arXiv:2609.25458
Random Gaussian Augmented Matrix Product States
Stavya Puri, Poetri Sonya Tarabunga, Piotr Sierant
Abstract
Strong entanglement limits the reach of tensor-network descriptions of quantum many-body systems. We study Gaussian-augmented matrix product states (GAMPS), formed by applying a free-fermionic (Gaussian) unitary to a matrix product state. To explore the capabilities of this family, we consider random GAMPS}, a minimally structured ensemble obtained by independently sampling random matrix product states and Gaussian unitaries. Exploiting the algebraic structure of Gaussian unitaries, we develop a replica tensor-network method that computes averages over the random GAMPS ensemble for hundreds of qubits. Random GAMPS exhibit nearly maximal volume-law entanglement already at modest bond dimension, while other quantum resource quantifiers rapidly approach Haar-random state values as the bond dimension increases. The entanglement scaling suggests an exponential reduction in the bond dimension needed to approach the volume-law entropy compared with unaugmented matrix product states. We find a state two-design behavior of random GAMPS at fixed accuracy with a bond dimension independent of system size. Our results reveal how Gaussian augmentation expands the expressive power of tensor networks.
AI Summary
How much can a fermionic Gaussian layer expand the expressive power of a matrix product state? This paper studies random Gaussian-augmented MPS (GAMPS), $|\Psi\rangle=U_G|\Phi_\chi\rangle$, with independently sampled random MPS and Gaussian unitaries, including parity reflections. A two-replica tensor-network method combines local Haar integration for the MPS with the Gaussian commutant. Pairing tensors have MPO bond dimension $n+1$, while the averaged two-copy MPS has bond dimension two. Their contraction enables ensemble calculations for hundreds of qubits and polynomial-cost evaluation of the two-copy Gaussian twirl for arbitrary MPS inputs. Numerical thermodynamic extrapolations indicate relative inverse-participation-ratio deviations and trace distance from the Haar second moment scaling as $\chi^{-2}$; the relative second-order frame-potential deviation scales as $\chi^{-4}$. These support approximate state two-design behavior at fixed accuracy with bond dimension independent of system size. An analytical transfer-matrix calculation gives the fermionic-antiflatness deficit per site as $5/(N\chi)+5/(2\chi^2)$ plus subleading terms. The annealed half-chain Rényi-2 entropy is extensive and approaches the Haar value. Reaching a relative entropy deficit below $10^{-3}$ numerically requires a crossover bond dimension $\chi_c\sim2^{0.04N}$, versus approximately $2^{0.5N}$ for unaugmented MPS. Parity-resolved calculations show analogous entanglement and anticoncentration trends. The design and entanglement scalings rely on numerical extrapolation rather than uniform analytical bounds. Entropy calculations primarily concern annealed averages, supplemented by numerical annealed–quenched comparisons. Higher-design properties and the practical classical simulation cost of GAMPS-based algorithms remain open.
Error found: The entropy crossover χ_c is defined by a relative deficit of at most 10^−3, but the stated formula a=a_Haar(1−χ_c/χ) instead gives unit relative deficit at χ=χ_c and reaches that tolerance only at χ=1000χ_c. The reported N^−1 relative annealed–quenched entropy difference supports equality of entropy densities but does not establish the paper's literal assertion that the unnormalized entropies coincide.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 25 pages, 6 figures