PaperlayerView exact revision on arXiv

Source-linked AI summary

Optimal Low-Rank Quantum State Tomography with Bounded-Sample Joint Measurements

Ashwin Nayak, Xingyu Zhou

arXiv:2609.10514v1quant-phcs.DScs.ITcs.LG

TL;DR

The paper asks how efficiently bounded-rank quantum states can be learned when joint measurements are limited by experimental resources and each measurement acts on at most t samples. It characterizes the optimal rank-dependent rate for adaptive protocols and matches it with a nonadaptive upper bound, showing when joint measurements reach unrestricted performance.

  • Problem

    The paper studies whether mixed quantum states with bounded rank can be learned using fewer samples than Θ(d^2), while accounting for the practical resource demands of joint measurements.

  • Method

    The lower bound controls Fisher information for joint measurements on t samples, extends it across adaptive transcripts using a conditional score chain rule, and converts it to a trace-norm bound via the van Trees inequality.

  • Results

    Jointly measuring t samples improves optimal sample complexity over single-sample tomography by a factor of order √t until t reaches order r^2, where the unrestricted collective rate is attained; a nonadaptive protocol matches the lower bound up to constants.

  • Takeaways & Limitations

    Measuring order r^2 samples jointly is necessary and sufficient for unrestricted optimal sample complexity, and classical adaptivity provides no further improvement beyond constant factors.

  • Takeaways & Limitations

    The results concern protocols retaining only classical information between rounds; persistent quantum memory changes the relevant resource and leaves broader tradeoffs open.

Abstract

from arXiv · show

We determine the optimal sample complexity of low-rank quantum state tomography when each measurement may act jointly on at most $t$ samples. For sufficiently small $\varepsilon$, estimating an unknown state on $\mathbb{C}^d$ of rank at most $r$ to trace norm error $\varepsilon$ with constant success probability requires, and is achievable with, $$ Θ\left( \frac{dr}{\varepsilon^2} \max\left\{1,\frac r{\sqrt t}\right\} \right)$$ samples. The lower bound allows the protocol to choose each joint measurement adaptively using all previous classical outcomes; the matching upper bound is nonadaptive. Thus joint measurements on at most $t$ samples improve the complexity of algorithms making single-sample measurements by at most a factor $\sqrt t$. Further, measuring order $r^2$ samples jointly is necessary and sufficient to attain the unrestricted collective rate. For the lower bound, we vary the support of a state with fixed uniform spectrum and bound the Fisher information trace of every joint measurement on $t$ samples. The adaptive Fisher chain rule and the van Trees inequality then give the trace norm lower bound. For the upper bound, we construct and analyze a nonadaptive tomography protocol based on a Gaussian joint measurement. An explicit second moment identity and a conditional Gaussian law outside the state's support give a rank-dependent error analysis, yielding the matching rate.

1 Introduction

This work characterizes low-rank tomography when measurements act jointly on at most t samples, covering adaptive and nonadaptive strategies. It establishes matching bounds across the interpolation from single-sample to unrestricted collective measurements.

  • 1 Introduction: Jointly measuring t samples improves the single-sample rate by at most a factor of order √t.For t = 1, arbitrary classical adaptivity does not improve the optimal nonadaptive rate.
  • 1 Introduction: Measuring order r^2 samples jointly is necessary and sufficient to attain the unrestricted collective sample complexity O(dr/ε^2).The rate reaches O(dr/ε^2) once t reaches order r^2.
  • 1 Introduction: The result gives the first matching rank-dependent characterization for arbitrary adaptive joint measurements on at most t samples.It resolves rank-dependent adaptive questions at t = 1 and the interpolation for intermediate r and t.
  • 1 Introduction: The lower bound combines Fisher-information trace bounds, an adaptive score chain rule, and the van Trees inequality.The construction varies the support while keeping the spectrum uniformly fixed, then converts the information bound into trace-norm error.
  • 1 Introduction: The upper bound uses a nonadaptive Gaussian joint measurement, second-moment control, and projection onto rank-at-most-r density matrices.The analysis separately controls errors on the support and outside it.

2 Proof overview

The lower bound parameterizes rank-r states by rotations of a fixed support, bounds transcript Fisher information under adaptive bounded-sample measurements, and applies van Trees to obtain trace-norm loss. The upper bound uses a nonadaptive Gaussian joint measurement whose moment and tail estimates yield the claimed error rate.

  • Lower bound: The lower-bound proof uses van Trees to convert bounded transcript Fisher information into a tomography lower bound.The argument first controls Fisher information and then translates parameter-estimation error into trace norm.
  • Lower bound: Operator-norm localization converts van Trees’ Frobenius-loss bound into expected trace-norm loss, producing the lower-bound contradiction for accurate tomography.The construction bounds the prior Fisher information and relates matrix estimation error to trace norm.
  • Lower bound: The hard family fixes the uniform spectrum and varies only the r-dimensional support through a small matrix rotation.The rotation matrix has 2r(d-r) real parameters and locally describes nearby support subspaces.
  • Lower bound: For every joint POVM on t samples, the Fisher-information estimate applies at each adaptive decision-tree node with the node’s sample count.The adaptive chain rule sums the information of the fixed POVMs selected along the realized transcript.
  • Upper bound: The upper-bound protocol uses the same Gaussian joint POVM in every repetition and is therefore nonadaptive.Its estimator is an unbiased matrix average followed by projection onto rank-at-most-r density matrices.
  • Upper bound: The Gaussian measurement’s second-moment and tail estimates control support and off-support errors, while Eρ[J] = O(√s) yields the joint-measurement gain.The resulting bounds show that increasing the joint block size beyond the stated scale changes the rate by at most a constant factor.

3 Preliminaries

This section establishes notation for finite-dimensional quantum states, measurements, matrix spaces, and the classical and quantum Fisher information used later. It also specifies the classical-memory restriction on adaptive protocols.

  • A density matrix is a positive semidefinite operator with trace one, and its support is the image of that operator.
  • A POVM is a countably additive map from measurable outcome events to positive semidefinite operators, normalized by the identity.
  • Classical Fisher information: The classical Fisher information matrix is derived from the likelihood gradient of the measurement outcome distribution.
  • Quantum Fisher information: The quantum Fisher information depends directly on the parameterized state family, is measurement-independent, and upper-bounds information obtainable from any parameter-independent POVM.
  • Adaptive measurement protocols: The theorem restricts protocols to classical information between rounds, excluding persistent quantum memory that could combine rounds into a larger coherent measurement.

4 From Fisher information bounds to the adaptive lower bound

The lower-bound proof proceeds from a hard local family through Fisher-information control, adaptive transcript accumulation, and a van Trees conversion to trace-norm risk.

  • The proof constructs a hard parameterized family, bounds Fisher information for t-sample measurements, and accumulates the bounds along adaptive transcripts.
  • The van Trees inequality converts the accumulated Fisher-information bound into a lower bound on expected trace-norm loss.

4.1 The hard family and its local coordinates

The hard family fixes a uniform rank-r spectrum and varies only the support through nearby subspace rotations. The resulting local coordinates have 2r(d −r) real dimensions and are controlled in trace norm.

  • The hard states have rank r, are maximally mixed on their supports, and vary only through support rotations with fixed eigenvalues.
  • The construction uses an orthogonal decomposition C^d = S ⊕ S⊥ with dim S = r and dim S⊥ = d −r.
  • Small-operator-norm matrices X parameterize r-dimensional supports near the reference support S, with orthonormal basis columns supplied by V_X.
  • The real and imaginary parts of X ∈ C^(d−r)×r provide 2r(d −r) real tangent coordinates at X = 0.
  • For ∥X∥op, ∥Y∥op ≤ 1/4, support-projector trace distance controls parameter distance through ∥X −Y∥1 ≤ 2∥Π_X −Π_Y∥1 = 2r∥ρ_X −ρ_Y∥1.

4.2 Fisher information trace for joint measurements

This section bounds the Fisher information trace available from arbitrary joint POVMs on t copies of the hard family. The proof analyzes tangent directions through zero- and one-excitation subspaces and obtains different regimes depending on t relative to r^2.

  • The main proposition bounds the classical Fisher-information trace for every joint POVM on t samples and every support parameter X.
  • The argument controls cross terms among tensor positions by reducing them to occupation-number operators and bounding their quadratic forms uniformly over directions.
  • The tangent directions connect the zero-excitation subspace only to one-excitation subspaces, enabling a blockwise analysis of joint measurements.

4.3 Adaptive accumulation across measurement rounds

The adaptive protocol is represented as a decision tree whose transcript Fisher information equals the accumulated conditional Fisher information across measurement rounds. A pathwise sample bound then yields a global information bound.

  • Pathwise bound: The adaptive Fisher information bound applies to protocols whose history-dependent POVM acts jointly on at most t fresh samples per round.The resulting bound controls how much the full transcript can reveal about the support parameter.
  • Transcript representation: Each transcript is padded to N rounds, with histories containing the private seed and previous outcomes; dummy rounds use no samples or Fisher information.This makes all transcripts have a common length without changing the protocol.
  • Fisher accumulation: The adaptive Fisher chain rule decomposes transcript Fisher information into the expected sum of roundwise conditional Fisher information matrices.Both transcript and roundwise matrices use the same 2mr real support coordinates.
  • Fisher accumulation: Cross terms vanish because each conditional score has mean zero given the preceding history, while diagonal terms reduce by total expectation.The parameter-independent seed contributes no Fisher information.

4.4 From a Fisher information bound to expected trace norm loss

The Fisher bound is converted into trace-norm risk using van Trees, a smooth prior over bounded support parameters, and an operator-norm constraint on the estimator. The prior has controlled Fisher information, enabling a quantitative lower bound on estimation loss.

  • Risk conversion: Van Trees converts limited transcript Fisher information into a lower bound on expected squared Frobenius error for support-parameter estimation.Restricting both parameters and estimators in operator norm converts this bound into trace-norm loss.
  • Smooth prior: The parameter dimension is p = 2mr, and the prior is supported where ∥X∥op < a with Fisher information bounded by a universal constant times pd/a^2.Here m = d − r and the prior is constructed on the rectangular support parameter X.
  • Smooth prior: The density π = ψ^2 is normalized, infinitely differentiable, compactly supported inside {X : ∥X∥op < a}, and has controlled Fisher information.The construction uses a cutoff and Gaussian smoothing to obtain the required regularity.
  • Smooth prior: The smooth prior is built by multiplying a Gaussian density by a cutoff that vanishes before ∥X∥op reaches a, while keeping the cutoff gradient O(1/a).This preserves compact support and keeps the prior Fisher information controlled.
  • Risk conversion: For bounded parameter and estimator operator norms, singular-value inequalities relate trace norm to Frobenius error, and expectation over the observation then applies van Trees.This establishes the trace-norm consequence used in the lower-bound argument.

4.5 Completion of the lower bound

The lower-bound proof amplifies confidence, postprocesses tomography outputs into bounded support estimates, and applies van Trees to the adaptive transcript. A depolarizing rank lift extends the result from ranks at most d/2 to all ranks.

  • Confidence amplification: Confidence amplification repeats the estimator and selects a representative output, converting success probability 2/3 within η into probability 1 − δ within 3η.Hoeffding’s inequality gives K = O(1 + log(1/δ)) repetitions.
  • Bounded support estimator: A measurable postprocessing maps an accurate density-matrix estimate to a support parameter in Xa while preserving a controlled estimation error.The construction uses compactness of Xa and continuity of X ↦ ρX.
  • Lower-bound contradiction: Applying van Trees to the amplified transcript shows that too few samples would force expected trace-norm loss of order ar, contradicting the estimator’s upper bound.The contradiction uses the adaptive Fisher bound, the smooth prior, and the trace-norm conversion.
  • Lower-bound contradiction: The lower-bound proposition applies to adaptive protocols using at most t fresh samples per round and establishes the required sample lower bound for r ≤ d/2.The argument compares the upper loss from successful tomography with the lower loss forced by insufficient Fisher information.
  • Rank lifting: Depolarizing rank lifting embeds lower-rank states into rank-exactly-r states, scales trace distances by one half, and preserves adaptive transcript laws and sample counts.Applying the adjoint channel to each POVM simulates lifted measurements on the original states.

5 Rank-sensitive tomography

The upper-bound protocol uses a nonadaptive Gaussian joint measurement on t samples, producing unbiased matrix estimates whose second moments support rank-sensitive error analysis. Averaging independent estimates and projecting onto rank at most r yields the output.

  • Measurement construction: The protocol constructs a POVM Mt on t copies with outcomes (J, G), where J ≤ min{d, t} and G is a complex Gaussian matrix.The POVM components are built from nested support projectors and pseudoinverse factors.
  • Rank sensitivity: For rank at most r, the expected outcome size satisfies a rank-dependent bound, so increasing joint-measurement size improves the rate only until t reaches r^2.The protocol therefore sets the joint measurement size according to the rank-sensitive regime.
  • Estimator construction: Independent applications of the measurement produce Hermitian estimates that are averaged and then projected in Frobenius norm onto density matrices of rank at most r.The same joint measurement is used in every repetition, making the complete protocol nonadaptive.
  • Estimator properties: The measurement outcome is converted into a Hermitian matrix estimator that is exactly unbiased for every state ρ.This unbiasedness is the starting point for the second-moment error analysis.
  • Estimator properties: The estimator has an exact uncentered tensor second moment involving ρ⊗2, the swap operator F, and the random outcome size J.The identity supplies the quantitative control used in the protocol analysis.
  • Rank-sensitive analysis: A conditional Gaussian law states that, outside the support of ρ, the projected Gaussian columns are independent standard complex Gaussian vectors independent of the supported columns.This structure enables the separate control of error outside the state’s support.

5.2 Properties of the joint measurement

The joint measurement is organized through nested support subspaces of Gaussian moment operators, yielding a valid POVM and an outcome index J controlled by the state rank. Antisymmetrizer identities establish the rank-dependent support and moment bounds used later in the estimator analysis.

  • Validity of the joint measurement: The nested projectors produce orthogonal differences Δℓ,t that sum to the identity, so the Gaussian densities define a valid POVM.The proof uses positivity and the spanning property of the largest support space.
  • Support of the Gaussian moment operators: The Gaussian moment operator Γℓ,t is supported on Vℓ,t, whose projector commutes with tensor permutations and A^⊗t.The support spaces are nested from the zero subspace to the full tensor-product space.
  • Moment and conditional-law properties: The measurement analysis derives Gaussian moment identities, bounds the tail and mean of J, and establishes a conditional Gaussian law outside the state support.Conditional on J, the orthogonal-complement Gaussian component remains independent with standard complex Gaussian columns.
  • Characterization by antisymmetrizers: Antisymmetrizers characterize the complements of the support spaces Vq−1,t through tensors antisymmetric on q selected positions.This converts the support question into a statement about rank-q minors and determinantal varieties.
  • Mean of J: If rank(ρ) ≤r, every outcome index J with positive probability is at most r.The support of ρ^⊗t lies in the r-dimensional support space, making the measurement components for ℓ>r vanish.

5.3 Analysis of the estimator

The estimator’s error is decomposed into supported, off-diagonal, and orthogonal-complement blocks, each controlled by second-moment or Gaussian concentration bounds. Choosing the joint-measurement size as s=min{t,r^2} yields the matching upper-bound rate, and larger blocks do not improve the asymptotics.

  • Blockwise error analysis: The error decomposition controls supported and off-diagonal blocks in Frobenius norm and the orthogonal-complement block in operator norm.A rank-constrained Frobenius projection combines these blockwise estimates into a trace-norm error bound.
  • Second-moment bounds: The estimator’s second moment identity bounds the supported and off-diagonal contributions after averaging independent measurement outcomes.The bounds are obtained by summing over Frobenius-orthonormal bases of the relevant Hermitian blocks.
  • Orthogonal-complement control: Conditional on the outcome indices, the orthogonal-complement Gaussian columns are independent standard complex Gaussian vectors, enabling operator-norm control.The resulting deviation is analyzed through a standard complex Gaussian matrix and its sample covariance.
  • Completion of the upper bound: s=min{t,r^2} explains the protocol choice because the first two error terms reach order dr/N at s=r^2.Increasing s beyond r^2 changes their sum by at most a constant factor.
  • Completion of the upper bound: Jointly measuring more than r^2 samples at a time does not improve the resulting asymptotic rate.The final sample-complexity bound follows by combining the block estimates and applying Markov’s inequality.

6 Discussion

The matching bounds show that rank determines how much bounded joint measurement helps: improvements continue until t reaches order r^2, where the unrestricted collective rate is attained. The discussion also identifies proof ingredients and scope boundaries involving continuous outcomes and quantum memory.

  • Discussion: Jointly measuring t samples improves the optimal sample complexity over single-sample tomography by a factor of order √t until t reaches order r^2.At that threshold, the sample complexity matches unrestricted collective measurements.
  • Discussion: The nonadaptive upper bound shows that classical adaptivity offers no further improvement beyond constant factors.The lower-bound framework nevertheless accommodates adaptive protocols through a conditional score chain rule.
  • Discussion: The Fisher-information trace bound for support rotations and the conditional score chain rule are highlighted as reusable ingredients for quantum estimation.The chain rule lets adaptive analyses build on a bound for each joint measurement.
  • Limitations and future work: The Gaussian protocol uses continuous-outcome joint measurements, leaving finite-outcome implementations with the same optimal complexity as an open direction.The number of outcomes required for such implementations is also left open.
  • Limitations and future work: The results retain only classical information between rounds, while persistent quantum memory can combine rounds into a larger coherent measurement.With quantum memory, the number of fresh samples measured in one round no longer captures the full quantum resource.

7 AI disclosure

The authors disclose using GPT-5.5 and GPT-5.6 Sol for mathematical assistance, reformulation, literature searches, and editorial revisions. They state that the authors independently verified the mathematical arguments and references and take responsibility for the paper.

  • Mathematical assistance: GPT-5.5 and GPT-5.6 Sol assisted with proofs of two lower-bound lemmas concerning Fisher-information bounds for t>1 joint measurements.The authors describe these lemmas as crucial to the t-dependence of the lower bound.
  • Mathematical assistance: The models helped reformulate earlier tomography algorithms and analyses without representation theory for the upper-bound presentation.The paper identifies the rank-dependent error analysis of the Gaussian formulation as its key additional contribution.
  • Verification and responsibility: The models were also used for general mathematical discussions, literature searches, and editorial revisions.The authors state that they independently verified all mathematical arguments and references.

A Statistical formalism

The appendix establishes the measure-theoretic foundations for continuous-outcome POVMs and adaptive transcripts, then verifies likelihood regularity needed for Fisher-information arguments and van Trees. It also constructs a parameter-independent dominating measure for adaptive protocols and proves the required differentiability and integrability properties.

  • A Statistical formalism: A fixed, state-independent reference measure supports likelihood densities, null-event arguments, and Fisher-information derivatives throughout the dominated model.The reference measure is σ-finite, and the Radon–Nikodym theorem supplies the relevant densities.
  • A.1 Dominated models and finite-dimensional POVMs: For finite-dimensional POVMs, the trace measure dominates every state-induced outcome law and yields an almost-everywhere operator-valued density with the required positivity and trace identity.Absolute continuity follows because a positive semidefinite POVM element with zero trace must vanish.
  • A.1 Dominated models and finite-dimensional POVMs: Zero-likelihood outcomes contribute zero score and Fisher integrands because differentiability of a nonnegative likelihood forces its gradient to vanish at interior zeros.This convention is used in the Fisher-information analysis.
  • A.2 Adaptive transcript domination: Adaptive protocols are represented by measurable history-dependent POVM kernels, with each round using at most t samples and measurable dependence on prior outcomes.Protocols are padded to a fixed number of rounds, and zero-sample rounds are included.
  • A.2 Adaptive transcript domination: Each adaptive POVM kernel admits a jointly measurable positive semidefinite density whose trace equals the dimension of the measured sample space.The construction uses entrywise Radon–Nikodym derivatives, finite partitions, martingale convergence, and measurable exceptional-set repairs.
  • A.2 Adaptive transcript domination: The resulting parameter-independent measure dominates the full adaptive transcript law for every input state.Conditional densities are iterated through the Ionescu–Tulcea construction and the conditional Born rule.
  • A.3 Likelihood regularity: Likelihoods are C1 pointwise and in the relevant L1 spaces, with jointly measurable derivatives that integrate to zero under both conditional and full transcript measures.The proof uses trace/operator norm duality, dominated convergence, and differentiation under integrals.
  • A.3 Likelihood regularity: On compact parameter sets, transcript likelihoods satisfy 0 ≤ qN,X ≤ d^tN and |D_HqN,X| ≤ N C_K d^tN ∥H∥F.These bounds follow from the product structure of the N conditional likelihoods.

B Measurable projection onto low-rank states

The appendix proves that nearest-state projection can be implemented measurably when estimating a low-rank quantum state from a Hermitian matrix. A Borel selection resolves possible ties among Frobenius-norm minimizers.

  • B Measurable projection onto low-rank states: The upper-bound estimator projects a Hermitian matrix to a nearest state in the Frobenius norm.The projection is restricted to the set of states with rank at most r.
  • B Measurable projection onto low-rank states: For every 1 ≤ r ≤ d, a Borel map assigns each Hermitian matrix a minimizer over rank-at-most-r density operators.This provides a measurable estimator despite potentially nonunique minimizers.
  • B Measurable projection onto low-rank states: Measurability follows from the measurable maximum theorem applied to the compact low-rank-state correspondence and continuous negative Frobenius-distance objective.The theorem supplies a measurable selection of minimizers.
Loading 2609.10514v1…