Iterative tensor network transformations for element-wise evaluation of elementary and filtering functions
A recent preprint on arXiv (2608.17135) presents iterative tensor network transformations (ITNTs), a comprehensive algorithmic framework designed for the element-wise assessment of both elementary and nonlinear filtering functions applied to data represented as tensor trains (TTs). This technique functions entirely within the compressed domain, allowing for efficient calculations on exponentially large datasets while maintaining manageable computational costs. The framework is illustrated through two significant applications: assessing highly nonlinear functions within a 3D reactive flow field for precise reaction rate calculations and region filtering, as well as identifying extrema in intricate optimization challenges, including Max-SAT problems across configurations as vast as 2^70. This method overcomes prior challenges associated with executing nonlinear operations on tensor networks, potentially expanding their applicability in large-scale data processing.
Key facts
- The paper introduces iterative tensor network transformations (ITNTs).
- ITNTs enable element-wise evaluation of elementary and nonlinear filtering functions.
- Data is encoded as tensor trains (TTs), a class of tensor networks.
- The method operates entirely in the compressed domain.
- It maintains controlled computational cost on exponentially large datasets.
- Application area I: evaluating nonlinear functions on a 3D reactive flow field.
- Application area II: finding extrema in optimization problems.
- The framework solves Max-SAT instances on spaces up to 2^70 configurations.
- The paper is available on arXiv with identifier 2608.17135.
Entities
—