Readout Orientation Controls Measurement-Accessible Quantum Tangent Geometry

AI summary

Overview

Research area: Quantum information / variational quantum circuits — specifically measurement-induced tangent geometry (classical Fisher geometry after a fixed measurement) and how a restricted readout projects it.

Technical level: Advanced. The core idea is intuitive, but the paper relies on Grassmann-manifold projector moments, spectral bounds, Ky Fan variational arguments, and cross-fitted estimation.

Scope (one sentence): The paper shows that, for a fixed quantum measurement, the orientation of a rank-r readout subspace relative to the tangent-score covariance — not the rank itself — determines how much measurement-accessible tangent information survives, and demonstrates an operational consequence at fixed rank and shot budget.

Author: Marwan Ait Haddou (Independent Researcher, Morocco), arXiv:2608.17085v1 [quant-ph], 17 Aug 2026.

What This Paper Is About

A variational quantum circuit is only ever accessed through measurement, and then through a restricted family of observables or features (a "readout"). Prior work shows that fixing a measurement contracts the quantum Fisher geometry to a classical one; this paper studies the second contraction, where only a rank-r readout subspace is retained. The central question: after the measurement and the readout dimension are both fixed, how much tangent information is actually accessible to that particular readout subspace?

Key Contributions

1. A rank-controlled accounting of readout accessibility. Defines retained tangent mass R(P,C) = Tr(PC) for a trace-one tangent covariance C and rank-r projector P, and the normalized ratio ρ = R/(r/N), where r/N is the rank-only random-orientation reference. The identity Tr(PC) is shown to equal the mean retained squared score norm, so R is operational rather than abstract.

2. A variance bound that separates rank from isotropy. Derives Var(ρ) = 2(N−r)[N Tr(C²)−1] / (r(N−1)(N+2)) ≤ 2/(r·d_eff) with d_eff = 1/Tr(C²), together with a Chebyshev-type concentration bound. The key consequence: rank-typical overlap only requires the product r·d_eff to be large, so near-baseline retention can coexist with strongly anisotropic covariance. The author explicitly frames this as a null model built from standard Grassmann geometry, not a new random-projection theorem.

3. Equal-rank, cross-fitted comparisons against physical and random readouts. Uses the Ky Fan variational principle to define the optimal rank-r subspace, cross-fits it on independent tangent data, and compares it with the physical low-weight span and a random rank-matched subspace at fixed circuit, measurement record, rank, and shot budget.

4. A structured symmetry case study with sector-corrected nulls. Provides a half-filled U(1)-conserving RZ–XY family in which physical low-weight Z readouts are already strongly aligned with leading tangent directions, using a random-orientation null drawn inside the same half-filled sector rather than in the full 2^n score space.

Main Findings

• Generic low-weight readouts stay near the rank baseline through n = 16. For the family-balanced generic ensemble, one- and two-body readouts remain within the preregistered rank-equivalence band through n = 16. A targeted Haar-U(4) stress test at n = 18 gives ρ₁ = 0.929 [0.910, 0.950] and ρ₂ = 0.958 [0.952, 0.966]; the paper notes this is a single-architecture stress test, not the family-balanced estimator.

• Near-baseline retention does not mean isotropy. The normalized purity N Tr(C²) grows from approximately 1.35 at n = 6 to 6.17 at n = 12, 17.52 at n = 14, and 49.84 at n = 16, with the n = 18 Haar-U(4) stress point at approximately 146.8. Correspondingly, d_eff/N falls from 0.746 at n = 6 to 0.1767 at n = 12 and 0.0233 at n = 16, with the n = 18 Haar-U(4) value near 0.00683.

• Architecture-resolved deviations occur on both sides of the reference. The RY–RZ–CZ family develops a persistent one-body deficit, with mean ρ₁ ≃ 0.874 at n = 14 and 0.877 at n = 16. The full computational-basis Fisher fraction remains order one, so the deficit is specific to the low-weight projection rather than to disappearance of measured tangent signal; the paper assigns no microscopic gate-level cause.

• Equal rank does not imply equal accessibility (Haar-U(4), n = 12). The physical one-body readout retains R_phys = 2.71×10⁻³, a rank-matched cross-fitted leading tangent subspace retains R_xfit = 3.06×10⁻², and the same-sample Ky Fan optimum retains 9.15×10⁻².

• The U(1) family behaves differently at the same size. At n = 12, the physical one-body readout retains 0.292, the cross-fitted aligned subspace 0.375, and the same-sample Ky Fan subspace 0.444 — the physical readout is not optimal but is already much closer to the leading tangent subspace than in generic Haar-U(4).

• Operational gains at fixed rank and shot budget (Haar-U(4)). Replacing the physical one-body readout with the cross-fitted aligned subspace increases the mean directional gradient-energy proxy by factors 2.963, 4.639, and 9.584 for n = 8, 10, 12, with corresponding finite-shot SNR gains of 1.734, 2.125, and 3.111. A random rank-matched subspace remains near the physical generic baseline.

• Smaller gains in the U(1) family. Aligned-to-physical gradient-energy ratios are approximately 1.148, 1.189, and 1.210 at n = 8, 10, 12, with SNR gains 1.068, 1.097, and 1.112.

• Persistent low-weight alignment in the half-filled U(1) family. At n = 18, one-body retention is approximately 0.215 versus 6.38×10⁻⁵ for Haar-U(4); two-body values are approximately 0.401 versus 6.25×10⁻⁴. Against a sector-corrected random-orientation null with means 0.00035 (k = 1) and 0.00313 (k ≤ 2), the observed subspace overlaps are approximately 811× and 148× larger.

• Finite-size scaling favors a power form over an exponential one, descriptively. Over n = 8–18, the cross-fitted one-body alignment A₁ takes values 0.552, 0.450, 0.396, 0.341, 0.309, and 0.284; A₁(18) = 0.283586 [0.277044, 0.290865] and A_≤2(18) = 0.463538 [0.457425, 0.469591], each with 20 independent circuit instances. Power-law exponents are 0.8223 [0.7694, 0.8735] for A₁ and 0.7041 [0.6798, 0.7281] for A_≤2, with ΔAICc = 14.39 and 12.82 in favor of the power model. The author states these are finite-size model-discrimination results only, not asymptotic bounds or a hydrodynamic exponent.

• Symmetry-breaking sensitivity control at n = 8. Unperturbed one-body alignment A₁ = 0.517 [0.486, 0.549]; a symmetry-preserving R_Z perturbation of strength ε = 0.3 leaves A₁ = 0.481 [0.454, 0.518]; a nontrainable symmetry-breaking R_X perturbation at the same strength gives A₁ = 0.0295 [0.0248, 0.0338], while physical one-body retention falls from 0.430 to 0.0325. Error bars are 95% circuit-bootstrap intervals over 10 circuit instances; the author treats this as a sensitivity control, not mechanism identification.

• The full measurement record is not the bottleneck. Across the generic data, F_full/F_Q is approximately 0.49, while U(1) values remain near 0.47–0.48. The orders-of-magnitude separation between the two circuit classes appears primarily after projection onto the low-weight readout span.

• Scope limits stated explicitly. The paper does not claim a barren-plateau theorem for supervised losses, does not claim U(1) symmetry generically improves trainability, and does not infer a hydrodynamic exponent from finite-size data. Task-level trainability is said to additionally depend on encoding, labels, classical postprocessing, loss function, and optimization trajectory.

Methodology in Plain English

The author starts from a fixed computational-basis measurement on a parametrized quantum state. Varying the parameters produces small changes in the outcome probabilities; after removing the constant direction and normalizing each direction to unit length, these become "score vectors" in an N-dimensional real space. Averaging their outer products gives a covariance matrix C that says which score-space directions carry tangent variation and how strongly.

A restricted readout — for example, all diagonal Pauli-Z strings up to some weight — is a subspace, represented by a projector P of rank r. The average fraction of normalized score energy the readout keeps is exactly Tr(PC). To get a fair baseline, the author keeps C fixed and only randomizes the orientation of P (a Haar-uniform point on the Grassmann manifold). This gives mean r/N and an exactly computable variance involving N, r, and Tr(C²).

The numerics use computational-basis measurement, depth proportional to system size, and diagonal Z readouts through weight one and two, across RY–RZ–CZ, SU2–CNOT, SU2–CZ, random-matching SU2–CZ, and Haar-U(4) brickwork families, plus a half-filled number-conserving RZ–XY family. The circuit instance, not the individual tangent direction, is the independent statistical unit, with nonparametric bootstrap over instances, and the generic aggregate is family-balanced so no ansatz dominates by sample count.

For the operational test, three rank-matched readouts are compared with circuit, full measurement record, rank, and evaluation shot budget all held fixed: the physical low-weight span, an independent random subspace, and a cross-fitted aligned subspace. Cross-fitting means the leading eigenvectors are estimated from one set of tangents and retention is evaluated on a separate set, preventing the same samples from both selecting and evaluating the subspace. Finite-shot signal-to-noise is computed under a multinomial shot model at a fixed budget of 10⁴ shots. The appendix's frozen operational profile uses 128 alignment tangents and 128 independent evaluation tangents. The independent alignment sample is described as a separate calibration resource, not counted as free measurement cost in the fixed-evaluation-budget comparison. Code, frozen profiles, aggregate tables, shard-level outputs, and result summaries are archived in a public reproducibility repository.

Why This Matters

Impact on research. The paper supplies a clean three-way factorization of a question that is often conflated: rank sets a null accessibility scale, the covariance spectrum sets how widely random rank-matched subspaces fluctuate, and the physical orientation of the readout relative to the tangent eigenspaces sets the actual retained mass. This gives a diagnostic that distinguishes "there is no tangent information at this rank" from "the readout is pointed the wrong way" — a distinction with a measurable operational signature at fixed shot budget.

Real-world applications (as the framing implies):

• Readout design for variational quantum algorithms — choosing which measured features to feed a classical optimizer, at fixed observable weight and shot cost.
• Benchmarking hardware-efficient versus symmetry-aware ansätze — the A_k subspace-overlap statistic with a sector-corrected null gives a symmetry-respecting baseline for such comparisons.
• Classical shadows and randomized-measurement protocols — quantifying what a fixed measurement already exposes versus what a restricted classical postprocessing layer discards.
• Diagnostics for quantum Fisher information budgets — separating loss of measured information from loss at the readout projection step.

Industry relevance. For near-term quantum computing, hardware constraints typically force low-weight, diagonal readouts. The paper's central practical message — that replacing the physical low-weight readout with a learned same-rank subspace raised the gradient-energy proxy by up to a factor 9.584 and SNR by up to 3.111 in Haar-U(4) at n = 12 — targets exactly that constraint, without adding observables or shots.

Future Directions

• Identify the dominant mechanism in the U(1) case. The author lists symmetry-restricted controllability, fixed-charge geometry, readout–charge compatibility, and slow operator structure under conservation laws as plausible but non-equivalent explanations, and states that distinguishing them requires interventions such as varying locality independently of conservation, comparing against fixed-charge random baselines, or resolving depth scaling.

• Bridge measurement accessibility to task usefulness. The framework currently stops before the loss function; a supervised problem's task gradient may occupy only a subset of the accessible tangent geometry, and that additional layer is left for separate work.

• Separate the four candidate explanations via controlled interventions. The n = 8 symmetry-breaking pilot is explicitly labeled a sensitivity control that does not separate the candidate mechanisms — a natural template for a follow-up mechanism-identification experiment.

• Extend beyond the tested finite-size window. The power-versus-exponential comparison is descriptive over n = 8–18; the author declines to treat the fitted exponents as asymptotic bounds or as a hydrodynamic exponent, leaving the scaling question open.

Target Audience

Quantum information researchers working on variational quantum circuits, trainability, and measurement-induced geometry; quantum machine learning practitioners who design readout and feature layers; and mathematically inclined readers comfortable with Grassmann-manifold moments, spectral bounds, and bootstrap statistics. Readers seeking a barren-plateau theorem or a task-level trainability guarantee should note that the paper explicitly disclaims both — it is a controlled statement about measurement-accessible tangent geometry.

Authors’ abstract

A fixed quantum measurement can expose substantially more tangent information than a restricted observable readout retains. We study this second restriction. For a normalized covariance $C \succeq 0$, $\mathrm{Tr} C = 1$, of measurement-induced tangent scores in an $N$-dimensional centered score space, and a rank-$r$ readout projector $P$, we quantify retained tangent mass by $R=\mathrm{Tr}(PC)$. The ratio $ρ=R/(r/N)$ separates the actual retained mass from a rank-only random-orientation reference. Standard Grassmann averaging gives $\mathbb{E}ρ=1$ and $\mathrm{Var}(ρ) \le 2/(r d_{\mathrm{eff}})$, where $d_{\mathrm{eff}}=1/\mathrm{Tr}(C^2)$. We use this identity as a null model rather than as a new random-projection theorem. Numerically, family-balanced one- and two-body readouts remain close to the rank reference through $n=16$ even as the tangent covariance becomes strongly anisotropic. The decisive equal-rank comparison holds the circuit, measurement record, readout rank, and evaluation shot budget fixed. For Haar-$U(4)$ at $n=12$, cross-fitted alignment increases the mean directional gradient-energy proxy by a factor 9.584 and the finite-shot signal-to-noise ratio by a factor 3.111 relative to the physical one-body readout, while a random rank-matched subspace remains near the rank baseline. A half-filled $U(1)$-conserving family provides a structured counterexample to generic orientation: physical low-weight $Z$ readouts are already strongly aligned with leading tangent directions over the tested finite-size range. We treat this symmetry result as a case study, not as a claim that $U(1)$ symmetry generically prevents barren plateaus or that hydrodynamics is the established mechanism. The results isolate readout orientation as a degree of freedom invisible to rank alone that directly controls how much measured tangent information remains usable after readout restriction.