PaperlayerView exact revision on arXiv

Source-linked AI summary

Exact Posterior Prediction from Product Haar Measurements and a Randomized-Mesh Maximum-Likelihood Bridge

Naqueeb Ahmad Warsi, Ayanava Dasgupta, Masahito Hayashi

arXiv:2610.00117v1quant-phcs.ITmath.IT

TL;DR

Predicting an unmeasured copy after independent one-copy Haar measurements requires a density-matrix decision rule for the fixed outcomes. The paper evaluates that rule exactly and analyzes its finite-sample and fixed-dimensional asymptotic performance. A randomized fixed-mesh bridge yields sharper asymptotic bounds, but not a uniform linear-in-dimension finite-sample guarantee.

  • Problem

    The paper studies how to predict an unmeasured copy of an unknown finite-dimensional pure state after independently measuring the observed copies with the one-copy Haar POVM.

    Source: §I · p. 1

  • Method

    It evaluates the fixed-outcome posterior predictive state through Gram-matrix minor permanents and uses a randomized fixed mesh with a selected finite-alphabet MLE to analyze asymptotic performance.

    Source: §I · p. 2, E. Gram matrices and the marginal… · p. 5, §V · p. 8

  • Results

    O((d^2/epsilon) log(d/epsilon)) is a uniform finite-sample guarantee; iterated fixed-mesh limits yield the sharper fixed-dimensional scale (d/epsilon) log(d/epsilon), with a_k approaching d−1.

    Source: §VIII · p. 10

  • Takeaways & Limitations

    The exact posterior is full rank and minimax among decision rules based on the fixed product-Haar outcomes, while collective benchmarks provide entropy and posterior-purity comparisons.

    Source: §I · p. 2, §VIII · p. 10

    The sharper linear-in-dimension scale holds only after fixing the mesh and taking the large-sample limit before refinement; it is not a uniform finite-sample guarantee.

    Source: D. Finite-sample count bound · p. 7, §VIII · p. 10

Abstract

from arXiv · show

We study prediction of one unmeasured copy of an unknown finite-dimensional pure quantum state after independently measuring the observed copies with the one-copy Haar POVM. For the resulting fixed separable observation, the Bayes predictive state under quantum relative-entropy loss is the full-rank posterior mean. We evaluate it exactly through permanents of minors of the outcome Gram matrix, prove positivity and normalization of the permanent representation, and establish minimaxity among decision rules based on these fixed outcomes. Using Hilbert--Schmidt projection of the posterior mean and an exactly analyzed linear-inversion competitor, we prove a finite-sample posterior-purity bound and an explicit relative-entropy guarantee with sample count $O((d^2/\epsilon)\log(d/\epsilon))$. Using the exact collective benchmarks proved in a companion paper, we also obtain entropy and posterior-purity comparisons. To connect the fine product data to a regular ray-valued experiment without assuming deterministic finite-sample uniqueness, we draw one random orientation for a finite projective mesh, retain it for the entire sample, and apply a selected finite-alphabet MLE in the resulting reference experiment. A nested recovering mesh uniformly recovers the fine Fisher information. Consequently every sufficiently fine fixed mesh gives a regular finite-alphabet model. Its selected MLE has Haar-averaged inverse-Fisher coefficient $\overline a_k$, and refining the mesh after the fixed-mesh large-sample limit makes $\overline a_k$ approach $d-1$. This yields epsilon-optimal leading relative-entropy and posterior-overlap bounds for the fine posterior and the sharper fixed-dimensional asymptotic scale $(d/\epsilon)\log(d/\epsilon)$. The latter is not asserted as a uniform finite-sample guarantee. No deterministic finite-sample MLE uniqueness or triangular mesh schedule is used.

I. INTRODUCTION

The paper studies prediction of an unmeasured copy after product Haar measurements, deriving the exact fixed-outcome Bayes state and connecting its performance to finite-sample bounds and iterated mesh asymptotics.

I. INTRODUCTION

  • The exact product-Haar posterior predictive state is full rank, uses Gram-minor permanents plus a universal identity term, and is minimax among rules on the fixed outcomes.The general principle that the posterior mean is the Bayes predictive state is known; the contribution is its exact evaluation for product Haar data.Source: §I · pp. 1–2
  • O((d^2/epsilon) log(d/epsilon)) samples suffice for a uniform finite-sample relative-entropy guarantee, independent of the randomized-mesh asymptotics.The bound follows from Hilbert–Schmidt projection of the posterior mean and the exactly computed variance of a linear-inversion competitor.Source: §I · p. 2
  • At fixed dimension, a selected MLE on a fixed randomized mesh has coefficient a_k, and refining the mesh after the large-sample limit makes a_k approach d−1.One random orientation is retained for the entire sample; this iterated-limit result is not a uniform finite-sample guarantee or a triangular mesh schedule.Source: §I · p. 2
  • The local posterior has the same leading entropy and purity coefficients as exact collective benchmarks, despite different measurements and operational optimizations.The collective benchmarks are used as certified comparisons; the present paper fixes independent one-copy Haar measurements.Source: §I · p. 2

II. POSTERIOR PREDICTION FOR A FIXED OBSERVATION

  • For a fixed observation, the posterior mean is the Bayes act under quantum relative-entropy loss, with optimized conditional loss equal to its entropy.This fixed-observation result does not optimize the measurement.Source: §II · p. 2
  • Coarse-graining raises posterior entropy and lowers posterior purity; for a covariant ray-valued observation, these are determined by its infidelity.The posterior states satisfy ρY = E[ρX | Y], and the covariant ray-valued coarse posterior has entropy h_d(r) and purity p_d(r).Source: §II · pp. 2–3

III. POSTERIOR PREDICTIVE STATE FOR THE LOCAL HAAR-POVM MODEL

For fixed product-Haar outcomes, the section evaluates the posterior predictive state without optimizing the measurement. Its calculation reduces the posterior integrals to symmetric-subspace operations and uses a SWAP decomposition to evaluate the numerator.

A. The one-copy and product Haar-POVM setting

B. Posterior predictive state for product Haar-POVM data

C. Evaluating the denominator and Haar integrals

D. Evaluating the numerator via SWAP decomposition

E. Gram matrices and the marginal permanent

The section expresses the outcome-dependent marginal through Gram-matrix permanents, then normalizes and verifies it as a valid state for the exact posterior rule.

E. Gram matrices and the marginal permanent

F. Bayesian prediction and decision-rule minimax optimality for product Haar-POVM data

For fixed product Haar-POVM outcomes, the unique Bayes predictive state is the posterior mean in identity-regularized form. It is also minimax among decision rules based on those outcomes, not among measurement choices.

F. Bayesian prediction and decision-rule minimax optimality for product Haar-POVM data

  • The minimax claim optimizes the decision rule after the product Haar-POVM is fixed; it does not optimize over collective measurements or other POVMs.The comparison class maps the observed outcome sequence to a density matrix.Source: F. Bayesian prediction and… · p. 6
  • The unique Bayes act for quantum relative-entropy loss is the posterior mean, with an equivalent identity-regularized posterior form.The explicit evaluation applies to the fixed product Haar-POVM observation.Source: F. Bayesian prediction and… · p. 6
  • The local posterior predictive state is minimax within the class of decision rules based on product Haar-POVM outcomes.Covariance makes its frequentist risk constant over the unitary orbit, so its Haar-Bayes and worst-case risks coincide.Source: F. Bayesian prediction and… · p. 6

IV. FINITE-SAMPLE RISK AND SAMPLE COMPLEXITY

The section bounds finite-sample posterior overlap by comparing the posterior mean with a matrix-valued linear-inversion estimator. The posterior mean also minimizes Bayesian Hilbert–Schmidt risk over all square-integrable Hermitian estimators.

A. Hilbert–Schmidt projection and overlap

  • The posterior mean minimizes Bayesian Hilbert–Schmidt risk over all square-integrable Hermitian matrix-valued estimators measurable from the observed data.The comparison class is not restricted to density matrices.Source: A. Hilbert–Schmidt projection and… · p. 6

B. Empirical linear inversion and its exact variance

  • The linear-inversion estimator is used as a matrix-valued competitor, and its stated exact numerator is d^2 + d − 2.Its squared one-copy estimator has trace d^2 + d − 1.Source: B. Empirical linear inversion and its… · p. 7
  • Theorem 10 gives a finite-sample local posterior-overlap bound for every n ≥ 1 and d ≥ 2.When its displayed right-hand side is negative, the bound is weaker than the trivial posterior-overlap bound.Source: B. Empirical linear inversion and its… · p. 7

C. Explicit relative-entropy bound

The paper gives finite-sample relative-entropy and sample-count guarantees, then uses a shared-orientation finite mesh to connect the fine observation to a regular finite-alphabet model. The mesh argument takes the large-sample limit at fixed mesh before refinement.

C. Explicit relative-entropy bound

  • Theorem 11 establishes a finite-sample relative-entropy risk bound using the posterior's spectral floor and a convexity bound for −log x.The posterior has spectral floor I/(n + d).Source: C. Explicit relative-entropy bound · p. 7

D. Finite-sample count bound

  • The finite-sample count guarantee has quadratic dependence on d because it compares the posterior mean with the simple linear-inversion estimator.The sharper fixed-dimensional linear-d scale is asymptotic, not a uniform finite-sample theorem.Source: D. Finite-sample count bound · p. 7

V. RECOVERING FINITE MESHES AND THE SHARED-ORIENTATION EXPERIMENT

  • A sufficiently fine fixed mesh restores common positive support and regular finite-alphabet likelihood conditions, while nested refinement uniformly recovers the fine Fisher information.The recovering sequence preserves global identifiability through refinement.Source: §V · pp. 7–8
  • The asymptotic argument fixes a sufficiently fine mesh, lets n tend to infinity, and only then refines the mesh; it uses no triangular mesh schedule.Finite-sample MLE uniqueness is not assumed.Source: §V · p. 8
  • The construction draws one Haar orientation before sampling and retains it for all copies, producing a covariant ray-valued coarse observation.Its role is an auxiliary comparison device; its coarse posterior bounds the fine posterior's entropy from above and purity from below.Source: §V · p. 8

VI. FIXED-MESH EFFICIENCY AND HAAR-AVERAGED RISK

For every sufficiently fine fixed mesh, the model is regular enough for finite-alphabet MLE theory. Its risk bounds and the shared random orientation transfer Haar-averaged reference risk to pointwise physical risk.

VI. FIXED-MESH EFFICIENCY AND HAAR-AVERAGED RISK

  • A sufficiently fine fixed-mesh model is compact, smooth, identifiable, full-support, and has nondegenerate Fisher information.Positive cell probabilities, global identifiability, and uniform Fisher recovery supply the stated regularity conditions.Source: §VI · pp. 8–9
  • For a fixed mesh, the selected global MLE has mean squared geodesic error O_k(n^-1), with an exponential localization bound outside any fixed neighborhood.The argument uses a domination estimate, not a parameter-uniform risk expansion.Source: §VI · p. 9
  • The shared Haar orientation converts Haar-averaged reference risk into pointwise physical risk for every fixed physical ray.The reference parameter is Haar distributed under the shared orientation, and unitary invariance supplies exact physical constancy.Source: §VI · p. 9

VII. POSTERIOR COMPARISON AND SHARP ASYMPTOTICS

Finite-mesh MLEs and posterior coarse-graining connect the fine product-Haar data to asymptotic comparisons for posterior entropy and purity. The section distinguishes these fixed-dimensional asymptotics from its uniform finite-sample guarantees.

B. Covariant coarse-graining of the posterior

  • Coarse-graining discards information, so posterior entropy can only increase and posterior purity can only decrease; the tower identities relate the coarse and fine posteriors.The MLE controls only the posterior conditioned on the reported ray, which bounds the posterior conditioned on the complete product-Haar data.Source: B. Covariant coarse-graining of the… · p. 9

D. Finite-sample numerical illustration

  • The finite-sample illustration compares the product-Haar posterior with the unrestricted collective benchmark and is consistent with their exact ordering.The Monte Carlo illustration concerns the exact local posterior, not the randomized-mesh MLE, and is not evidence that an asymptotic regime has been reached.Source: D. Finite-sample numerical illustration · p. 10

VIII. DISCUSSION

  • For fixed product observations, the paper gives an exact full-rank posterior predictive state with Gram-minor-permanent coefficients and decision-rule minimaxity within that experiment.The unrestricted collective optimization is handled separately in the companion paper and enters through its exact endpoints.Source: §VIII · p. 10
  • A shared random orientation lets one selected finite-alphabet MLE yield an exactly covariant physical risk, avoiding parameter-uniform risk expansions and control of every finite-sample argmax.Uniform Fisher recovery makes every sufficiently fine fixed mesh a regular finite-alphabet model.Source: §VIII · p. 10
  • The uniform finite-sample guarantee has scale O((d^2/ε) log(d/ε)), while the sharper fixed-dimensional asymptotic scale is (d/ε) log(d/ε), with no triangular mesh schedule.The asymptotic coefficient approaches d−1 only after fixing the mesh and taking the large-sample limit.Source: §VIII · p. 10

C. Sharp asymptotics through iterated limits

  • After fixing a sufficiently fine mesh and taking n → ∞, refining the mesh makes the inverse-Fisher coefficient approach d−1 and yields the sharper fixed-dimensional asymptotic scale.The exact collective endpoints provide the reverse liminf inequalities; no triangular schedule is used.Source: §VIII · p. 10, C. Sharp asymptotics through iterated… · p. 10

APPENDIX A CONSTRUCTION AND FISHER RECOVERY OF THE MESHES

The appendix constructs nested finite partitions that preserve informational completeness while their cells shrink, then proves that the associated meshes recover the fine Fisher information uniformly.

APPENDIX A CONSTRUCTION AND FISHER RECOVERY OF THE MESHES

  • A nested recovering mesh sequence exists with positive-measure cells, vanishing maximal diameter, globally identifiable finite-alphabet models, and uniform Fisher-information recovery.The proof builds an informationally complete initial partition, refines it while preserving spanning, and verifies Fisher recovery.Source: Appendix A · p. 10

A. An informationally complete initial partition

  • An informationally complete initial partition is formed so its cell operators span the Hermitian matrices.Its cell operators contain a spanning subfamily; rank-one projectors onto basis rays and real and imaginary superpositions supply diagonal and off-diagonal Hermitian components.Source: A. An informationally complete initial… · pp. 10–11

B. Nested refinement and identifiability

  • Nested refinements can have vanishing cell diameter while retaining positive measure and measure-zero boundaries, and refinement preserves the span of the cell operators.The construction permits omitted intersections only on a set of invariant measure zero.Source: B. Nested refinement and identifiability · pp. 11–12
  • When cell operators span the Hermitian matrices, the finite-alphabet parameter map is globally identifiable.Equality of measurement probabilities then forces the corresponding rank-one projectors, and hence the rays, to agree.Source: B. Nested refinement and identifiability · p. 12

C. Fine and mesh directional scores

Each finite mesh replaces the fine directional score by its conditional expectation on the containing cell. As nested meshes refine, their Fisher information recovers the fine information pointwise.

C. Fine and mesh directional scores

  • The mesh score construction handles the fine density’s parameter-dependent zero set by using a bounded density derivative for cell-probability differentiation.The logarithmic score may be singular on the zero hyperplane, which has zero probability under the model.Source: C. Fine and mesh directional scores · p. 12
  • The finite-mesh score is the conditional expectation of the fine directional score on its cell, and every mesh-cell probability is strictly positive.Differentiation under the cell integral is justified by the bounded derivative of the fine density.Source: C. Fine and mesh directional scores · p. 12

D. Pointwise Fisher recovery

  • Nested refinements monotonically restore Fisher information because conditional expectation is an L2 projection onto increasingly informative sigma-fields.The fine directional score is square-integrable, and the limiting sigma-field contains the fine observation, yielding pointwise recovery.Source: D. Pointwise Fisher recovery · p. 13

E. Uniform recovery on the unit tangent bundle

Compactness and Dini’s theorem upgrade pointwise Fisher recovery to uniform recovery on the unit tangent bundle. The appendix also verifies the trace-SWAP identity by reducing it to simple tensor operators.

E. Uniform recovery on the unit tangent bundle

  • The mesh Fisher quadratic forms are continuous and increase pointwise to the constant 2 on the compact unit tangent bundle, so Dini’s theorem gives uniform Fisher recovery.This establishes the uniform recovery condition used for sufficiently fine meshes.Source: E. Uniform recovery on the unit… · p. 13

APPENDIX B PROOF OF THE TRACE-SWAP IDENTITY

  • For a simple tensor operator, tracing the swapped pair reduces the factor to Tr Ak; linearity extends the trace-SWAP identity to every operator on H^⊗n.The swap acts between the kth tensor factor and the additional factor.Source: Appendix B · p. 13

APPENDIX C PROOF OF FACT 2

Appendix C proves Fact 2 by reducing the partial trace to a basis-index sum that replaces one tensor slot by the target vector. Appendix D proves Lemma 3 by expanding matrix elements and using a permutation bijection for the omitted row and target column.

APPENDIX C PROOF OF FACT 2

  • The partial-trace proof inserts a tensor-product basis, swaps the k-th entry with x, and resolves the k-th slot to replace it by y.This yields the matrix element with x in the bra and y in the ket at that slot.Source: Appendix C · p. 13

APPENDIX D PROOF OF LEMMA 3

  • The Lemma 3 proof expands the symmetrized matrix-element inner products and identifies the omitted row k and target column j = π(k).Source: Appendix D · p. 14
  • For fixed j = π(k), the restricted permutation is a bijection between the remaining row and column indices; the proof then concludes the identity for all x and y.Source: Appendix D · p. 14
Loading 2610.00117v1…