Exact Neural-Network Representations of the Motzkin States
AuthorsRunde Zha, Yuntian Gu, Chaohui Fan, Jia-lin Chen, Hai-Jun Liao, Tao Xiang
Resources
This work shows how several neural-network architectures can exactly encode exotic Motzkin quantum states with entanglement patterns that conventional tensor networks struggle to represent.
Key results
The colorless half-chain entropy scales as 1/2 ln N plus O(1).
The colorful transformer uses a 3-layer attention-plus-MLP architecture.
The colorful transformer embeds each input configuration into 9 dimensions.
What the paper found
“Exact Neural-Network Representations of the Motzkin States,” by researchers affiliated with the Chinese Academy of Sciences, Peking University, and ByteDance Seed, shows that neural quantum states can exactly encode quantum wavefunctions with entanglement beyond the matrix-product-state area law. The authors construct training-free, fixed-weight representations for both colorless and colorful Motzkin chains using four architectures: recurrent neural networks, feedforward networks, convolutional networks, and transformers. A causal prefix-sum module computes Motzkin heights, while position-selective ReLU gates enforce nonnegativity and endpoint closure. For colorful states, a causal stack or attention pointer implements the last-in-first-out color-matching rule. The colorless state has half-chain entropy proportional to 1/2 ln N, whereas colorful states exhibit supercritical square-root-of-N entanglement. Parameter complexity remains polynomial despite this anomalous entanglement: colorless constructions range from O(1) to O(N^2), and colorful ones from O(N) to O(N^3). The colorful transformer uses a 3-layer attention-plus-MLP design with 9-dimensional embeddings. Rather than learning amplitudes variationally, these networks recognize legal Motzkin configurations algorithmically and assign them uniform amplitudes, offering exact benchmarks and design principles for neural representations of constrained quantum states.
Original abstract
Motzkin spin chains are paradigmatic frustration-free one-dimensional quantum systems whose ground states feature exactly solvable combinatorial structures and exotic, area-law-violating entanglement scaling. Specifically, colorless Motzkin states exhibit critical logarithmic entanglement divergence \(\log N\) with system size \(N\), while their colorful counterparts host supercritical sublinear \(\sqrt{N}\) entanglement growth. Such unconventional entanglement behaviors place these states well beyond the expressive capability of standard matrix product states, which are fundamentally constrained by the entanglement area law. Here, we systematically construct exact, training-free neural-network representations for both colorless and colorful Motzkin states across four mainstream architectures, including recurrent, feedforward, convolutional, and transformer networks. Our core design leverages a causal prefix-sum module, implementable via recurrent updates, feedforward mappings, or masked attention layers, combined with position-selective rectified linear gates that enforce the Motzkin height constraints. For the colorful states, we further introduce a dedicated causal stack module that explicitly encodes the last-in-first-out color-matching rule. Our results demonstrate that neural architectures can accurately capture highly non-trivial entanglement features inaccessible to conventional tensor networks, providing prototypic examples for benchmarking and a constructive design framework for future neural-network quantum state developments targeting strongly entangled quantum systems.
Read the original paperMore in Neural Networks
Browse all 22 papers →End-to-End Hard-Label Cryptanalytic Model Extraction Using Efficient Sign Recovery
Akira Ito, Takayuki Miura, Yosuke Todo
A new query-efficient technique makes it possible to steal the parameters of small black-box neural networks using only their predicted labels.
Retrieving Individual Stems from Music Mixtures with Slot Embeddings
David Braun, Junyi Fan, Pranay Manocha, Donald S. Williamson, Adam Finkelstein
Stembed lets music producers search for individual instrument sounds hidden inside a full song by representing the mixture as multiple searchable stem-like embeddings.
The Linear Representation Hypothesis Needs a Group Action
Louie Hong Yao, Yuhao Li, Shengchao Liu
This paper argues that claims about linear representations only become meaningful once we specify which transformations leave a representation essentially unchanged.