrete <i>H</i><sup style="">1</sup>-stability of numerical displacement, we prove an <i>O </i>(<i>h^k</i>⁺²) approximation to the <i>L</i><sup style="">2</sup>-orthogonal projection of the eigenspace of exact displacement for the eigenvalue problem with a proper regularity assumption. Thus via postprocessing, we obtain a better approximation to the eigenspace of exact displacement for the eigenproblem than conventional methods. We also prove that numerical approximation to the eigenfunction of stress is locking free with respect to the Poisson ratio. We introduce a hybridization to reduce the mixed method to a condensed eigenproblem and prove an <i>O </i>(<i>h</i>²) initial approximation (independent of the inverse of the elasticity operator) of the eigenvalue for the nonlinear eigenproblem by using the discrete <i>H</i>¹-stability of numerical displacement, while only an <i>O </i>(<i>h</i>) approximation can be obtained if we use the traditional inf-sup condition. Finally, we report some numerical experiments. © 2023 The Author(s).