Density functional theory (DFT) is the workhorse method for computing how electrons arrange themselves in molecules and materials. Its central object is the electron density, the number of electrons per unit volume at each point in space. In exact quantum mechanics the density operators at different points all commute, so the density at every point can in principle be read out from a single measurement of a prepared state. Real calculations never work in the exact continuum, though. They expand everything in a finite set of atom-centred basis functions, and once the density operator is squeezed into such a basis, its values at two points generally stop commuting. This note proposes to quantify this basis-induced measurement incompatibility with the tools quantum information theory built for exactly this purpose, and to test whether it sets a real price for collecting density labels from a quantum computer. Such labels are what machine-learned DFT surrogates are trained on.
The simple version is this. The finite basis makes the density a set of mutually disturbing measurements, and we can compute how much noise must be added before they become jointly measurable. That number should predict how many repetitions of a quantum experiment are needed to learn a density. It should also tell us which bases and grids make the density cheap or expensive to learn. If the prediction holds, it is a design rule for quantum-generated training data. If it fails, the failure itself would show that the incompatibility literature's operational measures do not transfer to chemistry-scale observables.
Core Claim
In a basis of $K$ real functions $\chi_\mu$, the projected density operator at point $r$ is $n_K(r)=\sum_{\mu\nu}\chi_\mu(r)\chi_\nu(r)\,a^\dagger_\mu a_\nu$, where $a^\dagger_\mu$ and $a_\nu$ create and destroy an electron in basis function $\mu$ or $\nu$. Commutators of one-body operators reduce to commutators of their $K\times K$ coefficient matrices. The coefficient matrix at $r$ is the rank-one matrix $\chi(r)\chi(r)^{T}$. Two such matrices commute only when the vectors $\chi(r)$ and $\chi(r')$ are parallel or orthogonal. The relevant quantity is the normalized overlap $$\kappa_K(r,r')=\frac{\chi(r)\cdot\chi(r')}{\lVert\chi(r)\rVert\,\lVert\chi(r')\rVert},$$ the cosine of the angle between the basis-function vectors evaluated at the two points.
As the basis becomes complete, $\sum_\mu \chi_\mu(r)\chi_\mu(r')$ approaches a delta function. Distant points therefore become orthogonal and compatible, and very close points become nearly parallel and also compatible.
We claim three things. First, incompatibility is concentrated at separations comparable to the basis's spatial resolution. In the single-electron sector, pairwise noise robustness is a closed-form function of $\kappa_K$ because each pair spans a two-dimensional space and behaves like a qubit, and it peaks where $|\kappa_K|$ is intermediate. Second, this structure persists in small many-electron systems and in large sets of grid points. Third, a set-level incompatibility robustness predicts the shot cost of estimating the density better than the commutator norms chemists would reach for first.
This is the mirror image of the subspace phenomenon in Quantum measurement incompatibility in subspaces. That work projects incompatible measurements onto smaller spaces and asks when incompatibility disappears. Here, projecting compatible continuum observables onto a finite basis creates incompatibility.
Research Program
The work has three layers.
The first layer is analytical and runs in the one-electron sector. There we derive the pairwise robustness as a function of $\kappa_K$ under white noise and under the other noise models catalogued in Incompatibility robustness of quantum measurements: a unified framework. We then map $\kappa_K(r,r')$ for real molecular basis sets, STO-3G, 6-31G and Def2SVP, around H$_2$, LiH and H$_2$O. This yields a spatial "incompatibility kernel" for each basis.
The second layer is set-level and numerical. Measuring a density means measuring it at dozens to hundreds of grid points, which is far beyond the roughly 20 measurements at which semidefinite programs (SDPs), the standard convex solvers used to decide joint measurability, become intractable. For large sets we use the linear-programming hierarchy of Measurement incompatibility and quantum steering via linear programming. That hierarchy sandwiches the state space between inner and outer polytopes and scales linearly with the number of measurements.
This method has a known weakness. Its polytopes are practical only in very low dimension, so it cannot be applied to the full $K$-dimensional problem directly. We handle this in two ways. We use exact SDPs on subsets of up to about 15 points to anchor the scale. We apply the LP hierarchy to the effectively low-dimensional sub-blocks that the kernel identifies as the incompatible ones. Both upper and lower bounds are reported, so any gap between them is visible rather than hidden.
Many-electron checks use exact ground states of H$_2$, LiH and H$_4$ in minimal and split-valence bases. In these systems the Fock-space dimension still allows direct SDPs on small point sets.
The third layer is operational. From the SDP solution we extract an optimal parent measurement: a single measurement whose outcomes, after classical post-processing, reproduce noisy versions of all the target density measurements at once. We turn it into an estimator and compare its shot cost against the standard ways of measuring densities on quantum hardware.
Throughout, we compute at least two robustness measures, generalized robustness and white-noise robustness. The unified-framework paper showed that such measures can rank measurement sets differently, and we want to see whether any conclusion depends on that choice. A conclusion that holds under only one measure will be reported as measure-dependent.
Baselines
The naive reference measures each nonzero 1-RDM element separately in its own measurement setting and reconstructs the density from them. The 1-RDM, or one-particle reduced density matrix, is the table of expectation values $\langle a^\dagger_\mu a_\nu\rangle$ from which the density is assembled. This naive scheme sets the ceiling on shot cost that every method must beat.
The strongest practical competitor is fermionic classical shadow tomography together with commuting-operator grouping, as used in Efficient quantum computation of molecular forces and other energy gradients. Classical shadows, in the sense of New Bounds on Quantum Sample Complexity of Measurement Classes, are classical snapshots from randomized measurements that can be reused to estimate many observables. The O'Brien et al. scheme already exploits shared structure among many observables of one state, and it is the approach a quantum-chemistry group would actually deploy. We run it on the same exact ground states and the same grids, with the same total shot budgets, and with grouping heuristics tuned on a held-out molecule.
The density-accuracy target is anchored to the best classical surrogate in the file. A Function-Centric Graph Neural Network Approach for Predicting Electron Densities reports a normalized mean absolute error of 0.116% on QM9 (PySCF). Quantum-generated labels are only worth having as training data if their shot-noise error sits below this level. We therefore report the shots each method needs to reach a normalized error of 0.1%.
No reported result was found on Lacuna for the shot cost of estimating an electron density with fermionic shadows. Those numbers will be produced here under identical conditions for all methods.
The proposed parent-measurement estimator beats the baselines if it reaches 0.1% error with fewer shots than tuned shadows plus grouping on a majority of molecule–basis pairs.
The key ablation removes the ingredient that creates incompatibility. We re-express the same state in a grid-like basis in which the projected density operators are close to diagonal. If incompatibility is what drives cost, the gap between methods should close in that basis.
Why not just…?
Why not just use commutator norms instead of incompatibility robustness? Commutator norms are cheap, and for pairs of one-body observables they are closely tied to $\kappa_K$. They therefore may capture most of the pairwise structure, and we treat this as a real threat. The unified framework shows that robustness measures encode how much noise a joint measurement must tolerate, which a commutator does not. The subspace paper shows that joint measurability and coexistence already come apart for qubits, a distinction commutators cannot express. Whether this matters for densities is an empirical question. The decisive test is the rank correlation between each quantity and the measured shot cost across molecule–basis–grid combinations. If commutator norms predict cost as well as robustness does, we will say so and keep only the structural result.
Why not just rotate into a real-space or grid basis where the density commutes? This is the most serious alternative, and it is exactly the ablation above. The trade-off it hides is that grid-like bases need many more functions to reach the same energy accuracy, and their added qubits and circuit depth cost resources too. The robustness map says how far from diagonal the density can be before measurement cost grows. That tells us whether a cheap atom-centred basis with a modest grid rotation is good enough. If full rotation is always cheap and accurate, the practical case for the robustness machinery narrows to a diagnostic role.
Why not just use generic classical shadows, which already estimate everything at once? Heidari and Szpankowski show that shadow sample complexity is governed by a shadow norm that grows with the incompatibility of the measured class, so generic shadows are not free. What their result does not say is whether a measurement tailored to the density set can do better. The SDP parent measurement gives the best single-setting scheme for this specific set. Comparing its cost with shadows shows how much headroom exists, and a small gap would itself be a useful negative result.
Why not just train classical surrogates like BOA or DEQHNet on classical DFT data and skip quantum labels? For weakly correlated molecules this is the right answer. Infusing Self-Consistency into Density Functional Theory Hamiltonian Prediction via Deep Equilibrium Models reports close to a 50% error reduction over QHNet on QH9-dynamic. QH9: A Quantum Hamiltonian Prediction Benchmark for QM9 Molecules reports 76.31×10⁻⁶ Hartree mean absolute error on QH9-stable-id. The case for quantum labels rests on strongly correlated systems, where semilocal functionals fail. This proposal does not argue that case. It prices the data if someone wants it.
Proposed Experiments
The first experiment tests the kernel prediction. For H$_2$, LiH and H$_2$O in STO-3G, 6-31G and Def2SVP, we plot pairwise robustness against point separation. The prediction is a peak at a basis-dependent length that shrinks as the basis grows. A monotone curve, or a peak that does not move with basis size, would falsify the resolution-scale picture.
The second experiment checks set-level robustness. We compute it for grids of 10 to 400 points with exact SDPs and the LP hierarchy, and verify that the bounds bracket each other. We test whether the set value is predicted by aggregating pairwise kernels or requires genuinely collective computation. Collective incompatibility beyond pairs would be the more interesting outcome.
The third experiment is the shot-cost study. For each molecule–basis–grid combination, we measure the shots needed to reach 0.1% normalized density error with each of the naive per-element scheme, grouped fermionic shadows, the parent-measurement estimator, and the grid-rotated ablation. We also record the error of the energy estimated from the same shots. We then compute the rank correlation between shot cost and robustness, and between shot cost and commutator norm.
As a downstream check, noisy densities from each method seed PySCF self-consistent iterations, and we count iterations to convergence. BOA's 35% iteration reduction and DEQHNet's 2 to 3 iterations serve as reference points.
The single decisive check is the correlation comparison. Robustness must predict measured cost clearly better than commutator norms, under both robustness measures, for the incompatibility framing to earn its place in quantum chemistry practice.
References
- Quantum measurement incompatibility in subspaces
- Incompatibility robustness of quantum measurements: a unified framework
- Measurement incompatibility and quantum steering via linear programming
- Efficient quantum computation of molecular forces and other energy gradients
- New Bounds on Quantum Sample Complexity of Measurement Classes
- A Function-Centric Graph Neural Network Approach for Predicting Electron Densities
- Infusing Self-Consistency into Density Functional Theory Hamiltonian Prediction via Deep Equilibrium Models
- QH9: A Quantum Hamiltonian Prediction Benchmark for QM9 Molecules