Proper Quantum Scoring Rules Overview

• Proper quantum scoring rules are defined via two complementary formulations—measurement-explicit and operator-valued—that guarantee truthful state reporting by leveraging convex potentials.
• Convex analysis and duality underpin the methodology, linking proper scores to spectral functionals and quantum Bregman divergences analogous to classical convex geometric results.
• The framework finds practical application in quantum state estimation and tomography, where minimized scoring regrets lead to improved sample complexity and enhanced precision.

Proper quantum scoring rules are the noncommutative analogue of classical proper scoring rules: they are scoring mechanisms for quantum-state reports under which truthful reporting of the underlying density operator is optimal in expectation. In the finite-dimensional setting, the literature develops two closely related formulations. One makes the measurement stage explicit, defining a quantum score as a pair consisting of a payoff rule and a report-dependent measurement, with expected score $S(\hat\rho;\rho)$ computed from the outcome distribution induced by measuring the true mixed state $\rho$ after a report $\hat\rho$ (Frongillo, 2022). The other emphasizes an operator-valued state-level rule $\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$, evaluated directly by trace pairing as $\operatorname{Tr}(\rho\,\mathbf S(\sigma))$, and derives properness from a convex unitarily invariant Quantum Value Functional $V_Q(\rho)=\operatorname{Tr}(f(\rho))$ (AlMasri, 6 May 2026). Across both formulations, the central structural result is a quantum version of the classical convex-analytic correspondence between proper scores, supporting hyperplanes, and Bregman-type divergences.

1. Concept and formal definitions

In the measurement-explicit formulation, a quantum mixed state is a density matrix $\rho\in \mathrm{Dens}(X)$, where $X=\mathbb C^n$, and a quantum score is a pair $S=(s,\mu)$. Here $\mu(\hat\rho)$ is a POVM chosen as a function of the reported state $\rho$0, and the expected score is

$\rho$1

The score is called truthful if

$\rho$2

and strictly truthful if the inequality is strict whenever $\rho$3 (Frongillo, 2022). In the terminology used elsewhere, truthful and strictly truthful are the quantum analogues of proper and strictly proper scoring rules.

The operator-valued formulation replaces the classical scalar score $\rho$4 with an operator-valued rule

$\rho$5

defined on density operators $\rho$6 over a finite-dimensional complex Hilbert space $\rho$7. The realized scalar score against the true state $\rho$8 is

$\rho$9

and properness is the condition

$\hat\rho$0

The 2026 formulation identifies the primitive convex object as a Quantum Value Functional $\hat\rho$1 that is convex, unitarily invariant, and lower semicontinuous; by spectral calculus,

$\hat\rho$2

for a convex $\hat\rho$3 (AlMasri, 6 May 2026).

A notable conceptual difference from the classical case is that a density matrix does not directly generate an observable classical outcome. The 2022 framework therefore includes the measurement rule as part of the contract, while the 2026 framework packages the dependence on the report into the operator-valued score and evaluates it by trace pairing against the true state. This suggests two complementary viewpoints: one operational and measurement-level, the other convex-analytic and state-level.

2. Convex-analytic structure and duality

A central result of the finite-dimensional theory is that expected score is affine in the true state. In the measurement-based framework,

$\hat\rho$4

so every quantum score induces a linear functional of $\hat\rho$5 for fixed report $\hat\rho$6 (Frongillo, 2022). This is the quantum analogue of the classical fact that expected score is linear in the true distribution.

The main characterization theorem in that framework states that a quantum score $\hat\rho$7 is truthful if and only if there exists a convex function $\hat\rho$8 and a choice of subgradients $\hat\rho$9 such that

$\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$0

Strict truthfulness corresponds to strict convexity. The truthfulness gap is then the generalized Bregman divergence

$\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$1

(Frongillo, 2022).

The 2026 theory states an analogous correspondence in terms of Quantum Value Functionals. For differentiable spectral functionals,

$\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$2

and the induced quantum Bregman divergence is

$\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$3

Convexity yields $\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$4, and the paper uses the identity

$\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$5

to encode properness (AlMasri, 6 May 2026).

The duality theorem in the 2026 formulation centers on the correspondence

$\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$6

The explicit theorem establishes the forward direction: if $\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$7 is closed, convex, and unitarily invariant, then $\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$8 is proper. The abstract describes this as a “complete duality theory,” but the detailed statement presented there formalizes the forward construction most explicitly (AlMasri, 6 May 2026).

3. Spectral structure, operator convexity, and divergences

For spectral functionals $\mathbf S:D(H)\to \mathcal L(H)_{\rm Herm}$9, the Fréchet derivative can be computed through the Daleckii–Krein formula. If $\operatorname{Tr}(\rho\,\mathbf S(\sigma))$0, then

$\operatorname{Tr}(\rho\,\mathbf S(\sigma))$1

with first divided difference matrix

$\operatorname{Tr}(\rho\,\mathbf S(\sigma))$2

Identifying the Hilbert–Schmidt gradient yields $\operatorname{Tr}(\rho\,\mathbf S(\sigma))$3 (AlMasri, 6 May 2026).

The operator-convex case is singled out because it provides stronger order-theoretic and information-theoretic structure. If $\operatorname{Tr}(\rho\,\mathbf S(\sigma))$4 is operator convex, then

$\operatorname{Tr}(\rho\,\mathbf S(\sigma))$5

in Löwner order. Under the additional symmetry condition that

$\operatorname{Tr}(\rho\,\mathbf S(\sigma))$6

the induced Bregman divergence coincides with Petz’s quantum $\operatorname{Tr}(\rho\,\mathbf S(\sigma))$7-divergence,

$\operatorname{Tr}(\rho\,\mathbf S(\sigma))$8

(AlMasri, 6 May 2026).

The 2022 theory isolates a particularly elegant subclass, spectral scores. Writing a reported state as

$\operatorname{Tr}(\rho\,\mathbf S(\sigma))$9

a spectral score measures in the reported eigenbasis,

$V_Q(\rho)=\operatorname{Tr}(f(\rho))$0

and pays according to a classical scoring rule on the eigenvalue vector,

$V_Q(\rho)=\operatorname{Tr}(f(\rho))$1

Its expected score is

$V_Q(\rho)=\operatorname{Tr}(f(\rho))$2

The paper proves that a spectral quantum score is truthful if and only if the underlying classical scoring rule $V_Q(\rho)=\operatorname{Tr}(f(\rho))$3 is proper, and that a finite truthful quantum score is equivalent to a spectral score precisely when it is unitary-invariant (Frongillo, 2022).

These two lines of work converge on the same structural theme: unitary symmetry singles out spectral convex potentials and spectral proper scores, while noncommutativity forces one to distinguish ordinary convexity from operator convexity and to treat divergence notions that coincide classically as inequivalent in the quantum setting.

4. Relation to classical proper scoring rules

The quantum theories preserve the classical pattern with systematic replacements. The classical simplex of distributions becomes the convex set $V_Q(\rho)=\operatorname{Tr}(f(\rho))$4 or $V_Q(\rho)=\operatorname{Tr}(f(\rho))$5; scalar probabilities become density matrices; Euclidean pairings become Hilbert–Schmidt trace pairings; and classical gradients or subgradients become Fréchet derivatives or matrix subgradients (AlMasri, 6 May 2026).

In the commuting case, the quantum theory reduces to the classical one. If $V_Q(\rho)=\operatorname{Tr}(f(\rho))$6 and $V_Q(\rho)=\operatorname{Tr}(f(\rho))$7 commute, they are simultaneously diagonalizable: $V_Q(\rho)=\operatorname{Tr}(f(\rho))$8 Then

$V_Q(\rho)=\operatorname{Tr}(f(\rho))$9

and

$\rho\in \mathrm{Dens}(X)$0

The corresponding quantum Bregman divergence becomes

$\rho\in \mathrm{Dens}(X)$1

which is exactly the classical Bregman divergence (AlMasri, 6 May 2026).

The connection to classical convex-geometric results is not merely formal. The classical paper “Proper Scoring Rules and Domination” generalizes a theorem of Predd et al. to non-additive scoring rules on a finite state space, proving that under strict propriety and a closure or continuity condition, every non-probabilistic credence is pointwise dominated by some probabilistic one (Pruss, 2021). That work is entirely classical, but its relevance to the quantum setting lies in its emphasis that additivity is not essential: propriety can be studied through convex geometry of the global score set rather than coordinatewise decompositions. The 2022 and 2026 quantum theories likewise rely on convex structure, linearity in the true state, and non-additive state-level scoring objects rather than classical eventwise additivity.

A further distinction is that the 2022 framework makes report-dependent measurement explicit. Classically, one directly observes an outcome drawn from the true distribution; quantum mechanically, the report may also determine how the outcome distribution is generated. This report-dependent measurement is the main conceptual novelty of the operational formulation (Frongillo, 2022).

5. Estimation, tomography, and minimax theory

A major extension of proper quantum scoring rules is their use as loss functions for quantum state estimation. In the 2026 framework, the loss of estimating $\rho\in \mathrm{Dens}(X)$2 by $\rho\in \mathrm{Dens}(X)$3 is

$\rho\in \mathrm{Dens}(X)$4

and the corresponding minimax risk over $\rho\in \mathrm{Dens}(X)$5 copies is

$\rho\in \mathrm{Dens}(X)$6

This recasts tomography as minimax estimation under incentive-compatible scoring-rule regret (AlMasri, 6 May 2026).

The asymptotic lower bound derived there is the Quantum Cramér–Rao–McCarthy Bound,

$\rho\in \mathrm{Dens}(X)$7

The ingredients are the spectral curvature $\rho\in \mathrm{Dens}(X)$8 and the Symmetric Logarithmic Derivative Quantum Fisher Information. If $\rho\in \mathrm{Dens}(X)$9, then

$X=\mathbb C^n$0

The theorem therefore links minimax risk under McCarthy-type incentives to local information geometry and the curvature of the generating functional (AlMasri, 6 May 2026).

The mechanism of the bound is a second-order expansion of the Bregman divergence. For $X=\mathbb C^n$1,

$X=\mathbb C^n$2

with second variation

$X=\mathbb C^n$3

In the asymptotic approximation used there, this Hessian is represented by $X=\mathbb C^n$4 in the eigenbasis of $X=\mathbb C^n$5, so local risk behaves like

$X=\mathbb C^n$6

A sharper curvature therefore corresponds to harsher local penalization of estimation error (AlMasri, 6 May 2026).

For the logarithmic generator $X=\mathbb C^n$7, one has $X=\mathbb C^n$8, hence $X=\mathbb C^n$9 on the support of $S=(s,\mu)$0. The paper states that asymptotic equality is achieved in this case by the maximum likelihood estimator when $S=(s,\mu)$1 is nearly pure or the parameterization is locally orthogonal in the QFI metric (AlMasri, 6 May 2026).

6. Examples, resources, and scope

The canonical example is the quantum logarithmic score. In the spectral functional approach, for

$S=(s,\mu)$2

the induced score is

$S=(s,\mu)$3

and the truthful expected score is

$S=(s,\mu)$4

where $S=(s,\mu)$5 is von Neumann entropy (AlMasri, 6 May 2026). In the 2022 spectral-score formulation, the corresponding expected score is

$S=(s,\mu)$6

the truthful value is $S=(s,\mu)$7, and the truthfulness gap is

$S=(s,\mu)$8

giving an operational interpretation of von Neumann relative entropy as the regret from misreporting under the logarithmic score (Frongillo, 2022).

A quadratic analogue also appears naturally. With $S=(s,\mu)$9,

$\mu(\hat\rho)$0

which is the Hilbert–Schmidt quadratic member of the spectral class (AlMasri, 6 May 2026). The 2022 paper also gives a Brier-type realization with expected score

$\mu(\hat\rho)$1

which equals $\mu(\hat\rho)$2 and is therefore strictly truthful (Frongillo, 2022).

The qubit coherence example offers a concrete operational illustration. For

$\mu(\hat\rho)$3

a classical forecaster restricted to the computational $\mu(\hat\rho)$4-basis effectively sees $\mu(\hat\rho)$5, while a quantum forecaster measuring in the $\mu(\hat\rho)$6-basis identifies the state perfectly. Under the logarithmic score, the truthful quantum report yields

$\mu(\hat\rho)$7

whereas the best diagonal classical report gives

$\mu(\hat\rho)$8

The forecasting gap is therefore

$\mu(\hat\rho)$9

which matches the relative entropy of coherence

$\rho$00

(AlMasri, 6 May 2026).

More generally, the 2026 work assigns operational value to quantum resources in proper-scoring-rule risk units. For noncommuting output channels it proves a separation between classical fixed-basis strategies and fully quantum joint-measurement strategies: $\rho$01 and states that quantum strategies can reduce sample complexity by a factor $\rho$02 in the worst case. For the logarithmic score, it also gives the coherence-risk tradeoff

$\rho$03

provided the dephased state $\rho$04 is the information projection of $\rho$05 onto the diagonal submanifold with respect to the relevant Bregman geometry. The paper summarizes qualitative resource scalings as coherence $\rho$06, entanglement $\rho$07, and adaptivity $\rho$08 (AlMasri, 6 May 2026).

The present theory is explicitly finite-dimensional and state-based. Its assumptions include density operators as the report space, unitarily invariant convex generators, and—when asymptotic estimation results are derived—i.i.d. copies, smooth local parameterization, differentiability, $\rho$09, and LAN/Holevo/SLD asymptotic machinery (AlMasri, 6 May 2026). The 2022 framework likewise remains finite-dimensional, uses finite POVMs or PVMs, and permits extended-real scores to accommodate logarithmic boundary behavior (Frongillo, 2022). Reported limitations include asymptotic rather than exact nonasymptotic analysis for much of the minimax theory, commutation symmetry conditions in some divergence-identification statements, restriction to states rather than channels or processes, omission of computational constraints, and the absence of an infinite-dimensional theory (AlMasri, 6 May 2026). Open directions highlighted in the literature include extensions to quantum channels, more systematic treatment of adaptive protocols, experimental validation on near-term devices, strategic interaction among multiple quantum agents, non-i.i.d. settings, infinite-dimensional systems, and robustness under noisy channels using optimized $\rho$10-divergences (AlMasri, 6 May 2026).