arXiv:2611.01079
Optimally Weighted Constrained Shadow Tomography for Reduced Density Matrices
Yuchen Wang, Irma Avdic, Michael Rose, Lillian I. Payne, Noah Garrett, Madhubani Mukherjee, Leah P. Weisburn, Nate Earnest-Noble, Kevin J. Sung, David A. Mazziotti
Abstract
Accurate measurement of reduced density matrices (RDMs) is central to quantum simulations of many-fermion systems, but the required measurement resources remain a significant challenge. Constrained shadow tomography (CST) combines randomized measurements with $N$-representability conditions to reconstruct physically valid RDMs from limited measurement data. Here we improve CST by orthogonalizing the measurement constraints and using the resulting singular values to determine the relative weight of each independent measurement direction according to how strongly it is sampled by the quantum computer. These optimally weighted constraints are incorporated into the semidefinite program through a least-squares objective implemented with a Schur-complement relaxation. We demonstrate the approach for hydrogen chains on a quantum simulator and for azobenzene on a quantum processor. In both cases, optimally weighted CST systematically improves the accuracy of the reconstructed 2-RDM without increasing the measurement budget, providing a more measurement-efficient approach to partial tomography for quantum simulations of correlated fermions.
AI Summary
The paper addresses reconstruction of fermionic two-particle reduced density matrices (2-RDMs) from limited randomized measurements. It modifies constrained shadow tomography (CST), which combines measured projections with semidefinite $N$-representability constraints, to account for unequal sampling of independent directions in 2-RDM space. The measurement map is orthogonalized by singular-value decomposition. Each retained direction receives weight $w_i=\sigma_i/\sigma_1$, and weighted squared residuals are minimized alongside the energy, subject to DQG positivity and normalization constraints. A Schur-complement formulation expresses the quadratic penalty as a semidefinite program. Measurement-operator overlaps permit construction of the Gram matrix without explicitly assembling the full measurement map; direct diagonalization costs $O(N_u^3r^6)$ for $N_u$ settings and orbital dimension $r$. Numerical hydrogen-chain benchmarks report lower Frobenius 2-RDM errors than the original CST across the tested shot counts and system sizes. For H$_6$, two weighted settings outperform three original-CST settings, each sampled with $10^4$ shots. Improvements are largest at lower shot counts, with both reconstructions approaching the same limiting accuracy as shots increase. Quantum-processor measurements for azobenzene in an eight-electron, eight-orbital active space use five settings and $10^4$ shots per setting. At the planar geometry, original and weighted CST reduce error relative to variational 2-RDM reconstruction by 43% and 54%, respectively; weighted CST also follows the reference excitation-gap profile more closely. The azobenzene calculation is restricted to the seniority-zero DOCI sector and does not describe its physical conical intersection. More scalable, structured orthogonalization is left for future work.
Central claim may not hold: The advertised optimality of singular-value weighting is unproved: without truncation its penalty is ordinary measurement-coordinate least squares up to a constant and scale, and no suitable statistical noise model is established. The numerical comparison changes the reconstruction penalty from a nuclear norm to squared measurement residuals, so it does not isolate the benefit of singular-value weighting. The implemented DQG constraints do not guarantee full N-representability, contrary to the unrestricted claim of physically valid reconstructed RDMs. The finite-λ energy-plus-squared-residual objective need not recover exact measurement projections even for complete noiseless data, invalidating the unconditional exact-data and infinite-shot assertions.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.