arXiv:2610.03989
Certification failure in variational-principle observables
Ganga Singh Manchanda
Abstract
Algorithms which prepare states via the variational principle assume that closeness in energy implies closeness in state; we formalise how this assumption can fail by showing that a ground-state energy estimate converged to within $ε$ of the true value is compatible with observable expectation values in error by up to $\sqrt{2εχ_O}$ at leading order, where $χ_O$ is the static susceptibility of the observable. As a result, the observable error can be orders of magnitude larger than the energy error, a mechanism we call non-linear error amplification, and because a variational algorithm cannot guarantee the correctness of these properties from energy alone, energy does not necessarily certify an observable. In light of our findings, we examine three algorithms: the variational quantum eigensolver on the transverse-field Ising chain, the density-matrix renormalisation group on a pair of weakly coupled hydrogen chains, and Krylov quantum diagonalisation on lithium fluoride. In each case we demonstrate that a well-converged energy can leave observables uncertified.
AI Summary
The paper asks what an energy-converged variational state guarantees about observables other than energy. It defines the certification radius $\delta_O(\mathcal M_\epsilon)$ as the largest observable error among all states with energy at most $E_0+\epsilon$, assuming a nondegenerate ground state and a bounded observable. Upper and constructive lower bounds establish $\delta_O=\sqrt{2\epsilon\chi_O}+\mathcal O(\epsilon)$ as $\epsilon\to0$ at fixed positive gap, where $\chi_O=2\sum_{n>0}|O_{0n}|^2/(E_n-E_0)$ is the static susceptibility. The finite-resolution upper bound separates an excited-population contribution, linear in $\epsilon$, from a ground–excited coherence contribution involving denominators $E_n-E_0+\epsilon$. The lower bound follows by displacing the ground state along its linear-response direction. Conserved observables lose the coherence contribution; symmetry-restricted state preparation can remove dangerous transitions. A target observable error $\delta_*$ consequently requires, at leading order, $\epsilon\lesssim\delta_*^2/(2\chi_O)$. Classical numerical demonstrations cover VQE on a 12-site Ising chain, DMRG on two separating hydrogen chains, and noisy Krylov diagonalisation of LiF. Reported examples include spin contamination $\langle S^2\rangle=1$ at $0.34\,\mathrm{m}E_h$ energy error and a 16-debye dipole error at $1.4\,\mathrm{m}E_h$. Symmetry-preserving or detuned comparisons avoid these failures. The radius is a worst-case compatibility statement, not a prediction that optimisation attains that error. The susceptibility asymptotic applies at small resolution relative to the relevant gap; finite-resolution bounds are capped when gaps become unresolved. Numerical figures select the largest observable error among near-best-energy runs, and the analysis assumes an energy-error certificate relative to the true ground energy.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 10 pages, 4 figures