Both statistics and quantum theory deal with prediction using probability. We will show that there can be established a connection between these two areas. This will at the same time suggest a new, less formalistic way of looking upon basic quantum theory.A total parameter space Φ, equipped with a group G of transformations, gives the mental image of some quantum system, in such a way that only certain components, functions of the total parameter φ can be estimated. Choose an experiment/question a, and get from this a parameter space Λa, perhaps after some model reduction compatible with the group structure. As in statistics, it is important always to distinguish between observations and parameters, in particular between (minimal) observations ta and state variables (parameters) λa. Let Ka be the L2 ‐space of functions of ta, and let Ha be the image of Ka under the expectation operator of the model. The measure determining the L2‐spaces is the invariant measure under the maximal subgroup which induces a transformation on Λa.There is a unitary connection between Ka and Ha, and then under natural conditions between Ha and Hb for a and b are different. Thus there exists a common Hilbert space H such that Ha equals UaH for a unitary Ua. In agreement with the common formulation of quantum mechanics, the vectors of H are taken to represent the states of the system. The state interpretation is then: An ideal experiment Ea has been performed, and the result of this experiment, disregarding measurement errors, is that the parameter λa is equal to some fixed number λka. This essentially statistical construction leads under natural assumptions to the basic axioms of quantum mechanics, and thus implies a new statistical interpretation of this traditionally very formal theory. The probabilities are introduced via Born’s formula, and this formula is proved from general, reasonable assumptions, essentially symmetry assumptions.The theory is illustrated by a simple macroscopic example, and by the example of a spin 12 particle. As a last example we show a connection to inference between related macroscropic experiments under symmetry. The result of this last example seems to have potential for being generalized considerably.
In classical statistics, a well known paradigm consists in establishing asymptotic equivalence between an experiment of i.i.d. observations and a Gaussian shift experiment, with the aim of obtaining optimal estimators in the former complicated model from the latter simpler model. In particular, a statistical experiment consisting of n i.i.d. observations from d-dimensional multinomial distributions can be well approximated by an experiment consisting of d−1 dimensional Gaussian distributions. In a quantum version of the result, it has been shown that a collection of n qudits (d-dimensional quantum states) of full rank can be well approximated by a quantum system containing a classical part, which is a d−1 dimensional Gaussian distribution, and a quantum part containing an ensemble of d(d−1)∕2 shifted thermal states. In this paper, we obtain a generalization of this result when the qudits are not of full rank. We show that when the rank of the qudits is r, then the limiting experiment consists of an r−1 dimensional Gaussian distribution and an ensemble of both shifted pure and shifted thermal states. For estimation purposes, we establish an asymptotic minimax result in the limiting Gaussian model. Analogous results are then obtained for estimation of a low rank qudit from an ensemble of identically prepared, independent quantum systems, using the local asymptotic equivalence result. We also consider the problem of estimation of a linear functional of the quantum state. We construct an estimator for the functional, analyze the risk and use quantum local asymptotic equivalence to show that our estimator is also optimal in the minimax sense.
The analysis of photon count data from the standard nitrogen vacancy (NV) measurement process is treated as a statistical inference problem. This has applications toward gaining better and more rigorous error bars for tasks such as parameter estimation (e.g. magnetometry), tomography, and randomized benchmarking. We start by providing a summary of the standard phenomenological model of the NV optical process in terms of Lindblad jump operators. This model is used to derive random variables describing emitted photons during measurement, to which finite visibility, dark counts, and imperfect state preparation are added. NV spin-state measurement is then stated as an abstract statistical inference problem consisting of an underlying biased coin obstructed by three Poisson rates. Relevant frequentist and Bayesian estimators are provided, discussed, and quantitatively compared. We show numerically that the risk of the maximum likelihood estimator is well approximated by the Cramér–Rao bound, for which we provide a simple formula. Of the estimators, we in particular promote the Bayes estimator, owing to its slightly better risk performance, and straightforward error propagation into more complex experiments. This is illustrated on experimental data, where quantum Hamiltonian learning is performed and cross-validated in a fully Bayesian setting, and compared to a more traditional weighted least squares fit.
This paper attempts to give an overview about sufficiency in the setting of quantum statistics. The basic concepts are treated in parallel to the measure theoretic case. It turns out that several classical examples and results have a noncommutative analogue. Some of the results are presented without proof (but with exact references) and the presentation is intended to be self-contained. The main examples discussed in the paper are related to the Weyl algebra and to the exponential family of states. The characterization of sufficiency in terms of quantum Fisher information is a new result.
In label-free microscopy, the experimenter's recurring question is not “Which named modality is best?” but “What operator should I use to estimate this physical quantity in this sample?” Named modalities are preset design families spanning illumination, sample interaction, post-sample optics, detector channels, readout, and dose allocation. Within a design, intrinsic minimax risk
≤ admissible-pipeline infimum
≤ selected deployed risk. Across designs, deployed-pipeline rankings do not determine Blackwell order. Given a declared physical labeling and prior, finite-dose prior-averaged squared risk splits exactly into within-component conditioning, between-component ambiguity, and posterior regret. Label refinement transfers an exact amount from conditioning to ambiguity. On a common-displacement branch, Theorem 2.2 bounds selected deployed risk by an asymptotic surrogate with a
1/N envelope of complete branch coefficients, dose-independent squared displacement, and decaying or persistent alias loss. The general fallback partition keeps
1/N separate while coupling displacement and ambiguity in §5.1's report-to-estimand block envelope. Neither corresponds termwise to the exact Bayes split. Finite upper certification adds a statistical remainder and, for approximate-generator inputs, a physical-fidelity allowance. In a coherent-displacement model with nonzero target and nuisance sensitivities, a Gram-proportional readout's target-information fraction equals the squared sine of their signal-phase separation. Any measurement on a
Kmode-mode coherent-displacement family jointly identifies at most
2Kmode displacement parameters on an open neighborhood. For two jointly unknown parameters at acute signal-phase separation
Δψ, simultaneous quadrature has penalty
2/(1+sinΔψH) relative to the best single-local-oscillator homodyne sequence, with
H the magnitude-symmetry factor. With nonsingular sensitivity Gram matrix, a local one-mode pure coherent-displacement model admits a homodyne allocation attaining the Holevo bound for every positive-definite quadratic weight. Under §0's detection laws, continuous rotational stabilizer generators of a rigid particle's detected response are rotationally Fisher-null. Independent fixed-registration fusion removes a null exactly when a positively allocated design supplies information in that direction. The mean fitted score cannot identify estimand displacement without an added relation, such as a calibrated regular-branch fidelity path or true-estimand identification from the full retained law. At a fixed endpoint with
0<F1eff<∞, the regular ambiguity exit is the generalized last-exit boundary of an exact pair Bayes floor above the local comparator. Once defined for a fixed witness pair, endpoint, and branch, the endpoint ambiguity diagnostic and variance–displacement crossover delimit an informative window but do not certify feasibility. Under a fixed design, a target-moving true-map fiber or other class with one complete detected law at every dose has intrinsic minimax risk at least its estimand image's squared Chebyshev radius,
h2/4 for a two-point gap
h. A pair sharing one common law within every candidate design at every dose makes that floor design-set-wide. Under transfer calibration, pair-kernel identification, and signed kernel normalization, a single-exchange calculation for homogeneous, isotropic, dilute, far-zone ensembles of weak, lossless, single-signed scalar scatterers and their co-polarized dipole extension certifies both signs of the pair-order ensemble-mean exact-minus-multislice forward-channel response-phase difference outside an endpoint allowance. At full retention, bounded variation of the pair correlation beyond a fixed exclusion radius bounds that allowance independently of the scalar radial pair-visibility window. Full-response means require omitted-mean bounds, whereas estimand-level means require the resolved phase-to-estimand bridge.
We address a broad class of optimization problems of finding quantum measurements, which includes the problems of finding an optimal measurement in the Bayes criterion and a measurement maximizing the average correct probability with a fixed rate of inconclusive results. Our approach can deal with any problem in which each of the objective and constraint functions is formulated by the sum of the traces of the multiplication of a Hermitian operator and a detection operator. We first derive dual problems and necessary and sufficient conditions for an optimal measurement. We also consider the minimax version of these problems and provide necessary and sufficient conditions for a minimax solution. Finally, for optimization problem having a certain symmetry, there exists an optimal solution with the same symmetry. Examples are shown to illustrate how our results can be used.