Coefficient-based regularization network with variance loss for error

Loss function is the key element of a learning algorithm. Based on the regression learning algorithm with an offset, the coefficient-based regularization network with variance loss is proposed. The variance loss is different from the usual least quare loss, hinge loss and pinball loss, it induces a kind of samples cross empirical risk. Also, our coefficient-based regularization only relies on general kernel, i.e. the kernel is required to possess continuity, boundedness and satisfy some mild differentiability condition. These two characteristics bring essential difficulties to the theoretical analysis of this learning scheme. By the hypothesis space strategy and the error decomposition technique in [L. Shi, Learning theory estimates for coefficient-based regularized regression, Appl. Comput. Harmon. Anal. 34 (2013) 252–265], a capacity-dependent error analysis is completed, satisfactory error bound and learning rates are then derived under a very mild regularity condition on the regression function. Also, we find an effective way to deal with the learning problem with samples cross empirical risk.

Convergence of a Two-Step Iterative Method for Nondifferentiable Operators in Banach Spaces

The semilocal and local convergence analyses of a two-step iterative method for nonlinear nondifferentiable operators are described in Banach spaces. The recurrence relations are derived under weaker conditions on the operator. For semilocal convergence, the domain of the parameters is obtained to ensure guaranteed convergence under suitable initial approximations. The applicability of local convergence is extended as the differentiability condition on the involved operator is avoided. The region of accessibility and a way to enlarge the convergence domain are provided. Theorems are given for the existence-uniqueness balls enclosing the unique solution. Finally, some numerical examples including nonlinear Hammerstein type integral equations are worked out to validate the theoretical results.

AN INVARIANCE PRINCIPLE FOR THE TAGGED PARTICLE PROCESS IN CONTINUUM WITH SINGULAR INTERACTION POTENTIAL

We consider the dynamics of a tagged particle in an infinite particle environment moving according to a stochastic gradient dynamics. For singular interaction potentials this tagged particle dynamics was constructed first in Ref. 7, using closures of pre-Dirichlet forms which were already proposed in Refs. 13 and 24. The environment dynamics and the coupled dynamics of the tagged particle and the environment were constructed separately. Here we continue the analysis of these processes: Proving an essential m-dissipativity result for the generator of the coupled dynamics from Ref. 7, we show that this dynamics does not only contain the environment dynamics (as one component), but is, given the latter, the only possible choice for being the coupled process. Moreover, we identify the uniform motion of the environment as the reversed motion of the tagged particle. (Since the dynamics are constructed as martingale solutions on configuration space, this is not immediate.) Furthermore, we prove ergodicity of the environment dynamics, whenever the underlying reference measure is a pure phase of the system. Finally, we show that these considerations are sufficient to apply Ref. 4 for proving an invariance principle for the tagged particle process. We remark that such an invariance principle was studied before in Ref. 13 for smooth potentials, and shown by abstract Dirichlet form methods in Ref. 24 for singular potentials. Our results apply for a general class of Ruelle measures corresponding to potentials possibly having infinite range, a non-integrable singularity at 0 and a nontrivial negative part, and fulfill merely a weak differentiability condition on ℝd\{0}.