arXiv:2610.07144
Decoupling Spectral Gaps from Entanglement Complexity in Quantum Annealing
Roopayan Ghosh, Sougato Bose
Abstract
Exponentially small spectral gaps are the central obstacle to adiabatic quantum annealing, yet whether such bottlenecks demand correspondingly complex entanglement remains unclear. We show that gap scaling and entanglement complexity need not be linked. Conventional first-order transitions combine exponentially small gaps with bounded $O(1)$ bipartite Schmidt rank, even though the critical states retain global multipartite correlations, while continuous transitions combine polynomially closing gaps with logarithmically growing entanglement. We show that each pairing can also be inverted: in a known first-order transition in a staggered-field antiferromagnetic Ising chain the gap closes polynomially despite bounded Schmidt rank, while we construct a mirror-twin manifold model whose gap closes exponentially even as its entanglement entropy grows logarithmically with system size. We derive bounds connecting Schmidt rank to Hamming-space geometry, and identify manifold density and connectivity as the ingredients that determine which combination occurs. These results show that entanglement complexity alone cannot characterize annealing difficulty, and point toward a broader search for quantum resources that capture the geometry of state evolution in Hilbert space.
AI Summary
The paper asks whether small spectral gaps in quantum annealing require complex ground-state entanglement. It compares bipartite entanglement and gap scaling in collective-spin models, optimization-inspired avoided crossings, a frustrated Ising chain, and a constructed three-local Hamiltonian. Its organizing distinction is between connectivity of competing low-energy sectors, which controls tunnelling, and independent Schmidt branches across a cut. For first-order $p$-spin transitions, numerical calculations find exponentially small gaps with bounded entropy, reported as $S_{\max}\to\ln2$ for $p=3$ and $\ln3$ for $p=4$. The continuous $p=2$ case instead has $\Delta\sim L^{-1/3}$ and logarithmic entropy. A known odd-length antiferromagnetic Ising chain provides polynomial gap closing with bounded entanglement: a mobile frustrated bond gives $\Delta\sim L^{-2}$ and at most four Schmidt branches in the low-energy description. The new “mirror-twin” construction contains two manifolds of $n=L/2$ configurations. Neighbours within each manifold hybridize at second order, while inter-manifold coupling requires $n+1$ spin flips. The resulting effective states have Schmidt rank $\Theta(L)$ and entropy $\Theta(\log L)$. Exact diagonalization for $n=3,\ldots,6$ gives an interior-gap fit $\Delta\sim e^{-1.017L}$. Support-size and coding-theoretic bounds supplement this geometric account. The construction’s exponential scaling is supported by perturbative reasoning and small-system numerics, without a sharp tunnelling prefactor. Its problem ground space is degenerate, so the quoted gap excludes the endpoint closing. Whether a geometric resource provides information beyond gaps and entanglement remains open.
Central claim may not hold: The abstract's exact O(1) Schmidt-rank claim is not established by bounded entropy or an approximate superposition of a few macroscopic branches. The mirror-twin exponential-gap asymptotic is supported by only four small-system sizes and a perturbative argument without control of the size-dependent path sum. The mirror-twin logarithmic-entropy argument omits entangling virtual excitations across an extensive number of A–B interactions, so its projected-manifold calculation does not establish logarithmic entropy for the full ground state. The antiferromagnetic chain's exact rank-four bound and entropy limit ln(4) are derived only after projection onto the single-defect manifold and need not hold at finite transverse field. The general tunnelling argument incorrectly treats shortest-path counting as polynomial even though configurations at Hamming distance d have d! shortest flip orderings.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 18 pages 5 figures