Projected Least Squares POVM Tomography

• Projected Least Squares POVM Tomography is a two-step method that first applies unconstrained least-squares estimation and then projects onto the set of valid quantum measurements.
• It leverages informationally complete probe ensembles and closed-form expressions for 2-design probes to simplify computational complexity.
• The approach provides rigorous non-asymptotic guarantees for both worst-case and average-case errors, ensuring sample-optimal performance in quantum measurement reconstruction.

Searching arXiv for the cited papers to ground the article in the relevant literature. arXiv search: (Zambrano et al., 6 Jul 2025) "Fast quantum measurement tomography with dimension-optimal error bounds" Projected least squares POVM tomography is a two-step protocol for reconstructing an unknown quantum measurement from data obtained with a known informationally complete probe ensemble. In the formulation introduced in "Fast quantum measurement tomography with dimension-optimal error bounds" (Zambrano et al., 6 Jul 2025), one first applies least-squares estimation to obtain an unconstrained approximation of each POVM element and then projects the resulting collection onto the set of valid quantum measurements. For an $L$-outcome POVM acting on a $d$-dimensional system, the method is designed to retain light classical co-processing while achieving non-asymptotic error guarantees and sample complexity that is sample-optimal in the system dimension $d$ up to logarithmic factors (Zambrano et al., 6 Jul 2025).

1. Definition and estimation framework

The target of inference is an unknown $L$-outcome POVM

$E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$

with $E_j\ge 0$ and $\sum_{j=1}^L E_j=I$. The protocol assumes access to a known informationally complete ensemble $\{\ket{\psi_i}\}_{i=1}^M$. In each of $N$ independent rounds, one chooses $i\in\{1,\ldots,M\}$ uniformly at random, prepares $d$0, measures $d$1, and records the observed outcome $d$2. The empirical frequencies are

$d$3

and satisfy

$d$4

where $d$5 is the linear measurement map (Zambrano et al., 6 Jul 2025).

The first step estimates each outcome operator independently by unconstrained least squares: $d$6 equivalently as the unique $d$7 minimizing $d$8. The second step enforces physicality jointly across all outcomes by solving

$d$9

where $d$0 is a chosen metric on POVMs (Zambrano et al., 6 Jul 2025).

A recurrent misconception is to identify POVM projection with separate positivity corrections on the individual $d$1. The formulation above is stricter: the projection is onto the full feasible set $d$2, so the normalization constraint is imposed collectively rather than effect-by-effect. This suggests that the method is best understood as a constrained estimation procedure on the measurement object as a whole, not as $d$3 decoupled matrix denoisings.

2. Two-step projected least squares construction

Algorithmically, the protocol is stated in terms of precomputed frame operators $d$4. The raw least-squares estimate can be written as

$d$5

after which one solves the projection problem over valid POVMs. The resulting estimator is the physical POVM $d$6 (Zambrano et al., 6 Jul 2025).

The projection step is metric-dependent. The underlying paper allows either the worst-case operational distance or the average-case distance, and the projection can be carried out by an SDP or, for large $d$7, by block-coordinate or first-order methods (Zambrano et al., 6 Jul 2025). This separates the procedure into a linear inversion stage and a convex feasibility-and-proximity stage.

The construction is closely related to projected least squares in quantum state tomography. In "Fast state tomography with optimal error bounds" (Guta et al., 2018), one first computes the least-squares estimator and then projects onto the state space; in "A comparative study of estimation methods in quantum tomography" (Acharya et al., 2019), least squares and projected least squares are analyzed within a broader family of estimators that project data onto parameter spaces with respect to specific metrics. Projected least squares POVM tomography transfers that paradigm from density operators to measurement operators, but the geometry changes because the feasible set is a product of PSD constraints coupled by the identity-resolution condition.

3. Closed-form least squares for 2-design probes

A principal simplification occurs when the probe ensemble is a complex projective 2-design $d$8, satisfying

$d$9

In this case,

$L$0

and the raw estimator simplifies to

$L$1

The paper also states that the method admits an analytic form when using local 2-designs, including tensor-product 2-design probes on $L$2 qubits (Zambrano et al., 6 Jul 2025).

This closed form is the measurement-tomography analogue of the 2-design formulas that make projected least squares especially attractive in state tomography. For state estimation, the same 2-design structure yields explicit inverses for $L$3 and numerically cheap linear inversion (Guta et al., 2018). In the POVM setting, the same phenomenon removes any large-scale matrix inversion beyond the precomputed frame operators, which is central to the method’s intended low co-processing cost.

4. Distances and non-asymptotic guarantees

Two distances are used to quantify reconstruction error (Zambrano et al., 6 Jul 2025). The operational, or worst-case, distance is

$L$4

The average-case distance is

$L$5

The operational distance equals the maximum total-variation distance between outcome distributions over all possible input states. The average-case distance corresponds to the expected $L$6-error of the outcome distribution, averaged over input states drawn from a spherical 4-design, in the sense that

$L$7

For global 2-design probes, the non-asymptotic high-probability guarantees are as follows.

Target metric	Guarantee	Sufficient sample size
$L$8	confidence $L$9	$E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$0
$E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$1	confidence $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$2	$E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$3

The corresponding asymptotic forms stated in the abstract are $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$4 samples for worst-case distance and $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$5 samples for average-case distance (Zambrano et al., 6 Jul 2025). For local $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$6-qubit 2-designs, the theorem replaces $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$7 by $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$8 and $E=\{E_1,\dots,E_L\}\subset \mathbb{C}^{d\times d},$9 by $E_j\ge 0$0 in the worst-case numerator, and replaces $E_j\ge 0$1 by $E_j\ge 0$2 and $E_j\ge 0$3 by $E_j\ge 0$4 in the average-case bound (Zambrano et al., 6 Jul 2025).

The lower bounds establish that any non-adaptive, single-copy POVM tomography protocol must use at least

$E_j\ge 0$5

to achieve accuracy $E_j\ge 0$6 in $E_j\ge 0$7, and at least

$E_j\ge 0$8

to achieve accuracy $E_j\ge 0$9 in $\sum_{j=1}^L E_j=I$0 (Zambrano et al., 6 Jul 2025). Accordingly, the protocol is sample-optimal in dimension $\sum_{j=1}^L E_j=I$1 up to logarithmic factors, but the lower-bound comparison is explicitly restricted to non-adaptive, single-copy schemes. This suggests that the optimality claim is dimension-theoretic rather than universal across all conceivable tomography models.

5. Proof architecture and computational profile

The proof strategy follows the same general template as projected least squares state tomography, but now at the level of measurement effects. For the worst-case bound, one shows by matrix-Bernstein that

$\sum_{j=1}^L E_j=I$2

once $\sum_{j=1}^L E_j=I$3. A union bound over $\sum_{j=1}^L E_j=I$4 subsets then controls $\sum_{j=1}^L E_j=I$5, and the projection onto POVMs can only improve the worst-case distance by triangle inequality and the definition of projection (Zambrano et al., 6 Jul 2025). For the average-case theorem, the Frobenius part and trace part in $\sum_{j=1}^L E_j=I$6 are bounded separately via matrix-Bernstein and Hoeffding’s inequality and then combined (Zambrano et al., 6 Jul 2025).

The lower bounds are obtained by packing-and-Fano arguments: one constructs an exponentially large family of well-separated POVMs, encodes a random label, upper-bounds the information gained per measurement via a $\sum_{j=1}^L E_j=I$7-divergence argument, and applies Fano’s inequality (Zambrano et al., 6 Jul 2025).

The classical post-processing cost is polynomial in $\sum_{j=1}^L E_j=I$8. Given $\sum_{j=1}^L E_j=I$9 and precomputed $\{\ket{\psi_i}\}_{i=1}^M$0, each raw $\{\ket{\psi_i}\}_{i=1}^M$1 is a weighted sum of $\{\ket{\psi_i}\}_{i=1}^M$2 rank-1 terms, with cost $\{\ket{\psi_i}\}_{i=1}^M$3 per outcome or $\{\ket{\psi_i}\}_{i=1}^M$4 overall; if $\{\ket{\psi_i}\}_{i=1}^M$5 for a global 2-design, this becomes $\{\ket{\psi_i}\}_{i=1}^M$6. The projection step is an SDP of size $\{\ket{\psi_i}\}_{i=1}^M$7. Off-the-shelf interior-point solvers run in $\{\ket{\psi_i}\}_{i=1}^M$8 worst-case, while first-order or alternating-projection schemes reduce it in practice to roughly $\{\ket{\psi_i}\}_{i=1}^M$9 per iteration (Zambrano et al., 6 Jul 2025). Thus the method avoids any large-scale matrix inversion beyond the closed-form 2-design formulas, but its overall runtime still depends materially on the projection stage.

6. Empirical demonstration and related developments

The protocol has been demonstrated on two flux-tunable transmon qubits with chip $N$0, $N$1, readout fidelities $N$2–$N$3, and single-qubit gate errors $N$4 (Zambrano et al., 6 Jul 2025). In one experiment, a single-qubit SIC-POVM was implemented via an ancilla qubit and controlled rotations. Using $N$5 random input states from a single-qubit 2-design, each POVM element was recovered with worst-case distance $N$6 from the ideal. The same study reconstructed the corresponding half-sided measurement channel in the Pauli basis, which was needed for classical-shadow error-mitigation, and reported that the experimental inversion closely matches the ideal depolarizing-channel inversion. Additional two-qubit POVMs—identity, Hadamard, Bell, and Haar-random—were studied with $N$7 shots, showing high fidelity to the targets and behavior consistent with the $N$8 scaling predicted by the non-asymptotic bounds (Zambrano et al., 6 Jul 2025).

Within the broader projected least-squares literature, this work extends a pattern already visible in state tomography. The state-tomography version provides rigorous non-asymptotic confidence regions in trace distance and, for the uniform POVM, saturates known lower bounds up to constants (Guta et al., 2018). Comparative analysis of quantum tomography estimators further shows that projected linear estimators can combine strong statistical behavior with high computational efficiency, especially in low-rank regimes (Acharya et al., 2019). A distinct line of work applies constrained least squares to structured state classes such as matrix product operators, using informationally complete POVMs including SIC-POVMs and spherical $N$9-designs and obtaining sample complexity proportional, up to logarithms, to the number of independent parameters (Qin et al., 2024). Taken together, these developments place projected least squares POVM tomography within a larger family of projection-based quantum estimation methods in which closed-form linear inversion, concentration inequalities, and convex projection are the defining ingredients.