Quantum Measurement Verification

• Quantum Measurement Verification (QMV) is a framework that uses measurement data to evaluate if quantum devices and processes perform as intended without full tomography.
• The approach integrates protocols like randomized measurements and Hamiltonian learning to reliably assess device performance despite noise and SPAM errors.
• QMV advances quantum computing by offering efficient, scalable verification metrics that adapt to various hardware models and trust assumptions.

Searching arXiv for recent and foundational papers on quantum measurement verification and adjacent verification paradigms. Quantum Measurement Verification (QMV) denotes the use of measurement data to decide whether a quantum device, measurement process, or measurement-mediated computation behaves as intended, typically without full tomography and often without a trustworthy classical simulator. In current usage it appears inside broader programs of quantum verification and quantum characterization, verification, and validation (QCVV): Hamiltonian learning for analog simulators, cross-device comparison by randomized measurements, output verification of untrusted processors through delegated basis measurements, and tomography- or instrument-based characterization of noisy readout and mid-circuit measurements (Carrasco et al., 2021, Blume-Kohout et al., 20 Mar 2025).

1. Conceptual scope

QMV is not a single protocol family with a single metric. In the QCVV framing, it sits between characterization techniques that estimate predictive models for a device’s behavior from data and benchmarking techniques that assess overall performance of a device. Measurement is treated as one of the three core objects of QCVV alongside states and control operations, so verification can target terminal readout, mid-circuit measurements, or the measurement subroutines that underpin delegated and interactive verification protocols (Blume-Kohout et al., 20 Mar 2025).

The standard operational model is noisy open-system quantum mechanics. Ideal measurements are represented by projective measurements, whereas noisy readout is modeled by POVMs with outcome probabilities

$\Pr(k|\psi)=\operatorname{Tr}[E_k\rho].$

Within this framework, verification may ask whether the observed statistics match an advertised POVM, whether the post-measurement state update is reliable, or whether a measurement instrument embedded in a larger device remains valid in the presence of SPAM errors, decoherence, and temporal correlations. The same source stresses that non-Markovian dynamics can invalidate QCVV protocols that rely on Markovian assumptions, so QMV is unavoidably model-dependent when memory effects are present (Blume-Kohout et al., 20 Mar 2025).

The quantum-verification perspective broadens the scope further. For noisy intermediate-scale quantum devices, verification often relies on indirect, measurement-based protocols rather than full tomography or brute-force classical simulation. The three verification problems emphasized in this perspective are analog quantum simulation, cross-device verification, and output verification of an untrusted quantum processor; measurement is central to all three, but with different operational roles in each case (Carrasco et al., 2021).

A still broader extension appears in Quantum Measure Theory. There the generalized measure $\mu(E)$ assigned to a history-space event $E$ is not generally the expectation value of any selfadjoint operator or POVM. Nonetheless, an ancilla-assisted protocol can be arranged so that a final “yes” probability is proportional to the quantum measure. For an event with $k$ histories,

$P(E)=\frac{\mu(E)}{k},$

so $\mu(E)=k\,P(E)$. This places QMV-like reasoning outside the ordinary observable/POVM setting and into history-based operational tests (Frauca et al., 2016).

2. Figures of merit and statistical decision rules

No universal measurement-verification metric is singled out across the literature. The QCVV review instead presents a menu of quality measures—state fidelity, trace distance, and diamond norm—and emphasizes that certification typically amounts to testing whether a measured quality parameter exceeds a threshold with confidence bounds (Blume-Kohout et al., 20 Mar 2025).

For direct verification of an ideal projective measurement

$$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$

against an implemented POVM

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$

a recent QMV framework defines the measurement fidelity

$$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$

The bad case is specified by

$F(\mathcal P,\mathcal M)\le 1-\varepsilon,$

and a protocol $\mu(E)$0 is judged by its worst-case passing probability $\mu(E)$1. For homogeneous protocols,

$\mu(E)$2

so the required number of rounds obeys

$\mu(E)$3

This brings measurement verification into the same hypothesis-testing form used in quantum state verification (Wang et al., 19 Jun 2026).

In target-state verification, the standard object is the verification operator

$\mu(E)$4

For small infidelity $\mu(E)$5, the required sample size scales as

$\mu(E)$6

while the globally optimal projective test achieves

$\mu(E)$7

In experiments with imperfect states, the observed acceptance ratio $\mu(E)$8 can be converted into a confidence bound through

$\mu(E)$9

where $E$0 is the binary relative entropy. This is a concrete decision rule: accept a verification claim when the observed pass fraction is high enough, reject otherwise (Zhang et al., 2020).

Cross-device verification uses related but distinct state-comparison quantities. One perspective highlights the experimentally friendly

$E$1

which defines a metric and coincides with the Uhlmann fidelity if one state is pure. In that setting, fidelities, overlaps, and purities are reconstructed from randomized measurements rather than from a full state model (Carrasco et al., 2021).

3. Protocol families

For analog quantum simulators, QMV takes the form of Hamiltonian learning. If a prepared stationary state $E$2 commutes with the parent Hamiltonian,

$E$3

then for observables $E$4 and an ansatz

$E$5

one obtains linear constraints

$E$6

The reconstructed Hamiltonian is the lowest right singular vector of $E$7, and verification is performed by comparing the learned coefficients with the target coefficients. For the Fermi–Hubbard example, the required observables are accessible through local current measurements such as

$E$8

The practical point is locality: the number of measurements required for fixed accuracy scales polynomially with system size (Carrasco et al., 2021).

For cross-device verification, randomized measurements are central. A random unitary $E$9 is applied from a tomographically complete set and the system is then measured in the computational basis. From the probabilities

$k$0

one estimates overlaps through the Hamming-distance kernel

$k$1

The same framework yields purities and reduced-state properties, and it connects naturally to shadow tomography, where randomized measurements are used to predict expectation values directly rather than reconstructing the whole state. A proposed central data repository would turn such randomized-measurement data into a standardized cross-device verification infrastructure (Carrasco et al., 2021).

Adaptive local protocols improve the efficiency of target-state verification without requiring entangled measurements. In an experimental study on the two-qubit state

$k$2

three strategies were compared: LO, Uni-LOCC, and Bi-LOCC. Classical communication improves the constant prefactor in the $k$3 law: the overhead relative to the global optimum is $k$4 for LO, $k$5 for Uni-LOCC, and $k$6 for Bi-LOCC. Experimentally, the constant factor was reduced from roughly $k$7 to $k$8, so only about 60% of the measurements were needed compared with the optimal non-adaptive local strategy (Zhang et al., 2020).

A related line of work optimizes not the measurement family but the measurement order. Given a fixed informationally complete set $k$9, verification is cast in terms of two stopping criteria: certify non-accuracy when

$P(E)=\frac{\mu(E)}{k},$0

and certify accuracy when

$P(E)=\frac{\mu(E)}{k},$1

The proposed OS, IOS, IAS, and AV strategies show that ordering can significantly reduce the number of measurements needed, with the adaptive protocol particularly useful for faulty preparations (Liang et al., 2023).

Quantum memory changes the same optimization problem more radically. In memory-assisted verification, the verifier stores several copies and measures them collectively. For two-copy verification the efficiency is governed by a new intrinsic quantity $P(E)=\frac{\mu(E)}{k},$2, with

$P(E)=\frac{\mu(E)}{k},$3

For multi-qubit graph states, a two-copy strategy using only local Bell measurements satisfies $P(E)=\frac{\mu(E)}{k},$4 and reaches the global-optimum scaling

$P(E)=\frac{\mu(E)}{k},$5

For GHZ-like states, a dimension-expansion technique improves performance monotonically with memory size (Chen et al., 2023).

The most explicit recent QMV framework for entangled measurements uses only local state preparations. For locally transitive and irreducible projective measurements, symmetry reduces locality-constrained QMV to state verification of a single basis state. The resulting homogeneous verification operators were worked out for generalized Bell measurements, single-parameter two-qubit measurements, elegant joint measurements, and stabilizer-state-induced measurements, together with closed-form success probabilities and sample complexities (Wang et al., 19 Jun 2026).

4. Distrust models: measurement-device-independent and device-independent verification

A major branch of QMV studies settings in which the measurement devices themselves are not trusted. In the quantum refereed steering game, both parties and their devices are treated as untrusted, and a loss-tolerant steering inequality due to Bennet et al. is converted into a measurement-device-independent criterion by replacing Bob’s trusted observables with referee-supplied states

$P(E)=\frac{\mu(E)}{k},$6

The definitive score is

$P(E)=\frac{\mu(E)}{k},$7

A central threshold is

$P(E)=\frac{\mu(E)}{k},$8

which is sufficient to certify steering with a maximally entangled two-qubit state, while Bob’s inefficiency only rescales the score by $P(E)=\frac{\mu(E)}{k},$9 and therefore does not change its sign as long as $\mu(E)=k\,P(E)$0 (Jeon et al., 2018).

Measurement-device-independent channel verification uses a semi-quantum game in time. Alice trusts only her preparation device and sends two non-orthogonal quantum inputs, taken in the experimental realization as two sets of four SIC qubit states. Bob’s channel and detector are treated as a black box. The payoff

$\mu(E)=k\,P(E)$1

implements an entanglement witness for the channel’s Choi matrix, and the channel is certified quantum when

$\mu(E)=k\,P(E)$2

In the ideal construction $\mu(E)=k\,P(E)$3; in the experiment the effective entanglement-breaking threshold was corrected upward to about $\mu(E)=k\,P(E)$4 to account for multiphoton events and other imperfections. The method remained effective under depolarizing noise, dephasing noise, imperfect two-photon interference, and realistic detector inefficiencies (Graffitti et al., 2019).

Device-independent QMV removes even the trusted-preparation side. In a Bell scenario with one three-outcome setting on Alice’s side, the correlation expression

$\mu(E)=k\,P(E)$5

satisfies

$\mu(E)=k\,P(E)$6

for all binary quantum measurements in any dimension, while the maximal quantum value is

$\mu(E)=k\,P(E)$7

Experimentally, the reported value

$\mu(E)=k\,P(E)$8

excluded binary-measurement models by more than $\mu(E)=k\,P(E)$9 standard deviations. Combined with a CHSH-based self-test certifying a qubit–qubit maximally entangled state to fidelity at least $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$0 at $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$1, this yielded the first device-independent certification of a nonprojective qubit measurement (Gómez et al., 2016).

5. Restricted verifiers and delegated measurement protocols

QMV is also central in complexity-theoretic and delegated-computation settings, where the verifier is deliberately weak. One result shows that QMA verification remains possible when Arthur receives the proof qubit-by-qubit and performs only single-qubit measurements, with no quantum memory and no multiqubit operations. Two independent proofs are given: an MBQC construction with a stabilizer test, which also works for QMA$$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$2, and a Local Hamiltonian construction based on sampling Pauli terms and estimating energy using single-qubit Pauli measurements (Morimae et al., 2015).

The cryptographic version of the same idea culminates in a classical-verifier measurement protocol. Under an extended trapdoor claw-free family based on LWE hardness, a classical verifier can force a quantum prover to commit to an $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$3-qubit state, later measure each qubit in either the standard basis or the Hadamard basis, and report the results. For every accepted prover there exists an underlying state $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$4 such that the prover’s output distribution is computationally indistinguishable from the honest measurement distribution of $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$5 in the chosen bases. This measurement-verification primitive is then combined with Hamiltonian-energy estimation, yielding the statement

$$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$6

under the stated cryptographic assumption (Mahadev, 2018).

Measurement-only blind quantum computing supplies another restricted-verifier model. In the original verification protocol, Alice can only measure qubits, while Bob prepares the resource. Two constructions are given. The trap-based scheme hides the computation register inside

$$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$7

with secret permutation $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$8 and trap qubits measured in the $$\mathcal P=\{\ket{\psi_\theta}\!\bra{\psi_\theta}\}_{\theta\in\Theta}$$9 or $\mathcal M=\{M_\theta\}_{\theta\in\Theta},$0 bases. If Bob changes the logical outcome and still avoids detection, the cheating probability is at most

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$1

A second, topological-code-based protocol removes explicit traps and gives the bound

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$2

In both cases blindness follows from the no-signaling principle (Morimae, 2012).

A later extension adds quantum input verification. Alice hides an unknown $\mathcal M=\{M_\theta\}_{\theta\in\Theta},$3-qubit input inside

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$4

then applies a random permutation and Pauli one-time pad before sending the state to Bob. In an input-state test, hidden traps allow Alice to decode the state and check the projector

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$5

The protocol verifies the graph-state computation resource, the correctness of the quantum input, and blindness, while retaining an inverse-polynomial completeness-soundness gap (Morimae, 2016).

6. Resource-state, fault-tolerant, and on-chip verification

Verifiable fault tolerance in MBQC treats QMV as a test of whether an arbitrary resource state lies inside a correctable error set rather than as a full noise-identification problem. For a two-colorable graph state, the verification tests are

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$6

If the protocol passes with significance level $\mathcal M=\{M_\theta\}_{\theta\in\Theta},$7, then the state $\mathcal M=\{M_\theta\}_{\theta\in\Theta},$8 of the computation block satisfies

$\mathcal M=\{M_\theta\}_{\theta\in\Theta},$9

For any POVM $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$0, the corresponding output distribution is close to that of the corrected state, and the verification overhead including classical processing is linear in the size of the quantum computation (Fujii et al., 2016).

Hypergraph-state verification extends the single-qubit-measurement paradigm beyond graph states. For hypergraph states

$$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$1

the stabilizers contain $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$2 terms and are therefore not directly local Pauli observables. By expanding each $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$3 into a sum of Pauli operators, the protocol reduces each stabilizer test to sequential single-qubit $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$4 and $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$5 measurements. With explicit choices

$$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$6

the protocol achieves completeness greater than $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$7 and guarantees that, upon acceptance, the computation register has fidelity at least $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$8 with the target hypergraph state with probability greater than $$F(\mathcal P,\mathcal M):=\frac{1}{|\Theta|}\sum_{\theta\in\Theta}\bra{\psi_\theta}M_\theta\ket{\psi_\theta}.$$9. The same construction supports verified blind quantum computing with hypergraph states and verified quantum supremacy demonstrations (Morimae et al., 2017).

Self-testing ideas can be used to certify the measurement devices themselves, not only the resource state. A self-guaranteed MBQC protocol assumes no prior-trusted devices—neither the measurement basis nor the entangled state—and combines Bell-pair self-tests with graph-state structure. For the graph-state protocol, the observable error satisfies

$F(\mathcal P,\mathcal M)\le 1-\varepsilon,$0

and the final output error is bounded by terms of order $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$1. The stated overhead is $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$2 copies of the graph-state resource (Hayashi et al., 2016).

Recent experimental work has moved trap-based verified MBQC into an on-chip regime. On the 20-qubit Quantinuum H1-1 trapped-ion QCCD device, qubit measurements and resets were used to realize verified MBQC patterns with up to 52 vertices. The protocol assumes honest classical control and secret-independent preparation noise, the latter justified by single-qubit tomography over eight YZ-plane preparation angles. In the deterministic case $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$3 with graph-coloring parameter $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$4, the tolerated test-failure threshold is

$F(\mathcal P,\mathcal M)\le 1-\varepsilon,$5

This is presented as the largest verified measurement-based quantum computation performed to date (Gustiani et al., 2024).

7. Dynamical, temporal, and practical limits

QMV has also acquired a temporal-logic form. For a quantum Markov chain

$F(\mathcal P,\mathcal M)\le 1-\varepsilon,$6

measurement-based linear-time temporal logic (MLTL) introduces atomic propositions $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$7 that are true when the measurement probability associated with $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$8 lies in an interval $F(\mathcal P,\mathcal M)\le 1-\varepsilon,$9. Agrawal et al.’s symbolic-dynamics methods for stochastic matrices are extended to quantum super-operators through eigenvalue analysis, periodic stability is characterized spectrally, and the period can be computed in $\mu(E)$00. Approximate model checking is then reduced to Büchi-automata operations. Applied to quantum walks, the method recovers the asymptotic advantage reported by Ambainis et al. and identifies additional quantum-only phenomena such as failure of classical symmetry properties (Guan et al., 2024).

Several practical limitations recur across the literature. Randomized-measurement fidelity estimation scales exponentially with subsystem size, roughly as

$\mu(E)$01

so present-day experiments can compare fidelities only for tens of qubits or for smaller reduced subsystems of much larger systems. This is substantially better than tomography, but it is not a full certification of an entire large quantum state. Likewise, no single verification metric covers all tasks: cross-device comparison, instrument tomography, delegated output verification, and device-independent certification use different figures of merit and different trust assumptions (Carrasco et al., 2021, Blume-Kohout et al., 20 Mar 2025).

The trust model is itself a decisive practical constraint. Measurement-device-independent protocols still rely on trusted state preparation on at least one side, whereas device-independent certification demands Bell-type nonclassical correlations and is correspondingly more stringent. Classical communication or quantum memory can sharply improve efficiency, but they do so in specific ways: LOCC improves the constant prefactor in $\mu(E)$02 state verification, while memory-assisted collective local measurements can reach global-optimum scaling for graph states (Zhang et al., 2020, Chen et al., 2023).

Experimental scope is correspondingly diverse. Hamiltonian learning, randomized measurements, and delegated-measurement protocols are presented as viable on trapped ions, optical lattices, Rydberg arrays, and superconducting devices. On-chip verification further removes the need for a separate quantum client, but only under explicit assumptions about honest control flow and secret-independent state-preparation noise. QMV therefore remains best understood not as a universal certification oracle, but as a layered family of measurement-centered tests whose validity depends on the hardware model, the adversarial model, and the quantity being certified (Carrasco et al., 2021, Gustiani et al., 2024).