arXiv:2610.34696
Rank-one Factored Gradient Descent without Restricted Isometry: Application to Quantum Pure State Tomography on Multi-Qubits
Takahiro Kajisa
Abstract
Factored gradient descent (FGD) is a scalable nonconvex method for low-rank positive semidefinite matrix recovery, in which the rank and positivity constraints are satisfied automatically by the factorization. Most of the existing theory that certifies benignity of its loss landscape, however, relies on a restricted isometry property (RIP) of the measurement map. In quantum state tomography, one of its applications, the measurements that can be acquired with stable statistics are limited by detection efficiency and drift to Pauli observables of low weight; we first show that under such partial-weight Pauli measurements RIP fails in the worst possible way. We then bundle the mutually commuting observables setting by setting, pushing the state forward to the probability distribution of each setting, and introduce a max-sliced total-variation loss. On the stratum where the setting attaining the maximum is unique there is no spurious local minimum, and provided only that the measurement map is injective on generic pure states, the global minimum is the truth up to a phase. Finally we verify these claims numerically on four qubits and compare against FGD with the ordinary loss.
AI Summary
The paper studies rank-one factored optimization for pure-state tomography when only partial-weight Pauli measurements are available. For $n=2N$ qubits, it considers all Pauli expectations of weight at most $N$, invoking generic identifiability among pure states. A GHZ-phase family shows that measurements supported on proper subsets have a nonzero rank-two kernel element, ruling out rank-two restricted isometry with $\delta_2<1$. The proposed loss groups commuting observables into measurement settings and minimizes the maximum, over settings, of the $\ell^1$ discrepancy between predicted and observed outcome distributions. Its landscape theorem applies at positive-loss points with a unique maximizing setting: stationarity holds exactly when every deficient outcome cell has zero predicted probability. Such stationary points admit an explicit second-order descent direction, so no positive-loss local minimum occurs on this stratum. The proof characterizes realizable first-order probability flows and constructs descent by populating an empty deficient cell. With exact data and an identifiable truth, zero loss determines the state up to phase. Four-qubit simulations use 54 settings and perturbed cluster-state targets. On noiseless data, momentum-assisted max-sliced descent succeeds in 100/100 trials, versus 75/100 for Pauli least squares. With finite shots, max-sliced initialization followed by least-squares polishing performs best among the tested schemes. The theorem leaves multiple-maximizer tie points unresolved; no global convergence guarantee is established. Experiments are limited to four qubits, omit detection inefficiency and drift, and report some near-tie stalling at low shot counts.
we used generative AI tools for implementing the numerical experiments, for translating and copy-editing the manuscript, and for producing the figure sources.
Error found: Generic injectivity guarantees a unique zero-loss solution only for an identifiable truth, a qualification missing from the abstract and required by the paper's own GHZ counterexample. The generic-identifiability proposition and global-uniqueness corollary incorrectly include two qubits, whose generic pure states retain an undetermined relative Schmidt phase when only one-qubit marginals are measured.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.