Abstract:We introduce a change of perspective on tensor network states that is defined by the computational graph of the contraction of an amplitude. The resulting class of states, which we refer to as tensor network functions, inherit the conceptual advantages of tensor network states while removing computational restrictions arising from the need to converge approximate contractions. We use tensor network functions to compute strict variational estimates of the energy on loopy graphs, analyze their expressive power for ground-states, show that we can capture aspects of volume law time evolution, and provide a mapping of general feed-forward neural nets onto efficient tensor network functions. Our work expands the realm of computable tensor networks to ones where accurate contraction methods are not available, and opens up new avenues to use tensor networks.

What problem does this paper attempt to address?

This paper aims to solve the computational limitations faced by traditional Tensor Network States (TNS) when dealing with complex quantum many - body systems. Specifically, traditional methods require high precision when performing tensor network contractions, which usually means a large auxiliary bond dimension (\(\chi\)), leading to an exponential increase in computational cost. This limitation makes tensor networks impractical when dealing with certain graph structures with high degrees of entanglement. To solve this problem, the author proposes a new perspective - Tensor Network Functions (TNFs). The core idea of TNFs is to consider the approximate contraction scheme itself as part of the tensor network state, rather than as a method that must converge to an exact solution. In this way, TNFs can still provide effective computations with smaller values of \(\chi\) and can maintain the variational property, that is, the calculated energy will not be lower than the actual ground - state energy of the system. ### Main contributions: 1. **Strict variational estimation**: TNFs can provide strict variational energy estimates on graphs with loops, even when using smaller values of \(\chi\). 2. **Efficient representational ability**: TNFs can efficiently represent neural - network quantum states, thus expanding the application range of tensor networks. 3. **Volume - law behavior**: TNFs can capture the behavior of volume - law time evolution, which is particularly important for studying the time evolution of chaotic systems. 4. **Removal of computational limitations**: TNFs eliminates the strict convergence requirement for the auxiliary bond dimension \(\chi\), making more types of graph structures available for defining tensor networks. ### Application examples: - **Quantum spin model**: Through TNFs, the author shows how to obtain strict variational energy estimates in the two - dimensional Heisenberg model. - **Fermi - Hubbard model**: In the Fermi - Hubbard model, TNFs can achieve high precision at a lower computational cost. - **Random circuits**: TNFs can effectively simulate the dynamics of random circuits and capture volume - law behavior. ### Conclusion: By introducing TNFs, the author not only expands the application range of tensor networks but also provides a new computational framework, making it more flexible and efficient when dealing with complex quantum many - body systems. This method is expected to open up new avenues in future research, especially in the cross - field of quantum computing and machine learning.

Tensor Network Computations That Capture Strict Variationality, Volume Law Behavior, and the Efficient Representation of Neural Network States