State Discrimination, Hypothesis Testing and Estimation

Introduction

Quantum information processing rarely provides direct access to an unknown quantum state. A qubit may have been prepared by one of several physical processes, transmitted through a noisy channel, or produced by a sensor. The receiver must then make an inference from measurement outcomes. Three closely related tasks arise:

1. State discrimination: deciding which of several possible states was prepared.
2. Quantum hypothesis testing: choosing between competing explanations of observed data.
3. Parameter estimation: inferring an unknown continuous quantity, such as a phase, probability or field strength.

These tasks connect quantum theory with statistics, signal processing, communications, medical sensing and cybersecurity. The central difficulty is that nonorthogonal quantum states cannot generally be distinguished perfectly. Measurement changes the state, and the information available from a single copy is limited.

This chapter develops the principal tools used to quantify and manage that limitation: trace distance, fidelity, the Helstrom measurement and bound, quantum hypothesis-testing error probabilities, and estimation theory based on Fisher information and the quantum Cramér–Rao bound.

Learning objectives

By the end of this chapter, you should be able to:

• explain why nonorthogonal states cannot be perfectly discriminated;
• calculate trace distance and interpret it operationally;
• calculate fidelity and distinguish it from trace distance;
• use the Helstrom bound to determine the minimum error probability for binary discrimination;
• formulate a quantum hypothesis test and distinguish type-I and type-II errors;
• explain the role of likelihood ratios and quantum relative entropy;
• describe quantum parameter estimation and apply the quantum Fisher-information bound;
• design a simple measurement strategy and evaluate its limitations;
• identify practical, ethical and governance issues in measurement-based decision systems.

1. Quantum states as statistical evidence

A quantum state is represented by a density operator ρ satisfying three conditions:

• ρ is Hermitian;
• ρ is positive semidefinite;
• Tr(ρ) = 1.

A measurement is described by a positive-operator-valued measure (POVM), a collection {Eᵢ} such that Eᵢ ≥ 0 and ∑ᵢ Eᵢ = I. If the system is in state ρ, the probability of outcome i is given by the Born rule:

Here, Eᵢ is the effect associated with outcome i, I is the identity operator, and Tr denotes the trace. The formula assumes an ideal measurement model and gives probabilities, not deterministic outcomes.

A measurement converts a quantum state into a classical probability distribution. Consequently, every quantum decision problem can be viewed as a search for a measurement that extracts the most useful evidence while respecting quantum-mechanical constraints.

Suppose a source prepares state ρ₀ with prior probability π₀ or state ρ₁ with prior probability π₁, where π₀ + π₁ = 1. A decision rule assigns each measurement outcome to one of the hypotheses. The average probability of error depends on the states, their priors and the measurement.

Orthogonal states, such as |0⟩ and |1⟩, can be perfectly distinguished using a measurement in the computational basis. Nonorthogonal states, such as |0⟩ and |+⟩ = (|0⟩ + |1⟩)/√2, cannot be perfectly distinguished from one copy. This is not a technological imperfection; it is a structural property of quantum theory.

2. Trace distance: distinguishability of states

For density operators ρ and σ, the trace distance is

where ||A||₁ = Tr(√(A†A)) is the trace norm. For density operators, 0 ≤ D(ρ,σ) ≤ 1. A value of 0 means that the states are identical. A value of 1 means that they have orthogonal supports and can be perfectly distinguished by a suitable measurement.

The trace distance has a direct operational interpretation. For any two-outcome measurement, the largest possible difference between the probabilities of a selected outcome is

The maximization is over all valid measurement effects E. Thus trace distance measures the maximum statistical advantage available to a measurement when deciding whether a state is ρ or σ.

For pure states |ψ⟩ and |φ⟩, the expression simplifies to

This formula assumes normalized pure states. If the states are identical up to a global phase, their overlap has magnitude 1 and the distance is zero. If they are orthogonal, the overlap is zero and the distance is one.

Example 1: trace distance between two qubit states

Consider |ψ⟩ = |0⟩ and |φ⟩ = |+⟩. Their overlap is

⟨0|+⟩ = 1/√2

Therefore,

The states are substantially distinguishable, but not perfectly distinguishable. A single measurement cannot identify the preparation with certainty in every trial.

Trace distance also obeys contractivity under a quantum channel 𝓔:

A quantum channel is a completely positive, trace-preserving map. The inequality means that physical processing cannot increase distinguishability. Noise, loss and imperfect transmission cannot create evidence that was absent initially.

3. Fidelity: closeness and overlap

A commonly used measure of similarity is the quantum fidelity. One standard convention is

Under this convention, 0 ≤ F(ρ,σ) ≤ 1, with F = 1 for identical states and F = 0 for states with orthogonal supports. Some texts use the unsquared quantity Tr(√(√ρσ√ρ)); therefore, the convention should always be stated.

For pure states, fidelity becomes

Fidelity and trace distance answer related but different questions. Fidelity measures overlap or similarity, while trace distance measures the maximum operational distinguishability. They satisfy the Fuchs–van de Graaf inequalities:

These inequalities apply when fidelity uses the squared convention above. They give bounds rather than, in general, an equality. For pure states, the upper bound is attained because D = √(1 − F).

Example 2: fidelity and distance

For |0⟩ and |+⟩, F = 1/2 and D = 1/√2 ≈ 0.707. The inequalities give

1 − √(1/2) ≈ 0.293 ≤ D ≤ √(1 − 1/2) ≈ 0.707

The actual distance equals the upper bound. A fidelity of 0.5 should not be interpreted as a 50% probability that two physical preparation procedures are identical; it is a state-overlap measure.

4. Binary state discrimination and the Helstrom bound

Assume that ρ₀ and ρ₁ are prepared with prior probabilities π₀ and π₁. A binary measurement has effects E₀ and E₁ satisfying E₀ + E₁ = I. The probability of correct classification is

Since E₁ = I − E₀, maximizing this expression gives the Helstrom result:

Equivalently,

The formula assumes binary hypotheses, known priors, and unrestricted quantum measurements on one copy. The minimum is over all POVMs.

For equal priors, π₀ = π₁ = 1/2, the result simplifies to

Thus trace distance directly determines the best possible single-copy decision performance under equal priors.

The Helstrom measurement

Define the Hermitian operator

The optimal decision rule measures in the eigenbasis of Λ. The subspace associated with positive eigenvalues is assigned to hypothesis 0, while the subspace associated with negative eigenvalues is assigned to hypothesis 1. Zero-eigenvalue subspaces may be assigned arbitrarily without changing the optimum.

This result is important in practice: it converts an abstract optimization over all measurements into an eigenvalue problem. However, the required measurement may be difficult to implement experimentally, especially for high-dimensional systems or unknown prior probabilities.

Example 3: equal-prior discrimination of |0⟩ and |+⟩

For equal priors and the states in the previous examples, D = 1/√2. Therefore,

The maximum success probability is approximately 0.854. Even the optimal measurement makes an error in about 14.6% of single-copy trials.

5. Multiple copies and collective measurements

If n identical copies are available, the hypotheses become ρ₀⊗ⁿ and ρ₁⊗ⁿ. More copies generally improve discrimination because they provide more evidence. For pure states, the fidelity satisfies

when the stated fidelity convention is used. If 0 < F < 1, the overlap decreases exponentially with n.

A measurement may be performed separately on each copy, or jointly across all copies. Collective measurements can outperform independent measurements for some tasks, although they may require substantial experimental resources. In communications, this creates a trade-off between theoretical optimality and implementability.

The quantum Chernoff bound describes the asymptotic exponential decay of the minimum error probability for many copies. For two states, the Chernoff exponent is determined by

The expression assumes finite-dimensional density operators and natural logarithms unless another base is specified. As n becomes large, the optimal error probability is governed asymptotically by exp(−nξ_Q), up to subexponential factors. This is an asymptotic result, not an exact finite-sample formula.

6. Quantum hypothesis testing

Hypothesis testing formalizes decisions between competing claims. Let H₀ represent a null hypothesis with state ρ and H₁ represent an alternative with state σ. A test is an effect T, where outcome T means “accept H₁” and I − T means “accept H₀”.

Two error probabilities are central:

is the type-I error: rejecting H₀ when H₀ is true. Also,

is the type-II error: accepting H₀ when H₁ is true.

Unlike the symmetric Helstrom problem, hypothesis testing often fixes a maximum acceptable α and minimizes β. The optimal type-II error at significance level α is

The notation β_α(ρ||σ) denotes the best achievable type-II error under the stated type-I constraint. The ordering of ρ and σ matters.

A quantum analogue of the likelihood ratio is obtained through the operator structure of the two states. In commuting cases, ρ and σ can be diagonalized simultaneously, and quantum hypothesis testing reduces to classical testing of probability distributions. In noncommuting cases, no single basis simultaneously reveals all classical likelihood information, and the optimization over quantum measurements is essential.

Quantum relative entropy

Quantum relative entropy is

provided the support of ρ is contained in the support of σ; otherwise it is defined as +∞. Logarithms may be base e, giving units of nats, or base 2, giving units of bits. Relative entropy is nonnegative but is not symmetric: generally D(ρ||σ) ≠ D(σ||ρ).

It quantifies the asymptotic distinguishability of hypotheses. The quantum Stein lemma states, informally, that when the type-I error is held below a fixed value less than one and many independent copies are available, the optimal type-II error exponent approaches D(ρ||σ). This result explains why relative entropy is fundamental in quantum detection, channel discrimination and security analysis.

7. Measurement design in practice

An operational measurement-design workflow is:

1. Specify the candidate states, their priors or decision costs.
2. Decide whether the objective is symmetric error minimization or constrained hypothesis testing.
3. Calculate a distinguishability measure such as D or F.
4. Construct the Helstrom operator Λ for a binary minimum-error problem.
5. Determine whether the optimal POVM is physically implementable.
6. If not, restrict the measurement family and optimize within that family.
7. Validate performance using empirical error rates and confidence intervals.
8. Monitor drift, calibration error, state-preparation error and prior-probability changes.

A measurement that is mathematically optimal under a model may be inferior in the laboratory if it is sensitive to loss or calibration. Robust design therefore evaluates performance under model uncertainty, not only under ideal assumptions.

8. Quantum parameter estimation

State discrimination concerns a finite set of alternatives. Estimation concerns an unknown continuous parameter θ. A parameterized quantum model is a family of states ρ_θ. A measurement produces data with probability

where {E_x} is a POVM indexed by outcome x.

An estimator θ̂(x) maps outcomes to numerical estimates. Its mean-square error is

where expectation is over p(x|θ). Bias is

An unbiased estimator has b(θ̂) = 0 for all relevant θ. In practice, slightly biased estimators may have lower mean-square error, so unbiasedness should not be treated as an unconditional requirement.

Classical Fisher information

For a specified measurement, the classical Fisher information is

for discrete outcomes, with the sum replaced by an integral for continuous outcomes. The classical Cramér–Rao bound states that for an unbiased estimator based on n independent trials,

The bound assumes regularity conditions, including differentiability and an appropriate parameter-independent outcome space. It is a lower bound and need not be attainable at finite sample size.

Quantum Fisher information

Optimizing classical Fisher information over all POVMs defines the quantum Fisher information I_Q(θ). For the symmetric logarithmic derivative L_θ, defined implicitly by

quantum Fisher information is

The quantum Cramér–Rao bound is

for locally unbiased estimators under the usual regularity conditions. It represents the best precision permitted by the state model, while I_C represents the information obtained by a particular measurement. Always,

Example 4: phase estimation with a qubit

Consider the pure state

|ψ_θ⟩ = (|0⟩ + e^iθ|1⟩)/√2

and measure in the basis |+⟩, |−⟩. The probabilities are

Differentiating and substituting into the classical Fisher-information formula gives I_C(θ) = 1 wherever the probabilities are nonzero, by continuity at the limiting points. Thus n independent measurements satisfy

for an unbiased estimator that uses this measurement and operates in a locally identifiable region. The phase is periodic, so a globally unique estimate requires a stated interval or additional measurements; otherwise θ and θ + 2π represent the same physical state.

9. Applied case: quantum-enhanced optical inspection

Illustrative case. A manufacturing company uses a low-light optical sensor to decide whether a microscopic component contains a defect. A probe interacts with the component and returns one of two approximate quantum states: ρ₀ for “no defect” and ρ₁ for “defect.” The states are not orthogonal because background light and detector noise cause overlap.

Background and problem

The company initially uses a fixed threshold on detector intensity. The defect rate is low, so the prior probabilities are unequal: π₀ = 0.95 and π₁ = 0.05. Missing a defect is more costly than inspecting a nondefective component, so symmetric error minimization may not reflect the business objective.

Analysis

Engineers characterize ρ₀ and ρ₁ by calibration experiments. They calculate the trace distance to estimate the maximum distinguishability and the fidelity to monitor drift. They then form

Λ = π₀ρ₀ − π₁ρ₁

for minimum average error. Because false negatives are especially costly, they also evaluate a cost-sensitive decision rule rather than relying only on the Helstrom criterion.

Approach and solution

The laboratory implements an approximate Helstrom measurement using an interferometer and calibrated photodetectors. A validation set is used to estimate type-I and type-II error rates. The threshold is selected so that the probability of missing a defect remains below a specified engineering limit. Periodic recalibration is introduced because optical loss changes the effective states.

Outcome

The approximate measurement improves detection at the same false-alarm rate. It does not reach the theoretical Helstrom bound because of detector inefficiency and imperfect state preparation. Nevertheless, the trace-distance analysis identifies the maximum possible improvement, while the hypothesis-testing analysis selects a threshold consistent with the company’s risk tolerance.

Learning points

• Priors and costs must be stated explicitly.
• The Helstrom bound is a benchmark, not a guarantee of laboratory performance.
• Type-I and type-II errors should be reported separately.
• Calibration drift can invalidate a measurement designed for an earlier state model.
• A technically superior detector must still be evaluated for safety, traceability and operational reliability.

10. Limitations, governance and responsible use

The mathematical model may be wrong or incomplete. State-preparation errors, correlated noise, detector dark counts and finite calibration data can produce biased estimates of ρ and σ. Confidence intervals or Bayesian posterior distributions should accompany reported performance estimates.

In safety-critical applications, a low average error is insufficient. Rare but severe errors may require conservative thresholds, human review and fail-safe procedures. If quantum measurements are used in medical, security or employment decisions, governance must address privacy, fairness, auditability and the right to challenge an automated decision.

Quantum advantage should not be claimed merely because a quantum state or detector is involved. The relevant comparison is with the best practical classical baseline under equal resource, cost and data constraints. In addition, quantum estimation may be locally optimal but globally ambiguous, particularly for periodic parameters such as phase.

11. Best practices

• State the density-operator convention and fidelity convention explicitly.
• Report priors, costs, sample size and whether measurements are collective.
• Distinguish theoretical bounds from achieved experimental results.
• Use trace distance for operational distinguishability and fidelity for overlap or similarity.
• Report type-I and type-II errors separately in hypothesis testing.
• Check identifiability before estimating a parameter.
• Include calibration uncertainty and model mismatch in performance analysis.
• Use independent validation data rather than evaluating a detector only on its design data.
• Prefer interpretable decision rules where errors affect people or safety.
• Reassess the measurement when noise, priors or operating conditions change.

Chapter Summary

State discrimination asks which member of a known set of quantum states was prepared. Trace distance quantifies maximum operational distinguishability, while fidelity quantifies state overlap. For binary minimum-error discrimination, the Helstrom bound gives the exact optimum over all single-copy POVMs, and the Helstrom operator identifies the corresponding measurement structure.

Quantum hypothesis testing separates type-I and type-II errors and permits one error to be constrained while the other is minimized. Quantum relative entropy governs asymptotic discrimination rates. Parameter estimation instead infers a continuous quantity from measurement outcomes. Classical Fisher information describes the information supplied by a particular measurement, whereas quantum Fisher information gives the best information allowed by the quantum model. The quantum Cramér–Rao bound converts information into a precision limit.

In professional applications, optimality must be balanced against implementability, calibration, cost, risk and governance. A reliable quantum decision system therefore combines theoretical bounds, experimentally validated measurements, explicit uncertainty analysis and responsible operational controls.

Key Terms

Trace distance: A metric measuring the maximum distinguishability of two quantum states.

Trace norm: The sum of the singular values of an operator.

Fidelity: A measure of similarity or overlap between quantum states.

Density operator: A positive, unit-trace operator representing a quantum state.

POVM: A set of positive measurement effects summing to the identity.

Effect: A positive operator representing one measurement outcome.

Born rule: The rule p(i) = Tr(ρEᵢ) for measurement probabilities.

State discrimination: Decision-making among a finite set of candidate states.

Helstrom bound: The minimum binary discrimination error over all POVMs.

Helstrom operator: Λ = π₀ρ₀ − π₁ρ₁ for a binary decision problem.

Prior probability: Probability assigned to a hypothesis before measurement data.

Type-I error: Rejection of a true null hypothesis.

Type-II error: Failure to reject a false null hypothesis.

Quantum hypothesis test: A quantum measurement-based procedure for choosing between hypotheses.

Quantum relative entropy: An asymmetric measure of distinguishability based on logarithms of density operators.

Parameter estimation: Inferring an unknown continuous parameter from data.

Fisher information: A measure of information about a parameter contained in a probability model.

Quantum Fisher information: The maximum Fisher information over quantum measurements.

Cramér–Rao bound: A lower bound on the variance of an unbiased estimator.

Identifiability: The property that different parameter values produce distinguishable statistical models.

Review Questions

Short Answer Questions

1. What operational quantity does trace distance measure?
2. State the squared-fidelity formula for two pure states.
3. What is the Helstrom bound for equal-prior binary discrimination?
4. Distinguish type-I error from type-II error.
5. Why can a quantum channel not increase trace distance?

Long Answer Questions

1. Explain the relationship between trace distance, fidelity and the Fuchs–van de Graaf inequalities.
2. Derive the equal-prior form of the Helstrom minimum-error probability from the general binary formula.
3. Explain how the Helstrom operator determines an optimal binary measurement.
4. Compare classical Fisher information with quantum Fisher information and explain their roles in estimation.
5. Discuss why a theoretically optimal quantum measurement may not be the best practical measurement.

Analytical / Application Questions

1. Two pure states have overlap magnitude 0.8 and equal prior probabilities. Calculate their fidelity, trace distance and minimum single-copy error probability.
2. Design a hypothesis-testing procedure for a sensor that must keep the type-I error below 1% while minimizing type-II error. State the data, measurement, validation and reporting requirements.
3. A phase-estimation system has quantum Fisher information 25 per copy and uses 100 independent copies. Calculate the quantum Cramér–Rao lower bound on variance for a locally unbiased estimator, and explain two reasons why an experiment may fail to attain it.

Multiple Choice Questions with Answers

Question 1

Which quantity equals the maximum difference in the probability of a selected measurement outcome between two states?

A. Trace distance

B. Quantum relative entropy

C. Von Neumann entropy

D. Purity

Correct Answer: A. Trace distance

Question 2

For equal priors, the minimum binary discrimination error is expressed as

A. 1/2[1 + D(ρ,σ)]

B. 1/2[1 − D(ρ,σ)]

C. D(ρ,σ)²

D. 1 − 2D(ρ,σ)

Correct Answer: B. 1/2[1 − D(ρ,σ)]

Question 3

Under the squared fidelity convention, what is the fidelity between |0⟩ and |+⟩?

A. 0

B. 1/√2

C. 1/2

D. 1

Correct Answer: C. 1/2

Question 4

The Helstrom operator for priors π₀ and π₁ is

A. π₀ρ₀ + π₁ρ₁

B. ρ₀ρ₁ − ρ₁ρ₀

C. π₀ρ₀ − π₁ρ₁

D. ρ₀ + ρ₁ − I

Correct Answer: C. π₀ρ₀ − π₁ρ₁

Question 5

A type-I error occurs when a test

A. Accepts a true null hypothesis

B. Rejects a true null hypothesis

C. Accepts a false alternative hypothesis

D. Uses an invalid POVM

Correct Answer: B. Rejects a true null hypothesis

Question 6

If two states are orthogonal, their trace distance is

A. 0

B. 1/2

C. √2

D. 1

Correct Answer: D. 1

Question 7

Which expression defines quantum relative entropy, when the support condition is satisfied?

A. Tr[ρ(log ρ − log σ)]

B. Tr[(ρ − σ)²]

C. 1 − F(ρ,σ)

D. Tr(ρσ)

Correct Answer: A. Tr[ρ(log ρ − log σ)]

Question 8

For n independent copies, the quantum Cramér–Rao bound for an unbiased estimator is

A. Var(θ̂) ≥ nI_Q(θ)

B. Var(θ̂) ≥ I_Q(θ)/n

C. Var(θ̂) ≥ 1/[nI_Q(θ)]

D. Var(θ̂) ≥ 1/I_Q(θ)²

Correct Answer: C. Var(θ̂) ≥ 1/[nI_Q(θ)]

Question 9

If two pure states have overlap magnitude 0.8, their trace distance is

A. 0.20

B. 0.36

C. 0.64

D. 0.80

Correct Answer: B. 0.36

Question 10

Which statement about a quantum channel 𝓔 is correct?

A. It always increases fidelity and trace distance

B. It can increase trace distance but not fidelity

C. It cannot increase trace distance

D. It makes all distinct states orthogonal

Correct Answer: C. It cannot increase trace distance