SSP-QST: Spectral Subspace Purification for Photonic Quantum State Tomography

AI summary

Overview

Research area: Quantum information science, specifically photonic quantum state tomography (QST) and the reconstruction of low-rank entangled probe states used in quantum sensing.

Technical level: Intermediate. The method itself is a short closed-form algorithm, but Sections III-A and III-C rely on matrix perturbation theory (Weyl's theorem, matrix Bernstein), Uhlmann fidelity, and the quantum Fisher information formalism, which push parts of the paper toward Advanced.

Scope (one sentence): The paper introduces SSP-QST, a single-eigendecomposition post-processing layer that sits on top of least-squares quantum state tomography and removes noise-inflated eigenmodes from the reconstructed density matrix without being told the true rank, evaluated in Qiskit Aer simulations for up to 7-rank probes on 4-qubit systems.

What This Paper Is About

Ideal photonic sensing probes such as GHZ, Bell, and NOON states are rank-1, but least-squares tomography reconstructs them as full-rank density matrices littered with small, noise-induced eigenvalues, which raises the estimated entropy and reduces the quantum Fisher information available downstream. Forcing the answer to be rank-1 fixes that problem but breaks when real hardware (beam-splitter imbalance, multiphoton emission, waveguide crosstalk, photon loss, source drift) gives the probe an effective rank greater than one.

The goal is to recover the correct low-rank signal subspace from the measured spectrum alone, with no rank prior, no iteration, and low enough latency to sit inside a feedback loop that corrects source drift.

Key Contributions

1. Data-driven rank selection for LS-QST. A closed-form post-processing layer that picks the reconstruction rank from the eigenspectrum of the least-squares estimate, using a Weyl-perturbation-motivated noise floor computed from the dominant eigenvalue and the shot count, with no knowledge of the true rank.

2. One rule that handles both pure and mixed low-rank probes. SSP-QST avoids the two opposite failure modes of non-adaptive reconstruction: unlike LS-QST it does not keep every noise-inflated eigenmode, and unlike rank-one purification it does not throw away valid signal modes when the effective probe rank is greater than one. A rank-1 override handles the effectively pure case.

3. Perturbative fidelity-gain analysis plus simulation validation. A proposition showing the fidelity gain scales as p(d−1)/d for a rank-1 target and as p(d−r)/d for a rank-r target, together with Qiskit Aer simulations showing the highest fidelity among the tested non-iterative baselines at every evaluated probe rank.

4. A low-latency primitive for feedback-oriented sensing. Only one eigendecomposition and a renormalisation, giving O(d^3) closed-form reconstruction, plus a parameter-shift fidelity-gradient controller that uses the purified state to correct coherent source drift.

Main Findings

• Best fidelity at every tested rank. Table II (4 qubits, d = 16, depolarising noise p = 0.06, 4096 shots per Pauli basis) reports SSP-QST at 1.000, 0.980, 0.993, 0.987, 0.978, 0.965, and 0.941 for ranks 1 through 7, versus LS-QST at 0.883, 0.892, 0.901, 0.909, 0.916, 0.924, 0.933 and spectral squaring at 0.997, 0.972, 0.957, 0.937, 0.940, 0.929, 0.926.

• Maximum fidelity gain of +0.584. The gain over top-eigenvector extraction grows from +0.204 at rank-2 to +0.584 at rank-7, reflecting the structural mismatch between rank-1 reconstruction and a rank-r signal subspace.

• Small but consistent edge over spectral squaring. The gain over spectral squaring is positive across all mixed-state ranks, between +0.008 and +0.051, peaking at +0.051 at rank-4.

• Advantage survives a noise sweep. Sweeping depolarising rate p over [0, 0.22] at fixed rank-3 and rank-5 probes keeps the gain over the strongest baseline strictly positive, widest at low noise and narrowing smoothly at high noise as eigenvalue contrasts compress toward p/d.

• At least 8× shot efficiency. For a rank-3 probe at p = 0.08, SSP-QST reaches fidelity 0.956 with 512 shots, while LS-QST reaches only 0.890 at 4096 shots; across the tested range LS-QST never matches that 512-shot fidelity, giving at least an 8× photon-budget reduction.

• Rank identification is accurate and errs safely. For p ≤ 0.10, SSP-QST identified all true ranks 2 through 6 exactly in every one of 20 independent trials; at higher noise the identified rank underestimates the truth by at most one mode, never adding unsupported signal modes.

• Threshold is not finely tuned. Across four threshold variants, the identified rank stayed correct for the default and nearby settings and reconstruction fidelity changed by less than 0.025.

• The rigorous worst-case bound was deliberately not used. A matrix Bernstein bound gives a worst-case deviation of 0.041 in this setting; substituting it into the threshold inflates the threshold by roughly 5× and over-truncates, dropping rank-7 fidelity from 0.941 to 0.786.

• Amplitude-damping envelope is provable but loose. The bound gives ||Δ||_op ≤ 2nγ(1+γ), while observed shifts stay within nγ, the largest being 0.64 nγ at n = 4 and γ = 0.06.

Methodology in Plain English

The researchers start from the standard least-squares reconstruction, which always comes back with as many eigenvalues as the Hilbert-space dimension no matter how simple the true state is. They then diagonalise that estimate and ask a simple question of each eigenvalue: is this real signal, or is it noise?

To answer that, they borrow a classical result from matrix analysis, Weyl's perturbation theorem, which bounds how far each eigenvalue of a noisy matrix can drift from the corresponding eigenvalue of the true state. For a nearly pure target under depolarising noise, the deviation from one of the largest eigenvalue is itself a decent estimate of the noise level, and noise eigenmodes should sit around that level divided by d−1. They add a small margin of half a standard deviation of shot noise to avoid deleting real signal by accident, then throw away everything below the resulting floor and renormalise what remains. A special case check catches effectively pure probes and returns the single dominant eigenvector exactly.

They then prove, perturbatively, that deleting those noise modes can only help fidelity, and that the benefit grows with noise strength and with the difference between Hilbert-space dimension and probe rank. For the feedback part, they treat the photonic source as a parameterised circuit (GHZ preparation plus a layer of adjustable rotations modelling wave plates) and ascend the purified-state fidelity using the parameter-shift rule, which is exact for that gate family.

The experiments are pure software: Qiskit 2.4.1 with Qiskit-Aer 0.17.2, the density_matrix backend for exact evolution under gate-level Kraus noise, analytic shot noise at 4096 shots per Pauli basis, and random seed 42. Random rank-r target states come from QR-orthonormalising a complex Ginibre matrix (giving a Haar-distributed signal subspace) with Dirichlet(1,...,1) weights (uniform over the simplex, so no spectral profile is favoured). Three channels are tested — depolarising, amplitude damping, and phase damping — and the head-to-head comparison is deliberately restricted to non-iterative baselines (LS-QST, spectral squaring, top-eigenvector extraction) because iterative MLE, compressed sensing, and Bayesian methods exceed the latency budget set by photonic source drift.

Why This Matters

Impact on research. The paper reframes a practical annoyance in tomography as something with a clean middle path: neither keep-all (LS-QST) nor keep-one (rank-1 purification), but keep-what-rises-above-the-noise-floor. The link it draws between eigenvalue inflation, entropy, and quantum Fisher information makes the case that reconstruction quality is not a cosmetic concern but a limit on sensing precision, since the quantum Cramér-Rao bound is evaluated on the reconstructed state. The perturbative gain scaling as p(d−r)/d suggests the benefit strengthens as photonic systems scale, and the explicit negative result — that the rigorous Bernstein bound over-truncates — is a useful practical caution for anyone reaching for worst-case concentration bounds.

Real-world applications (as motivated in the paper):

• Photonic phase sensing with GHZ, Bell, or NOON probes, where reconstruction fidelity feeds directly into the achievable precision.
• Integrated silicon-photonic chips, where evanescent coupling between adjacent waveguides mixes spatial modes and raises the effective rank in a way that fixed rank-1 methods cannot represent.
• SPDC-based sources, where simultaneous multi-pair emission creates higher-order Fock components that a rank-adaptive reconstruction can retain rather than discard.
• Adaptive or feedback-driven sensing protocols, where source parameter drift continuously displaces the prepared state and low-latency reconstruction is required.

Industry relevance. The O(d^3) cost with a single eigendecomposition, completing in sub-microsecond time for n = 4 on a modern processor, positions the method as a drop-in layer rather than a replacement pipeline. The paper points to recent FPGA-based QST implementations as evidence that comparable reconstruction workloads can be deployed on reconfigurable hardware, which matters for any deployment where photon production rate, detector dead time, or experiment duration caps the shot budget.

Future Directions

• A tight structured-noise derivation. The threshold's first term is exact only for the depolarising channel; under amplitude and phase damping the perturbation is structured rather than isotropic, and the paper is explicit that its use there is empirically justified rather than rigorously derived. The authors state a tight structured-noise derivation is future work.

• Closing the gap between the typical-case margin and the worst-case bound. The Bernstein floor is called the rigorous uniform fallback, but substituting it inflates the threshold roughly 5× and over-truncates. Reconciling the two would give a threshold with both guarantees and practical accuracy.

• Validation beyond simulation. All results are from Qiskit Aer software simulation, and the paper states no physical hardware is used or implied. Testing the rank-identification behaviour on real detectors, real dark counts, and real drift statistics is the natural next step, along with the FPGA deployment the authors gesture toward.

• Probing the high-rank, high-noise corner. The gain weakens monotonically as r approaches d, and rank identification begins underestimating at higher noise. How the method behaves for larger n and for probes whose rank is a substantial fraction of d is left open.

Target Audience

Experimental photonics groups and quantum-sensing engineers who already run tomography and want a cheap fix for full-rank reconstructions; quantum information researchers interested in spectral and perturbation-theoretic approaches to state estimation; and hardware/systems engineers building low-latency reconstruction pipelines for closed-loop control. Readers without linear algebra background will find the algorithm section readable but will need to work through Section III-A and III-C to follow the threshold justification and the fidelity bound.

Authors’ abstract

Photonic quantum sensing often uses low-rank entangled probes such as Greenberger-Horne-Zeilinger (GHZ), Bell, and NOON states. Although these probes are ideally rank-1, practical quantum state tomography (QST) can produce density-matrix estimates with many small finite-shot and noise-induced eigenmodes. This eigenvalue contamination can increase the estimated entropy of the reconstruction and reduce the quantum Fisher information (QFI) available for downstream sensing, while fixed rank-1 purification can discard valid signal modes when real probes acquire additional signal modes. We introduce Spectral Subspace Purification for Quantum State Tomography (SSP-QST), a rank-adaptive post-processing layer for least-squares quantum state tomography (LS-QST). SSP-QST eigendecomposes the least-squares estimate, computes a Weyl-motivated noise floor from the measured spectrum and shot count, removes eigenmodes below this floor, and renormalises the retained subspace. It requires no rank prior, no iterative optimisation, and only one eigendecomposition. In Qiskit Aer simulations, SSP-QST achieves the highest fidelity among the tested non-iterative baselines across the evaluated probe ranks, with a maximum fidelity gain of $+0.584$. It also improves shot efficiency by at least $8\times$ within the tested range. These results show that SSP-QST can make photonic QST more reliable under finite-shot noise while providing a lightweight reconstruction primitive for feedback-oriented quantum sensing pipelines.