Source-linked AI summary
Machine learning meets quantum physics
Sankar Das Sarma, Dong-Ling Deng, Lu-Ming Duan
TL;DR
Quantum many-body states are exponentially costly to describe, while existing tensor-network methods become inefficient for highly entangled states. The paper examines neural-network representations and their applications, showing that RBMs can efficiently represent certain highly entangled states, while deeper networks broaden the representable class. It also identifies unresolved questions for quantum-enhanced machine learning.
Problem
Fully describing arbitrary quantum many-body states requires exponentially much information, and conventional tensor networks are inefficient for highly entangled states.
Method
The paper reviews tensor-network and neural-network representations of quantum states, including RBM-based variational learning and related applications.
Results
Neural-network representations can efficiently describe certain highly entangled quantum states, and deep Boltzmann machines can represent almost all physical quantum states with at most polynomial parameter scaling.
Takeaways & Limitations
Neural-network approaches provide a route to solving many-body problems involving large entanglement that are challenging or unsolvable with conventional methods.
Takeaways & Limitations
Some physically interesting states lack efficient RBM descriptions, and quantum-enhanced machine learning has no unified theory.
Abstract
from arXiv · showhide
The marriage of machine learning and quantum physics may give birth to a new research frontier that could transform both.
TWO REPRESENTATIONS
Quantum many-body states require exponentially large descriptions, motivating representations that compress them efficiently. Tensor networks work well for area-law states, while neural-network representations can efficiently capture highly entangled states and support related quantum algorithms.
- The representation problem: An arbitrary N-qubit quantum state requires 2^N complex numbers for a complete description.Each qubit has two configurations, producing 2^N possible configurations.
- Tensor-network representation: Tensor networks efficiently represent many physical states whose information scales polynomially with system size.Their efficiency is associated with the entanglement area law.
- Neural-network representation: RBMs represent certain topological states with a number of parameters that scales linearly with the number of qubits.For the toric-code state, each hidden neuron connects to four nearby visible neurons, giving roughly four parameters per qubit.
- Scope and open questions: An additional hidden layer broadens applicability: deep Boltzmann machines can efficiently represent almost all physical quantum states with at most polynomial parameter scaling.The paper also notes that some physically interesting states lack efficient RBM descriptions.
- Neural-network representation: Without short-range connectivity, RBMs can represent volume-law states with only linearly scaling parameter counts.A construction with each visible neuron connected to at most three hidden neurons describes maximally entangled states efficiently.
- Entanglement and efficiency: Highly entangled states can make conventional tensor-network descriptions exponentially expensive.Volume-law states require exponentially many tensor-network parameters.
- Applications: RBM-based methods captured ground states and time evolution, reconstructed highly entangled states, and addressed long-range-interaction systems.These applications include variational learning and quantum state tomography.
- Scope and open questions: Quantum-enhanced machine learning still lacks a unified theory, including criteria for when quantum computers significantly expedite learning tasks.The paper identifies efficient analysis of large quantum data sets as another open question.