Universal low-rank matrix recovery from Pauli measurements

We study the problem of reconstructing an unknown matrix M of rank r and dimension d using O(rd poly log d) Pauli measurements. This has applications in quantum state tomography, and is a non-commutative analogue of a well-known problem in compressed sensing: recovering a sparse vector from a few of its Fourier coefficients. We show that almost all sets of O(rd log^6 d) Pauli measurements satisfy the rank-r restricted isometry property (RIP). This implies that M can be recovered from a fixed (“universal”) set of Pauli measurements, using nuclear-norm minimization (e.g., the matrix Lasso), with nearly-optimal bounds on the error. A similar result holds for any class of measurements that use an orthonormal operator basis whose elements have small operator norm. Our proof uses Dudley’s inequality for Gaussian processes, together with bounds on covering numbers obtained via entropy duality.

💡 Research Summary

The paper addresses the problem of recovering an unknown d × d matrix M of rank r (with r ≪ d) from a small number of linear measurements of the form tr(M · A_i). The measurements considered are Pauli observables, i.e., tensor products of single‑qubit Pauli matrices, which form an orthonormal basis of the space of d × d matrices. The authors show that a random selection of m = C·K²·r d log⁶ d Pauli matrices (with K the bound on the operator norm of the basis elements, K = 1 for Pauli) satisfies the rank‑r restricted isometry property (RIP) with high probability. RIP means that for every matrix X with rank at most r (or, more generally, satisfying ‖X‖_* ≤ √r‖X‖_F) the sampling operator A obeys
(1 − δ)‖X‖_F² ≤ ‖A(X)‖₂² ≤ (1 + δ)‖X‖_F².

The proof departs from the standard concentration‑of‑measure + union‑bound technique used for Gaussian measurements because Pauli matrices are highly structured. Instead, the authors map the problem to a Gaussian process indexed by low‑rank matrices and apply Dudley’s entropy integral to bound its supremum. The crucial step is to control the covering numbers of the nuclear‑norm ball (‖X‖_* ≤ √r‖X‖_F). This is achieved via entropy duality, a technique introduced by Guédon et al., which yields covering bounds of order O(r d log⁶ d). Consequently, the failure probability of RIP decays exponentially in δ² C.

Once RIP is established, the paper leverages existing results on nuclear‑norm minimization for low‑rank matrix recovery. Two convex programs are considered: the matrix Dantzig selector (minimize ‖X‖* subject to ‖A⁎(y − A(X))‖ ≤ λ) and the matrix Lasso (minimize ½‖A(X) − y‖₂² + μ‖X‖). In the noiseless case (λ = 0) the Dantzig selector recovers M with nuclear‑norm error bounded by a constant times the nuclear norm of the residual part M_c (the tail after the best rank‑r approximation). In the presence of Gaussian noise z ∼ N(0,σ²I), choosing λ = 8√d σ or μ = 16√d σ yields, with high probability, a Frobenius‑norm error bound
‖\hat M − M‖F ≤ C₀√(r d) σ + C₁‖M_c‖/√r.
Thus the reconstruction error scales with the noise level and the size of the tail, and is essentially optimal up to constant factors.

The authors discuss the implications for quantum state tomography. A quantum state of n qubits is described by a density matrix ρ (positive semidefinite, trace 1) that is often low‑rank or well‑approximated by a low‑rank matrix. Pauli measurements are experimentally convenient, and the RIP result guarantees that a fixed set of O(r d poly log d) Pauli settings suffices for accurate reconstruction of any such state. The error bounds translate into guarantees on both the trace distance (via nuclear‑norm error) and the Hilbert‑Schmidt distance (via Frobenius error), which are directly relevant for distinguishing quantum states and for assessing the coherence of the reconstructed state.

The paper also notes that the analysis extends to any orthonormal operator basis whose elements have small operator norm (order 1/√d). Moreover, it warns that selecting only commuting subsets of Pauli operators (e.g., stabilizer states) would lead to information loss, but such pathological subsets are unlikely under random sampling.

In summary, the work provides a rigorous, near‑optimal guarantee that a universal (i.e., measurement‑independent) set of Pauli measurements enables stable, efficient recovery of low‑rank matrices via convex optimization. The combination of RIP for structured measurements, Dudley’s entropy bound, and entropy duality constitutes a significant technical advance, bridging compressed sensing theory with practical quantum tomography and other applications where low‑rank matrix models arise.