Introduction
Quantum statistics studies inferential questions arising in quantum-mechanical experiments and has become central to uncertainty quantification in emerging quantum technologies. Its objectives, including estimation and hypothesis testing, mirror those of classical statistics, although the underlying models differ. Quantum states, which play the role of probability distributions in classical statistics, are represented by operators that need not commute. Further, statistical procedures involve both the design of quantum measurements and the subsequent analysis of the resulting classical data. Consequently, characterizing optimal procedures generally requires joint optimization over measurements and decision rules.
The decision-theoretic formulation of quantum inference dates back to the foundational work of Helstrom and Holevo [22 , 23 ] . The rapid development of quantum technologies has renewed interest in this framework. Over the last two decades, quantum statistical methods have become increasingly important in applications such as quantum state tomography, quantum imaging, quantum metrology, and quantum sensing. These developments have raised fundamental questions concerning the limits of statistical inference in quantum systems, including the determination of efficiency bounds and minimax risks and the construction of procedures that attain them.
Inference in quantum statistical models
A quantum statistical model is a family Q = { ρ θ : θ ∈ Θ } \mathcal{Q} = \{\rho_\theta: \theta\in\Theta\} Q = { ρ θ : θ ∈ Θ } , where the parameter space Θ \Theta Θ is a subset of a Euclidean space and each ρ θ \rho_\theta ρ θ is a density operator, that is, a positive semidefinite Hermitian operator of unit trace. We briefly outline the principal inferential problems arising in such models.
Given measurement outcomes from repeated preparations of a quantum system, one seeks to test statistical hypotheses concerning the underlying state or its unknown parameters. The hypotheses may also concern structural properties of the state, such as purity, separability, entanglement, low-rank structure, or membership in a specified family. Testing such properties without reconstructing the full state has attracted considerable attention in quantum information theory [36 ] .
A standard example of a quantum hypothesis-testing problem is quantum state discrimination, in which the objective is to distinguish between specified quantum states through a suitable measurement and decision rule. In the asymptotic setting of many independent, identically prepared copies, a central question concerns the optimal exponential rates at which testing errors decay. For discrimination between two fixed states, the quantum Stein lemma and the quantum Chernoff bound characterize these rates in asymmetric and symmetric testing, respectively [41 , 3 , 4 , 34 , 37 ] . More broadly, quantum decision theory provides a framework for studying the choice of measurements and decision rules, extending the classical theory of statistical experiments and decision procedures to noncommutative statistical models [22 , 23 ] .
Within this broader framework, the present paper considers hypotheses about unknown parameters, with an emphasis on the asymptotic efficiency of testing procedures. At a prescribed significance level, a central criterion is local power, the ability to detect alternatives that approach the null hypothesis as the number of repeated preparations increases. This perspective assesses the sensitivity of tests to small departures from the null and provides a basis for comparing their efficiency.
The study of local power is naturally related to quantum parameter estimation, since the precision of an estimator can govern the sensitivity of a test constructed from it. Quantum analogues of the classical Cramér–Rao inequality provide lower bounds on the covariance matrices of locally unbiased estimators and serve as benchmarks for attainable precision [5 , 22 , 23 ] . These bounds underpin the study of asymptotic efficiency in quantum estimation and play a central role in quantum metrology, imaging, and sensing. As in classical statistics, the connection between estimation and testing operates in both directions. Testing bounds can yield lower bounds on estimation risk, while asymptotically normal estimators can be used, under suitable regularity conditions and with consistent covariance estimation, to construct Wald-type tests. We study optimal testing through local asymptotic normality, building on its use in establishing the efficiency of quantum estimators [14 ] .
Local asymptotic normality in quantum statistical models
In classical statistics, one approach to establishing optimality is to approximate a complicated statistical experiment by a simpler limit experiment in the sense of Le Cam [33 ] . An important instance is local asymptotic normality (LAN). Under LAN, a regular i.i.d. model localized around a fixed parameter value has a Gaussian shift limit. Such approximations allow optimality bounds derived in the Gaussian limit experiment to be carried over to the original model. A procedure whose asymptotic risk or power attains the corresponding bound is then asymptotically optimal in the original problem.
Beyond local approximations, asymptotic equivalence has also been established for a variety of nonparametric models. Under suitable smoothness conditions, both nonparametric regression [6 ] and density estimation [39 ] are asymptotically equivalent to Gaussian white noise experiments. These equivalence results provide a unified framework for deriving minimax rates, efficiency bounds, and optimal procedures across seemingly different statistical models. LAN techniques have also proved useful in the study of functional estimation, where suitably regular functionals can often be analyzed through associated local parametric submodels [24 ] .
The quantum analogue of this theory, known as quantum local asymptotic normality (q-LAN), was developed in [19 , 27 , 45 , 13 ] . These results show that, under suitable regularity conditions, a collection of n n n independent, identically prepared finite-dimensional quantum systems can be approximated locally around a fixed state by a quantum Gaussian shift experiment. A quantum analogue of nonparametric asymptotic equivalence was established in [7 ] , where local ensembles of pure states in an infinite-dimensional Hilbert space were shown to be asymptotically equivalent to coherent-state models, providing a quantum counterpart of the classical Gaussian white noise experiment. Related asymptotic approximations have also been established for quantum Markov models [20 , 21 , 16 ] and quantum time series models [38 ] . Quantum LAN has become a fundamental tool for deriving asymptotic lower bounds and constructing procedures that attain them [14 , 32 , 31 , 18 ] . Much of these works concern parameter estimation. Here we use q-LAN to establish analogous optimality results for hypothesis testing.
Three related approaches to Gaussian approximation are relevant here. One uses a local expansion of the log-likelihood ratio in terms of a central sequence Δ n \Delta_n Δ n . The others study strong convergence in the Le Cam distance and weak convergence of statistical experiments. Quantum counterparts of these approaches were developed in [45 ] , [27 ] , and [19 ] , respectively. We use the likelihood-ratio formulation of [45 ] , together with the quantum asymptotic representation theorem and the quantum analogue of Le Cam’s third lemma. The representation theorem provides the comparison with the Gaussian experiment, and the third lemma gives the limiting laws of our test statistics under local alternatives. Their classical counterparts play the same roles in the analysis of testing efficiency in Chapter 15 of [44 ] .
Efficiency of tests in classical statistics
In the classical setup, Le Cam’s asymptotic representation theorem [44 Theorem 7.10] provides a link between the original statistical experiment and its Gaussian limit. We give the precise statement in Appendix § 11 . Roughly speaking, if a sequence of experiments converges weakly to a limit experiment and a sequence of statistics has limiting distributions under local alternatives, then a single, possibly randomized procedure in the limit experiment realizes these distributions.
This observation underlies several results in asymptotic optimality theory. In estimation, it leads to the local asymptotic minimax theorem, which gives a lower bound on local asymptotic risk in terms of the minimax risk in the Gaussian shift experiment. In testing, the same principle yields an upper bound on local power. For a one-sided hypothesis about a differentiable scalar functional, Theorem 15.4 of [44 ] bounds the asymptotic power of any sequence of level-α \alpha α tests at each fixed local alternative by the power of the corresponding Gaussian test. Addendum 15.5 gives a sufficient condition for attainment.
Fujiwara and Yamagata [14 ] proved a quantum version of the asymptotic representation theorem and used it to establish efficiency bounds for estimators in quantum models. Our goal is to develop an analogous framework for quantum hypothesis testing. In particular, we use q-LAN to derive asymptotically optimal tests by analyzing the corresponding quantum Gaussian limit experiment.
Our contributions
Our main contributions are as follows.
We first solve the problem of testing a linear functional of the mean in a quantum Gaussian model with known covariance. A shift of the mean that leaves the linear functional unchanged preserves the hypotheses. Using this symmetry, we show that restricting attention to tests invariant under these nuisance shifts preserves the greatest attainable worst-case power. The argument builds on the quantum Hunt–Stein method used by Kumagai and Hayashi for Gaussian tests with nuisance parameters [29 ] . In the present setting, every invariant test can be implemented by measuring the linear combination of Gaussian coordinates corresponding to the functional and applying a classical rule to its outcome. The outcome has a classical Gaussian distribution, so the upper-tail test of its mean gives the sharp one-sided power bound, even when competing tests can use arbitrary quantum measurements. The two-tailed normal test gives the corresponding bound within the class of unbiased tests.
We next transfer these bounds to q-LAN models whose central sequences are asymptotically D D D -invariant. The asymptotic representation theorem relates the power of tests under local alternatives to the power of tests in the corresponding quantum Gaussian experiment. Applying the Gaussian bound to the local linear approximation of a differentiable functional gives an upper bound on worst-case local power. For two-sided hypotheses, this comparison is within the unbiased class. In the independent-copy model, a test based on a weighted combination of the symmetric logarithmic derivatives, the
quantum analogues of the usual likelihood scores, attains this bound in local asymptotic power.
The weighted combination of observables on which the optimal test is based, typically depends on the unknown state, even when the null value of the functional is known. For smooth independent-copy models of constant rank, we give a two-stage adaptive procedure. A preliminary measurement on a vanishing fraction of the copies provides an estimate of the parameter. We use that estimate to choose the observable measured separately on each remaining copy. The average of the resulting observations supplies a one-step correction to the preliminary estimate of the functional. The resulting Wald tests have the same limiting power as the test designed at the true reference point, uniformly over compact sets of local perturbations. This construction is related to Fujiwara’s work on adaptive quantum estimation [11 ] .
We apply our results to testing smooth spectral functionals at states with distinct positive eigenvalues and a nonvanishing functional derivative. A natural measurement for this problem is weak Schur sampling (WSS), which originates in the study of symmetries of product states [2 , 28 ] and has been used in spectrum testing and entropy estimation [8 , 40 , 1 ] . The sufficiency result of [30 ] shows that WSS retains all information available through unitarily invariant measurements, although this need not imply optimality in restricted parametric models. We identify two distinct sources of efficiency loss: suboptimal processing of the WSS outcome and information discarded by the measurement itself. In particular, we characterize the best limiting power among linear statistics of the WSS fluctuations having the correct first-order mean in every local direction, and show how this improves upon the usual plug-in test. We then compare this power with that of the scalar score test, which performs at least as well and attains the local maximin bound under our q-LAN assumptions. We derive necessary and sufficient conditions for equality in both comparisons. These conditions have a geometric interpretation in terms of the structure of the tangent space, closely related to the projection-based approach to quantum semiparametric estimation developed in [43 ] .
Organization of the paper
§ 2 introduces the notation and background on quantum statistical models. We define q-LAN, present the version of the quantum asymptotic representation theorem used in our optimality proofs, and compare classical and quantum hypothesis testing. § 3 gives the optimal power bounds in quantum Gaussian models, and § 4 derives the corresponding bounds for q-LAN models, including the i.i.d. case. § 5 develops an adaptive version of the classical Wald test that attains the same limiting power while § 6 applies these results to tests of spectral functionals. § 7 discusses extensions of our framework. The proofs and auxiliary results are collected in the appendices.
Notation
For a vector v v v in a separable Hilbert space H \mathcal{H} H , we write ∣ v ⟩ \lvert v\rangle ∣ v ⟩ for the vector (the ket) and ⟨ v ∣ \langle v\rvert ⟨ v ∣ for the corresponding element v ∗ v^{*} v ∗ of the dual space H ∗ \mathcal{H}^{*} H ∗ (the bra). The inner product ⟨ u ∣ v ⟩ ∈ C \langle u\lvert v\rangle\in\mathbb{C} ⟨ u ∣ v ⟩ ∈ C is linear in its second argument and conjugate-linear in its first. The rank-one operator M : = ∣ u ⟩ ⟨ v ∣ M := \lvert u\rangle\langle v\rvert M := ∣ u ⟩ ⟨ v ∣ acts by M ∣ w ⟩ = ⟨ v ∣ w ⟩ ∣ u ⟩ M\lvert w\rangle= \langle v\lvert w\rangle\lvert u\rangle M ∣ w ⟩ = ⟨ v ∣ w ⟩ ∣ u ⟩ . For an operator A A A , we also write ⟨ u ∣ A ∣ v ⟩ \langle u\lvert A\lvert v\rangle ⟨ u ∣ A ∣ v ⟩ for ⟨ u ∣ A v ⟩ \langle u\lvert A v\rangle ⟨ u ∣ A v ⟩ .
The space of bounded linear operators on H \mathcal{H} H is denoted by L ( H ) \mathcal{L}(\mathcal{H}) L ( H ) . We write T 1 ( H ) ⊂ L ( H ) \mathcal{T}_1(\mathcal{H}) \subset\mathcal{L}(\mathcal{H}) T 1 ( H ) ⊂ L ( H ) for the trace-class operators, equipped with the trace norm
∥ A ∥ 1 : = Tr [ ( A † A ) 1 / 2 ] . \lVert A\rVert_1 := \operatorname{Tr}\left[(A^\dagger A)^{1/2}\right]. ∥ A ∥ 1 := Tr [ ( A † A ) 1/2 ] .
We write ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ for the Hilbert space norm and 1 \mathbf{1} 1 for the identity operator, with the underlying space understood from context. For probability measures μ \mu μ and ν \nu ν , we use ∥ μ − ν ∥ T V : = sup B ∣ μ ( B ) − ν ( B ) ∣ \lVert\mu-\nu\rVert_{\mathrm{TV}} := \sup_B\lvert\mu(B)-\nu(B)\rvert ∥ μ − ν ∥ TV := sup B ∣ μ ( B ) − ν ( B )∣ , where the supremum is over measurable sets B B B . For probability mass functions p p p and q q q on a countable space, ∥ p − q ∥ 1 : = ∑ x ∣ p ( x ) − q ( x ) ∣ = 2 ∥ p − q ∥ T V \lVert p-q\rVert_1 := \sum_x\lvert p(x)-q(x)\rvert= 2\lVert p-q\rVert_{\mathrm{TV}} ∥ p − q ∥ 1 := ∑ x ∣ p ( x ) − q ( x )∣ = 2 ∥ p − q ∥ TV . We write a ∨ b a\vee b a ∨ b and a ∧ b a\wedge b a ∧ b for max ( a , b ) \max(a,b) max ( a , b ) and min ( a , b ) \min(a,b) min ( a , b ) , respectively, and a + : = a ∨ 0 a_+ := a\vee0 a + := a ∨ 0 . The symbols ⌊ a ⌋ \lfloor a\rfloor ⌊ a ⌋ and ⌈ a ⌉ \lceil a\rceil ⌈ a ⌉ denote the largest integer less than or equal to a a a and the smallest integer greater than or equal to a a a , respectively. We write a n ≍ b n a_n\asymp b_n a n ≍ b n if c < lim inf n ( a n / b n ) ≤ lim sup n ( a n / b n ) < C c < \liminf_n(a_n/b_n) \leq\limsup_n(a_n/b_n) < C c < lim inf n ( a n / b n ) ≤ lim sup n ( a n / b n ) < C for some constants c , C > 0 c,C>0 c , C > 0 . Throughout, c c c and C C C denote generic constants. Finally, Φ \Phi Φ denotes the standard Gaussian distribution function, and Φ ( a , b ) \Phi(a,b) Φ ( a , b ) denotes the probability that a standard Gaussian random variable lies in [ a , b ] [a,b] [ a , b ] .
Preliminaries
States, Measurements and Observables
A state of a quantum system is described by a self-adjoint operator ρ \rho ρ on a complex Hilbert space H \mathcal{H} H , which is positive (ρ ≥ 0 \rho\geq0 ρ ≥ 0 ) and normalized to Tr ( ρ ) = 1 \operatorname{Tr}(\rho)=1 Tr ( ρ ) = 1 (a density operator). A state is called pure if it is of the form ρ = ∣ ψ ⟩ ⟨ ψ ∣ \rho=\lvert\psi\rangle\langle\psi\rvert ρ = ∣ ψ ⟩ ⟨ ψ ∣ , otherwise it is called a mixed state . We denote the set of states by S ( H ) \mathcal{S}(\mathcal{H}) S ( H ) . It can be shown that S ( H ) ⊂ T 1 ( H ) \mathcal{S}(\mathcal{H})\subset\mathcal{T}_1(\mathcal{H}) S ( H ) ⊂ T 1 ( H ) .
Data on a quantum system are obtained from observables which are self-adjoint operators in the Hilbert space H \mathcal{H} H . If S S S is a self-adjoint operator in H \mathcal{H} H with spectral decomposition S = ∑ j λ j Π j S=\sum_j\lambda_j\Pi_j S = ∑ j λ j Π j where Π j \Pi_j Π j are projectors, then a measurement generates a discrete random variable X S X_S X S taking values in the set of eigenvalues { λ 1 , λ 2 , … } \{\lambda_1,\lambda_2,\ldots\} { λ 1 , λ 2 , … } with probabilities p j = Tr ( ρ ⋅ Π j ) p_j=\operatorname{Tr}(\rho\cdot\Pi_j) p j = Tr ( ρ ⋅ Π j ) . Whenever the first absolute moment is finite, the expectation of X S X_S X S under the state ρ \rho ρ is given by the Born-von Neumann postulate :
E ρ X S = ∑ j λ j Tr [ ρ Π j ] = Tr [ ρ S ] . (1) \mathbb{E}_{\rho}X_S=\sum_j\lambda_j\operatorname{Tr}[\rho\Pi_j]=\operatorname{Tr}[\rho S]. \tag*{(1)} E ρ X S = j ∑ λ j Tr [ ρ Π j ] = Tr [ ρS ] . (1)
More generally, a measurement with outcomes in a measurable space ( Ω , B ) (\Omega,\mathcal{B}) ( Ω , B ) is described by a positive operator-valued measure (POVM).
Definition 1. A positive operator valued measure (POVM) is a map M : B → L ( H ) M:\mathcal{B}\to\mathcal{L}(\mathcal{H}) M : B → L ( H ) having the following properties
positivity: M ( B ) ≥ 0 M(B)\geq0 M ( B ) ≥ 0 for all events B ∈ B B\in\mathcal{B} B ∈ B (hence M ( B ) M(B) M ( B ) is self-adjoint)
σ \sigma σ -additivity: M ( ⋃ i B i ) = ∑ i M ( B i ) M(\bigcup_iB_i)=\sum_iM(B_i) M ( ⋃ i B i ) = ∑ i M ( B i ) for any countable set of mutually disjoint events B i B_i B i (here the convergence is in the weak operator topology of L ( H ) \mathcal{L}(\mathcal{H}) L ( H ) )
normalization: M ( Ω ) = 1 M(\Omega)=\mathbf{1} M ( Ω ) = 1 .
The simplest example of a POVM is given by an effect E E E , which is a bounded operator satisfying $0\leq E\leq 1$. Indeed { E , 1 − E } \{E,1-E\} { E , 1 − E } forms a two-outcome POVM. If the operators $M(B)
are also orthogonal projections, i.e. M ( A ) 2 = M ( A ) M(A)^2=M(A) M ( A ) 2 = M ( A ) and M ( B ) M ( A ) = 0 M(B)M(A)=0 M ( B ) M ( A ) = 0 when A ∩ B = ∅ A\cap B=\varnothing A ∩ B = ∅ , then it is called a simple measurement . The collection of projectors { Π j } \{\Pi_j\} { Π j } in the spectral decomposition S = ∑ j λ j Π j S=\sum_j\lambda_j\Pi_j S = ∑ j λ j Π j is an example of a simple measurement. The outcome of the measurement has probability distribution
P ρ ( B ) = Tr ( ρ M ( B ) ) , B ∈ B . (2) P_\rho(B)=\operatorname{Tr}(\rho M(B)),\qquad B\in\mathcal{B}. \tag*{(2)} P ρ ( B ) = Tr ( ρM ( B )) , B ∈ B . (2)
The spectral theorem associates with every self-adjoint operator S S S , with dense domain D ( S ) ⊂ H D(S)\subset\mathcal{H} D ( S ) ⊂ H , a projection-valued spectral measure M M M such that
S = ∫ σ ( S ) x M ( d x ) , S=\int_{\sigma(S)}x\,M(dx), S = ∫ σ ( S ) x M ( d x ) ,
in the sense of the spectral calculus. Here σ ( S ) \sigma(S) σ ( S ) is the spectrum of S S S and M M M is a POVM, also called spectral measure associated with the operator S S S . When S S S is an observable with a continuous spectrum, it generates a continuous random variable X S X_S X S with probabilities given by (2) . Whenever the first absolute moment is finite,
E ρ [ X S ] = ∫ σ ( S ) x Tr { ρ M ( d x ) } , E_\rho[X_S]=\int_{\sigma(S)}x\,\operatorname{Tr}\{\rho M(dx)\}, E ρ [ X S ] = ∫ σ ( S ) x Tr { ρM ( d x )} ,
which we also denote by Tr ( ρ S ) \operatorname{Tr}(\rho S) Tr ( ρS ) .
Quantum Gaussian states and convergence in distribution
Here we introduce the Gaussian models which arise as limiting experiments under q-LAN. We begin with the usual Hilbert-space representation of quantum Gaussian states and subsequently explain how classical Gaussian coordinates are incorporated when the commutation matrix is degenerate.
To define a one-mode Gaussian state, consider the Hilbert space H = L 2 ( R ) \mathcal{H}=L^2(\mathbb{R}) H = L 2 ( R ) and the position and momentum operators
( Q f ) ( x ) = x f ( x ) , ( P f ) ( x ) = − i d d x f ( x ) , (Qf)(x)=xf(x),\qquad(Pf)(x)=-i\frac{d}{dx}f(x), ( Q f ) ( x ) = x f ( x ) , ( P f ) ( x ) = − i d x d f ( x ) ,
defined on appropriate dense domains. They satisfy the canonical commutation relation [ Q , P ] = i I [Q,P]=iI [ Q , P ] = i I . For u = ( u 1 , u 2 ) T ∈ R 2 u=(u_1,u_2)^T\in\mathbb{R}^2 u = ( u 1 , u 2 ) T ∈ R 2 , the linear combination Z u = u 1 Q + u 2 P Z_u=u_1Q+u_2P Z u = u 1 Q + u 2 P is a self-adjoint observable. Given a density operator ρ \rho ρ on H \mathcal{H} H , its quantum characteristic function is
W ~ ρ ( u ) = Tr ρ e i Z u . \widetilde{W}_\rho(u)=\operatorname{Tr}\rho e^{iZ_u}. W ρ ( u ) = Tr ρ e i Z u .
More generally, for an m m m -mode system we consider H m = ⨂ j = 1 m L 2 ( R ) \mathcal{H}_m=\bigotimes_{j=1}^m L^2(\mathbb{R}) H m = ⨂ j = 1 m L 2 ( R ) and the canonical observables
R = ( Q 1 , P 1 , … , Q m , P m ) T . R=(Q_1,P_1,\ldots,Q_m,P_m)^T. R = ( Q 1 , P 1 , … , Q m , P m ) T .
Here Q j Q_j Q j and P j P_j P j act on the j j j th tensor factor and satisfy
[ Q j , P k ] = i δ j k I , [ Q j , Q k ] = [ P j , P k ] = 0. (3) [Q_j,P_k]=i\delta_{jk}I,\qquad[Q_j,Q_k]=[P_j,P_k]=0. \tag*{(3)} [ Q j , P k ] = i δ j k I , [ Q j , Q k ] = [ P j , P k ] = 0. (3)
Writing R = ( R 1 , … , R 2 m ) T R=(R_1,\ldots,R_{2m})^T R = ( R 1 , … , R 2 m ) T , these relations may be expressed as
[ R j , R k ] = 2 i S j k I . (4) [R_j,R_k]=2iS_{jk}I. \tag*{(4)} [ R j , R k ] = 2 i S j k I . (4)
where
S = 1 2 ⨁ j = 1 m Ω , Ω = ( 0 1 − 1 0 ) . S=\frac{1}{2}\bigoplus_{j=1}^{m}\Omega,\qquad\Omega=\begin{pmatrix}0&1\\-1&0\end{pmatrix}. S = 2 1 j = 1 ⨁ m Ω , Ω = ( 0 − 1 1 0 ) .
The matrix S S S is real, skew-symmetric, and nonsingular. For ξ ∈ R 2 m \xi\in\mathbb{R}^{2m} ξ ∈ R 2 m , define the quantum characteristic function by
W ~ ρ ( ξ ) : = Tr ( ρ e i ξ T R ) . \widetilde{W}_{\rho}(\xi):=\operatorname{Tr}\left(\rho e^{i\xi^{T}R}\right). W ρ ( ξ ) := Tr ( ρ e i ξ T R ) .
Definition 2 (Quantum Gaussian state). Let μ ∈ R 2 m \mu\in\mathbb{R}^{2m} μ ∈ R 2 m and let Σ = V + i S \Sigma=V+iS Σ = V + i S , where V V V is a real symmetric matrix and Σ ⪰ 0 \Sigma\succeq0 Σ ⪰ 0 . A density operator ρ \rho ρ on H m \mathcal{H}_{m} H m is called a quantum Gaussian state with mean μ \mu μ and covariance matrix Σ \Sigma Σ if
Tr ( ρ e i ξ T R ) = exp ( i ξ T μ − 1 2 ξ T V ξ ) , ξ ∈ R 2 m . (5) \operatorname{Tr}\left(\rho e^{i\xi^{T}R}\right)=\exp\left(i\xi^{T}\mu-\frac{1}{2}\xi^{T}V\xi\right),\qquad\xi\in\mathbb{R}^{2m}. \tag*{(5)} Tr ( ρ e i ξ T R ) = exp ( i ξ T μ − 2 1 ξ T V ξ ) , ξ ∈ R 2 m . (5) We use the notation ρ = N 2 m ( μ , Σ ) \rho=\mathcal{N}_{2m}(\mu,\Sigma) ρ = N 2 m ( μ , Σ ) .
The real part V = Re Σ V=\operatorname{Re}\Sigma V = Re Σ is the symmetrized covariance matrix, while S = Im Σ S=\operatorname{Im}\Sigma S = Im Σ encodes the commutation relations. With the convention
[ R j , R k ] = 2 i S j k I , Σ = V + i S , [R_j,R_k]=2iS_{jk}I,\qquad\Sigma=V+iS, [ R j , R k ] = 2 i S j k I , Σ = V + i S ,
the matrix Σ \Sigma Σ is also the ordered second-moment matrix of the centered Gaussian observables. More precisely,
Tr [ ρ ( R j − μ j I ) ( R k − μ k I ) ] = V j k + i S j k = Σ j k . (6) \operatorname{Tr}\left[\rho(R_j-\mu_jI)(R_k-\mu_kI)\right]=V_{jk}+iS_{jk}=\Sigma_{jk}. \tag*{(6)} Tr [ ρ ( R j − μ j I ) ( R k − μ k I ) ] = V j k + i S j k = Σ j k . (6)
Indeed, the symmetric part of the left-hand side is V j k V_{jk} V j k , while
1 2 Tr [ ρ [ R j , R k ] ] = i S j k . \frac{1}{2}\operatorname{Tr}\left[\rho[R_j,R_k]\right]=iS_{jk}. 2 1 Tr [ ρ [ R j , R k ] ] = i S j k .
Further since S S S is skew-symmetric, a T Σ a = a T V a a^{T}\Sigma a=a^{T}Va a T Σ a = a T V a for every a ∈ R 2 m a\in\mathbb{R}^{2m} a ∈ R 2 m . It follows immediately from the characteristic function that measuring the single observable a T R a^{T}R a T R under N 2 m ( μ , Σ ) \mathcal{N}_{2m}(\mu,\Sigma) N 2 m ( μ , Σ ) produces the ordinary Gaussian distribution
a T R ∼ N ( a T μ , a T V a ) . (7) a^{T}R\sim\mathcal{N}(a^{T}\mu,a^{T}Va). \tag*{(7)} a T R ∼ N ( a T μ , a T V a ) . (7)
Gaussian models with classical coordinates. The Gaussian limits arising in q-LAN need not have a nonsingular commutation matrix. We therefore allow a general real skew-symmetric matrix S ∈ R d × d S\in\mathbb{R}^{d\times d} S ∈ R d × d . For a k k k -mode quantum system, consider the canonical observables { Q i , P i } i = 1 k \{Q_i,P_i\}_{i=1}^{k} { Q i , P i } i = 1 k satisfying (3) . In addition, allow ℓ \ell ℓ classical Gaussian random variables C 1 , … , C ℓ C_1,\ldots,C_\ell C 1 , … , C ℓ , which commute with each other and with all the quantum coordinates. We write the resulting d = 2 k + ℓ d=2k+\ell d = 2 k + ℓ Gaussian coordinates as
( X 1 , … , X d ) ≡ ( C 1 , … , C ℓ , Q 1 , P 1 , … , Q k , P k ) . (8) (X_1,\ldots,X_d)\equiv(C_1,\ldots,C_\ell,Q_1,P_1,\ldots,Q_k,P_k). \tag*{(8)} ( X 1 , … , X d ) ≡ ( C 1 , … , C ℓ , Q 1 , P 1 , … , Q k , P k ) . (8)
Their commutation structure is represented compactly by [ X i , X j ] = 2 i S i j I [X_i,X_j]=2iS_{ij}I [ X i , X j ] = 2 i S ij I , where
S = 0 ℓ × ℓ ⊕ 1 2 ⨁ j = 1 k Ω . S=0_{\ell\times\ell}\oplus\frac{1}{2}\bigoplus_{j=1}^{k}\Omega. S = 0 ℓ × ℓ ⊕ 2 1 j = 1 ⨁ k Ω.
We represent a product classical–quantum experiment by ϱ = f ⊗ ρ \varrho=f\otimes\rho ϱ = f ⊗ ρ , where f f f is a probability density on R ℓ \mathbb{R}^{\ell} R ℓ and ρ \rho ρ is a density operator on H k = L 2 ( R ) ⊗ k \mathcal{H}_k=L^2(\mathbb{R})^{\otimes k} H k = L 2 ( R ) ⊗ k . Its characteristic function is defined by
E ϱ ( e i ∑ j = 1 d u j X j ) : = ∫ R ℓ Tr ( ρ e i ∑ j = ℓ + 1 d u j X j ) f ( y ) e i ∑ j = 1 ℓ u j y j d y . (9) E_{\varrho}\left(e^{i\sum_{j=1}^{d}u_jX_j}\right):=\int_{\mathbb{R}^{\ell}}\operatorname{Tr}\left(\rho e^{i\sum_{j=\ell+1}^{d}u_jX_j}\right)f(y)e^{i\sum_{j=1}^{\ell}u_jy_j}\,\mathrm{d}y. \tag*{(9)} E ϱ ( e i ∑ j = 1 d u j X j ) := ∫ R ℓ Tr ( ρ e i ∑ j = ℓ + 1 d u j X j ) f ( y ) e i ∑ j = 1 ℓ u j y j d y . (9)
Definition 3 (Classical–quantum Gaussian experiment). A product classical–quantum experiment ϱ = f ⊗ ρ \varrho=f\otimes\rho ϱ = f ⊗ ρ is called Gaussian if its characteristic function has the form
E ϱ ( e i ∑ j = 1 d u j X j ) = exp ( i u T μ − 1 2 u T V u ) , u ∈ R d , E_{\varrho}\left(e^{i\sum_{j=1}^{d}u_jX_j}\right)=\exp\left(iu^{T}\mu-\frac{1}{2}u^{T}Vu\right),\qquad u\in\mathbb{R}^{d}, E ϱ ( e i ∑ j = 1 d u j X j ) = exp ( i u T μ − 2 1 u T V u ) , u ∈ R d , for some μ ∈ R d \mu\in\mathbb{R}^{d} μ ∈ R d and real symmetric covariance matrix V V V , and we associate to it the complex covariance Σ = V + i S \Sigma=V+iS Σ = V + i S .
A Gaussian experiment with a possibly degenerate commutation matrix and $\operatorname{Re}\Sigma\succ 0 admits, after an invertible real linear transformation of its coordinates, a product classical–quantum representation. We use N ( μ , Σ ) N(\mu,\Sigma) N ( μ , Σ ) for both purely quantum and classical–quantum Gaussian experiments. When S = Im Σ S=\operatorname{Im}\Sigma S = Im Σ is nonsingular, this experiment is represented by the Gaussian density operator N d ( μ , Σ ) N_d(\mu,\Sigma) N d ( μ , Σ ) introduced above; when S S S is degenerate, it denotes the corresponding classical–quantum Gaussian experiment. In either case, we write φ μ , Σ \varphi_{\mu,\Sigma} φ μ , Σ for the associated normal state on the von Neumann algebra generated by the Gaussian Weyl operators. The precise von Neumann algebra formulation, including the degenerate case, is given in Appendix § 12 .
It is convenient to denote the Gaussian coordinates collectively by X = ( X 1 , … , X d ) T X=(X_1,\ldots,X_d)^T X = ( X 1 , … , X d ) T . In the classical–quantum case, some of the X j X_j X j are ordinary classical Gaussian random variables and the remaining coordinates are canonical quantum observables. For every a ∈ R d a\in\mathbb{R}^{d} a ∈ R d , the scalar a T X a^TX a T X has the ordinary Gaussian distribution
a T X ∼ N ( a T μ , a T Re ( Σ ) a ) . (10) a^TX\sim N\left(a^T\mu,a^T\operatorname{Re}(\Sigma)a\right). \tag*{(10)} a T X ∼ N ( a T μ , a T Re ( Σ ) a ) . (10)
Convergence in distribution. Let H ( n ) \mathcal{H}^{(n)} H ( n ) be a sequence of Hilbert spaces, ρ ( n ) \rho^{(n)} ρ ( n ) states on H ( n ) \mathcal{H}^{(n)} H ( n ) , and
R ( n ) = ( R 1 ( n ) , … , R d ( n ) ) R^{(n)}=\left(R_1^{(n)},\ldots,R_d^{(n)}\right) R ( n ) = ( R 1 ( n ) , … , R d ( n ) )
lists of observables. We say that
R ( n ) ⇝ ρ ( n ) N ( μ , Σ ) R^{(n)}\mathop{\rightsquigarrow}^{\rho^{(n)}}N(\mu,\Sigma) R ( n ) ⇝ ρ ( n ) N ( μ , Σ )
if their ordered quantum characteristic functions converge to those of the limiting Gaussian experiment. More precisely, for every r ≥ 1 r\geq1 r ≥ 1 and every ξ 1 , … , ξ r ∈ R d \xi_1,\ldots,\xi_r\in\mathbb{R}^{d} ξ 1 , … , ξ r ∈ R d ,
lim n → ∞ Tr [ ρ ( n ) ∏ t = 1 r e i ξ t T R ( n ) ] = E N ( μ , Σ ) [ ∏ t = 1 r e i ξ t T X ] . (11) \lim_{n\to\infty}\operatorname{Tr}\left[\rho^{(n)}\prod_{t=1}^{r}e^{i\xi_t^TR^{(n)}}\right]=\mathbb{E}_{N(\mu,\Sigma)}\left[\prod_{t=1}^{r}e^{i\xi_t^TX}\right]. \tag*{(11)} n → ∞ lim Tr [ ρ ( n ) t = 1 ∏ r e i ξ t T R ( n ) ] = E N ( μ , Σ ) [ t = 1 ∏ r e i ξ t T X ] . (11)
When the limiting model is purely quantum, the expectation on the right-hand side is the corresponding trace against the Gaussian density operator. When classical coordinates are present, the expectation additionally averages over the classical Gaussian component.
q-LAN and D-Extendibility
Next we formalize this using the q-LAN framework [13 , 14 ] , written here in the functional notation of Kumagai and Hayashi [29 ] . In classical statistics, an experiment is defined to be a family of probability measures on a sample space and denoted by E = { P θ , θ ∈ Θ } \mathcal{E} = \{P_\theta, \theta\in\Theta\} E = { P θ , θ ∈ Θ } where Θ \Theta Θ is the parameter space. Analogously, we define a quantum statistical model as follows:
Definition 4. A quantum statistical model over a parameter space Θ \Theta Θ consists of a family of quantum states Q = { ρ θ : θ ∈ Θ } \mathcal{Q} = \{\rho_\theta: \theta\in\Theta\} Q = { ρ θ : θ ∈ Θ } on a Hilbert space H \mathcal{H} H , indexed by an unknown parameter θ ∈ Θ \theta\in\Theta θ ∈ Θ .
For a differentiable parametric family of states S = { ρ θ : θ ∈ Θ ⊂ R d } \mathcal{S} = \{\rho_\theta: \theta\in\Theta\subset\mathbb{R}^d\} S = { ρ θ : θ ∈ Θ ⊂ R d } , the symmetric logarithmic derivative (SLD) in the j j j th parameter direction at θ \theta θ is a self-adjoint operator L θ , j L_{\theta,j} L θ , j satisfying
∂ j ρ θ = 1 2 ( ρ θ L θ , j + L θ , j ρ θ ) , j = 1 , … , d . (12) \partial_j \rho_\theta= \frac{1}{2}\left(\rho_\theta L_{\theta,j} + L_{\theta,j}\rho_\theta\right), \qquad j = 1,\ldots,d. \tag*{(12)} ∂ j ρ θ = 2 1 ( ρ θ L θ , j + L θ , j ρ θ ) , j = 1 , … , d . (12)
Taking the trace in (12) also gives
Tr ( ρ θ L θ , j ) = 0. (13) \operatorname{Tr}\left(\rho_\theta L_{\theta,j}\right) = 0. \tag*{(13)} Tr ( ρ θ L θ , j ) = 0. (13)
When ρ θ \rho_\theta ρ θ is strictly positive, the SLD is unique. Indeed, in an eigenbasis of ρ θ \rho_\theta ρ θ , with eigenvalues λ 1 , … , λ r > 0 \lambda_1,\ldots,\lambda_r > 0 λ 1 , … , λ r > 0 ,
( L θ , j ) a b = 2 ( ∂ j ρ θ ) a b λ a + λ b . (L_{\theta,j})_{ab} = \frac{2(\partial_j\rho_\theta)_{ab}}{\lambda_a + \lambda_b}. ( L θ , j ) ab = λ a + λ b 2 ( ∂ j ρ θ ) ab .
Whenever the SLDs exist, we define their ordered second-moment matrix and the real SLD Fisher information matrix by
J θ , j k : = Tr ( ρ θ L θ , j L θ , k ) , G θ , j k : = Re J θ , j k = 1 2 Tr [ ρ θ ( L θ , j L θ , k + L θ , k L θ , j ) ] . (14) J_{\theta,jk} := \operatorname{Tr}\left(\rho_\theta L_{\theta,j}L_{\theta,k}\right), \qquad G_{\theta,jk} := \operatorname{Re}J_{\theta,jk} = \frac{1}{2}\operatorname{Tr}\left[\rho_\theta\left(L_{\theta,j}L_{\theta,k} + L_{\theta,k}L_{\theta,j}\right)\right]. \tag*{(14)} J θ , j k := Tr ( ρ θ L θ , j L θ , k ) , G θ , j k := Re J θ , j k = 2 1 Tr [ ρ θ ( L θ , j L θ , k + L θ , k L θ , j ) ] . (14)
Thus G θ = Re J θ G_\theta= \operatorname{Re}J_\theta G θ = Re J θ is real symmetric and positive semidefinite.
Next for finite-dimensional states ρ \rho ρ and σ \sigma σ , we consider the square-root likelihood ratio R ( σ ∣ ρ ) \mathcal{R}(\sigma\mid\rho) R ( σ ∣ ρ ) and the quantum Lebesgue decomposition of [13 ] . A chosen representative R = R ( σ ∣ ρ ) \mathcal{R} = \mathcal{R}(\sigma\mid\rho) R = R ( σ ∣ ρ ) satisfies
σ = R ρ R + σ ⊥ , R ⪰ 0 , σ ⊥ ⪰ 0 , Tr ( ρ σ ⊥ ) = 0. \sigma= \mathcal{R}\rho\mathcal{R} + \sigma^\perp, \qquad\mathcal{R} \succeq0,\quad\sigma^\perp\succeq0,\quad\operatorname{Tr}(\rho\sigma^\perp) = 0. σ = R ρ R + σ ⊥ , R ⪰ 0 , σ ⊥ ⪰ 0 , Tr ( ρ σ ⊥ ) = 0.
When ρ \rho ρ and σ \sigma σ are faithful, the positive solution is unique and is
R = ρ − 1 / 2 ( ρ 1 / 2 σ ρ 1 / 2 ) 1 / 2 ρ − 1 / 2 . \mathcal{R} = \rho^{-1/2}\left(\rho^{1/2}\sigma\rho^{1/2}\right)^{1/2}\rho^{-1/2}. R = ρ − 1/2 ( ρ 1/2 σ ρ 1/2 ) 1/2 ρ − 1/2 .
At nonfaithful states the displayed decomposition alone does not assert uniqueness on the ambient Hilbert space. In the next definition we also use the notations o L 2 ( ρ ( n ) ) o_{L^2}(\rho^{(n)}) o L 2 ( ρ ( n ) ) and o D ( Z ( n ) , ρ ( n ) ) o_D(Z^{(n)}, \rho^{(n)}) o D ( Z ( n ) , ρ ( n ) ) to denote remainders infinitesimal in L 2 L^2 L 2 and in distribution, respectively, following [14 Definition 1.2] ; see also [13 Definition 5.5] for the precise definition of the latter.
Definition 5 (q-LAN). A sequence of quantum statistical models S ( n ) = { ρ θ ( n ) ∣ θ ∈ Θ ⊆ R d } \mathcal{S}^{(n)} = \{\rho_\theta^{(n)} \mid\theta\in\Theta\subseteq\mathbb{R}^d\} S ( n ) = { ρ θ ( n ) ∣ θ ∈ Θ ⊆ R d } is called q-LAN at θ 0 \theta_0 θ 0 if for every fixed local parameter h ∈ R d h \in\mathbb{R}^d h ∈ R d , the square-root likelihood ratio R h ( n ) = R ( ρ θ 0 + h / n ( n ) ∣ ρ θ 0 ( n ) ) \mathcal{R}_h^{(n)} = \mathcal{R}(\rho_{\theta_0+h/\sqrt{n}}^{(n)} \mid\rho_{\theta_0}^{(n)}) R h ( n ) = R ( ρ θ 0 + h / n ( n ) ∣ ρ θ 0 ( n ) ) can be expanded as:
log ( R h ( n ) + o L 2 ( ρ θ 0 ( n ) ) ) 2 = h T Δ ( n ) − 1 2 ( h T J h ) I ( n ) + o D ( h T Δ ( n ) , ρ θ 0 ( n ) ) (15) \log\left(\mathcal{R}_h^{(n)} + o_{L^2}(\rho_{\theta_0}^{(n)})\right)^2 = h^\mathsf{T}\Delta^{(n)} - \frac{1}{2}\left(h^\mathsf{T}Jh\right)I^{(n)} + o_D\left(h^\mathsf{T}\Delta^{(n)}, \rho_{\theta_0}^{(n)}\right) \tag*{(15)} log ( R h ( n ) + o L 2 ( ρ θ 0 ( n ) ) ) 2 = h T Δ ( n ) − 2 1 ( h T J h ) I ( n ) + o D ( h T Δ ( n ) , ρ θ 0 ( n ) ) (15)
where
Δ ( n ) = ( Δ 1 ( n ) , … , Δ d ( n ) ) T \Delta^{(n)}=\left(\Delta^{(n)}_1,\ldots,\Delta^{(n)}_d\right)^T Δ ( n ) = ( Δ 1 ( n ) , … , Δ d ( n ) ) T
is a d d d -tuple of self-adjoint observables, called the quantum central sequence, satisfying
Δ ( n ) → ρ 0 ( n ) N ( 0 , J ) , \Delta^{(n)} \xrightarrow{\rho_0^{(n)}} \mathcal{N}(0,J), Δ ( n ) ρ 0 ( n ) N ( 0 , J ) ,
where J ∈ C d × d J\in\mathbb{C}^{d\times d} J ∈ C d × d is Hermitian positive semidefinite and Re J ≻ 0 \operatorname{Re}J\succ0 Re J ≻ 0 .
Definition 6 (Holevo’s commutation operator). Let ρ \rho ρ be a density matrix on a finite-dimensional Hilbert space, and define the left- and right-multiplication superoperators by
L ρ ( X ) : = ρ X , R ρ ( X ) : = X ρ . L_\rho(X):=\rho X,\qquad R_\rho(X):=X\rho. L ρ ( X ) := ρX , R ρ ( X ) := X ρ . When ρ \rho ρ is faithful, Holevo’s commutation operator is the linear superoperator
D ρ : = − 1 ( L ρ + R ρ ) − 1 ( L ρ − R ρ ) . D_\rho:=\sqrt{-1}(L_\rho+R_\rho)^{-1}(L_\rho-R_\rho). D ρ := − 1 ( L ρ + R ρ ) − 1 ( L ρ − R ρ ) . Equivalently, it is characterized by
ρ D ρ ( X ) + D ρ ( X ) ρ = − 1 ( ρ X − X ρ ) . (16) \rho D_\rho(X)+D_\rho(X)\rho=\sqrt{-1}(\rho X-X\rho). \tag*{(16)} ρ D ρ ( X ) + D ρ ( X ) ρ = − 1 ( ρX − X ρ ) . (16) For a possibly nonfaithful state, we use the canonical extension obtained by replacing the inverse with the Moore–Penrose inverse. In an eigenbasis of ρ \rho ρ , with eigenvalues λ 1 , … , λ r \lambda_1,\ldots,\lambda_r λ 1 , … , λ r , this gives
( D ρ ( X ) ) a b = { − 1 λ a − λ b λ a + λ b X a b , λ a + λ b > 0 , 0 , λ a + λ b = 0. \left(D_\rho(X)\right)_{ab}=
\begin{cases}
\sqrt{-1}\dfrac{\lambda_a-\lambda_b}{\lambda_a+\lambda_b}X_{ab}, & \lambda_a+\lambda_b>0,\\
0, & \lambda_a+\lambda_b=0.
\end{cases} ( D ρ ( X ) ) ab = ⎩ ⎨ ⎧ − 1 λ a + λ b λ a − λ b X ab , 0 , λ a + λ b > 0 , λ a + λ b = 0. For nonfaithful ρ \rho ρ , this is the canonical representative of the operator on the quotient by
K ρ : = { X : X ρ = ρ X = 0 } ; K_\rho:=\{X:X\rho=\rho X=0\}; K ρ := { X : X ρ = ρX = 0 } ; see [14 Appendix C] .
Definition 7 (D \mathcal{D} D -extendibility). Let R ( n ) = ( R 1 ( n ) , … , R m ( n ) ) T R^{(n)} = (R^{(n)}_1,\ldots,R^{(n)}_m)^{\mathsf T} R ( n ) = ( R 1 ( n ) , … , R m ( n ) ) T be a list of self-adjoint observables. The sequence R ( n ) R^{(n)} R ( n ) is called asymptotically D \mathcal{D} D -invariant at θ 0 \theta_0 θ 0 if
R ( n ) ⟶ ρ θ 0 ( n ) N ( 0 , Σ ) , R^{(n)} \mathrel{\mathop{\longrightarrow}^{\rho_{\theta_0}^{(n)}}} \mathcal{N}(0,\Sigma), R ( n ) ⟶ ρ θ 0 ( n ) N ( 0 , Σ ) , where Σ ∈ C m × m \Sigma\in\mathbb{C}^{m\times m} Σ ∈ C m × m is Hermitian, Σ ⪰ 0 \Sigma\succeq0 Σ ⪰ 0 , and Re Σ ≻ 0 \operatorname{Re}\Sigma\succ0 Re Σ ≻ 0 , and if, for every ξ , η ∈ R m \xi,\eta\in\mathbb{R}^m ξ , η ∈ R m ,
lim n → ∞ Tr ( ρ θ 0 ( n ) e − i ξ T R ( n ) ρ θ 0 ( n ) e − i η T R ( n ) ) = exp [ − 1 2 ( ξ η ) T ( Σ Σ # Σ T Σ # Σ T Σ T ) ( ξ η ) ] . (17) \lim_{n\to\infty}\operatorname{Tr}\left(\sqrt{\rho_{\theta_0}^{(n)}}e^{-\mathrm{i}\xi^{\mathsf T}R^{(n)}}\sqrt{\rho_{\theta_0}^{(n)}}e^{-\mathrm{i}\eta^{\mathsf T}R^{(n)}}\right)
=\exp\left[-\frac{1}{2}\begin{pmatrix}\xi\\\eta\end{pmatrix}^{\mathsf T}
\begin{pmatrix}\Sigma&\Sigma\mathbin{\#}\Sigma^{\mathsf T}\\\Sigma\mathbin{\#}\Sigma^{\mathsf T}&\Sigma^{\mathsf T}\end{pmatrix}
\begin{pmatrix}\xi\\\eta\end{pmatrix}\right]. \tag*{(17)} n → ∞ lim Tr ( ρ θ 0 ( n ) e − i ξ T R ( n ) ρ θ 0 ( n ) e − i η T R ( n ) ) = exp [ − 2 1 ( ξ η ) T ( Σ Σ # Σ T Σ # Σ T Σ T ) ( ξ η ) ] . (17) Here # \# # denotes the operator geometric mean, extended to positive semidefinite matrices by continuity.
A q q q -LAN model S ( n ) \mathcal{S}^{(n)} S ( n ) with d d d -dimensional quantum central sequence Δ ( n ) \Delta^{(n)} Δ ( n ) is called D \mathcal{D} D -extendible at θ 0 \theta_0 θ 0 if there exist an asymptotically D \mathcal{D} D -invariant sequence R ( n ) R^{(n)} R ( n ) and a fixed real matrix F ∈ R m × d F\in\mathbb{R}^{m\times d} F ∈ R m × d such that
Δ ( n ) = F T R ( n ) . \Delta^{(n)}=F^{\mathsf T}R^{(n)}. Δ ( n ) = F T R ( n ) .
In this case, the covariance matrix of the central sequence is J = F T Σ F J=F^{\mathsf T}\Sigma F J = F T Σ F . We call the model asymptotically D \mathcal{D} D -invariant when the central sequence itself is asymptotically D \mathcal{D} D -invariant, that is, when the preceding conditions hold with m = d m=d m = d , R ( n ) = Δ ( n ) R^{(n)}=\Delta^{(n)} R ( n ) = Δ ( n ) , and Σ = J \Sigma=J Σ = J .
We are now ready to state the asymptotic representation theorem for the models which are q q q -LAN and D \mathcal{D} D -extendible. As shown in [14 ] , this theorem is crucial in establishing optimality of estimators in quantum i.i.d. models. We will show that this theorem also serves as one of the key ingredients in deriving upper bounds on local testing power in q q q -LAN models.
To describe the limiting Gaussian states let τ = Σ F \tau=\Sigma F τ = Σ F and define
ν h : = ( Re τ ) h , ϱ h : = φ ν h , Σ , h ∈ R d . \nu_h := (\operatorname{Re}\tau)h,\qquad\varrho_h:=\varphi_{\nu_h,\Sigma},\qquad h\in\mathbb{R}^d. ν h := ( Re τ ) h , ϱ h := φ ν h , Σ , h ∈ R d .
Theorem 2.1 (Quantum asymptotic representation theorem [14 ] ). Suppose that S ( n ) \mathcal{S}^{(n)} S ( n ) is q q q -LAN and D \mathcal{D} D -extendible at θ 0 \theta_0 θ 0 . Let M ( n ) M^{(n)} M ( n ) be POVMs with a common outcome space ( R s , B ( R s ) ) (\mathbb{R}^s,\mathcal{B}(\mathbb{R}^s)) ( R s , B ( R s )) , and define
L h ( n ) ( B ) : = Tr ( ρ θ 0 + h / n ( n ) M ( n ) ( B ) ) , B ∈ B ( R s ) . \mathcal{L}^{(n)}_h(B):=\operatorname{Tr}\left(\rho_{\theta_0+h/\sqrt{n}}^{(n)}M^{(n)}(B)\right),\qquad B\in\mathcal{B}(\mathbb{R}^s). L h ( n ) ( B ) := Tr ( ρ θ 0 + h / n ( n ) M ( n ) ( B ) ) , B ∈ B ( R s ) . Assume that L h ( n ) ⇒ L h \mathcal{L}^{(n)}_h\Rightarrow\mathcal{L}_h L h ( n ) ⇒ L h for every fixed h ∈ R d h\in\mathbb{R}^d h ∈ R d . Then there exists a POVM M ( ∞ ) M^{(\infty)} M ( ∞ ) on the limiting Gaussian experiment { ϱ h : h ∈ R d } \{\varrho_h:h\in\mathbb{R}^d\} { ϱ h : h ∈ R d } such that
L h ( B ) = ϱ h ( M ( ∞ ) ( B ) ) , h ∈ R d , B ∈ B ( R s ) . \mathcal{L}_h(B)=\varrho_h\left(M^{(\infty)}(B)\right),\qquad h\in\mathbb{R}^d,\quad B\in\mathcal{B}(\mathbb{R}^s). L h ( B ) = ϱ h ( M ( ∞ ) ( B ) ) , h ∈ R d , B ∈ B ( R s ) . The weak-convergence assumption can be replaced by tightness under each fixed local parameter if one passes to a common subsequence, as follows.
Theorem 2.2 (A subsequential version of Quantum Asymptotic Representation Theorem). Let
S ( n ) = { ρ θ ( n ) : θ ∈ Θ ⊂ R d } \mathcal{S}^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb{R}^d\} S ( n ) = { ρ θ ( n ) : θ ∈ Θ ⊂ R d }
be q q q -LAN and D \mathcal{D} D -extendible at θ 0 \theta_0 θ 0 , with D \mathcal{D} D -extension R ( n ) R^{(n)} R ( n ) , limiting covariance Σ \Sigma Σ , and τ = Σ F \tau=\Sigma F τ = Σ F . Let M ( n ) M^{(n)} M ( n ) be arbitrary locally tight POVMs on R s \mathbb{R}^s R s , i.e. assume that for every fixed h ∈ R d h\in\mathbb{R}^d h ∈ R d the laws
L h ( n ) ( B ) = Tr ρ θ 0 + h / n ( n ) M ( n ) ( B ) \mathcal{L}^{(n)}_h(B)=\operatorname{Tr}\rho^{(n)}_{\theta_0+h/\sqrt{n}}M^{(n)}(B) L h ( n ) ( B ) = Tr ρ θ 0 + h / n ( n ) M ( n ) ( B )
are tight. Then there is a subsequence n k n_k n k , chosen independently of h h h , and a POVM
M ( ∞ ) : B ( R s ) → CCR ( Im Σ ) + M^{(\infty)}:\mathcal{B}(\mathbb{R}^s)\to\operatorname{CCR}(\operatorname{Im}\Sigma)_+ M ( ∞ ) : B ( R s ) → CCR ( Im Σ ) +
such that, for every fixed h ∈ R d h\in\mathbb{R}^d h ∈ R d ,
L h ( n k ) = Tr ρ θ 0 + h / n k ( n k ) M ( n k ) ⇒ L h , \mathcal{L}^{(n_k)}_h=\operatorname{Tr}\rho^{(n_k)}_{\theta_0+h/\sqrt{n_k}}M^{(n_k)}\Rightarrow\mathcal{L}_h, L h ( n k ) = Tr ρ θ 0 + h / n k ( n k ) M ( n k ) ⇒ L h ,
where the limiting law is represented on the limiting quantum Gaussian shift by
L h ( B ) = ϱ h ( M ( ∞ ) ( B ) ) , \mathcal{L}_h(B)=\varrho_h\left(M^{(\infty)}(B)\right), L h ( B ) = ϱ h ( M ( ∞ ) ( B ) ) ,
In particular, every locally tight sequence of POVMs has a common subsequence along which its outcome laws converge under all fixed local alternatives.
Hypothesis testing in quantum statistical models
We begin by recalling the classical formulation of hypothesis testing before introducing its quantum analogue. The objective is to provide a common decision-theoretic framework that will later be used to formulate optimality criteria for quantum tests.
Classical hypothesis testing
Let P = { P θ : θ ∈ Θ } \mathcal{P}=\{P_\theta:\theta\in\Theta\} P = { P θ : θ ∈ Θ } be a parametric statistical model and suppose that X 1 , … , X n ∼ i . i . d . P θ X_1,\ldots,X_n\mathrel{\overset{\mathrm{i.i.d.}}{\sim}}P_\theta X 1 , … , X n ∼ i.i.d. P θ . The basic hypothesis testing problem consists of deciding between the hypotheses
H 0 : θ ∈ Θ 0 , H 1 : θ ∈ Θ 1 , H_0:\theta\in\Theta_0,\qquad H_1:\theta\in\Theta_1, H 0 : θ ∈ Θ 0 , H 1 : θ ∈ Θ 1 ,
where Θ 0 ∩ Θ 1 = ∅ \Theta_0\cap\Theta_1=\varnothing Θ 0 ∩ Θ 1 = ∅ .
More generally, many statistical testing problems concern a functional of the underlying distribution rather than the parameter itself. Let ϑ : P → R \vartheta:\mathcal{P}\to\mathbb{R} ϑ : P → R be a statistical functional and write ϑ ( P θ ) = ψ ( θ ) \vartheta(P_\theta)=\psi(\theta) ϑ ( P θ ) = ψ ( θ ) . Testing a property of the functional may then be formulated as
H 0 : ψ ( θ ) ∈ C , H 1 : ψ ( θ ) ∉ C , (18) H_0:\psi(\theta)\in C,\qquad H_1:\psi(\theta)\notin C, \tag*{(18)} H 0 : ψ ( θ ) ∈ C , H 1 : ψ ( θ ) ∈ / C , (18)
for a prescribed subset C ⊆ R C\subseteq\mathbb{R} C ⊆ R .
A (possibly randomized) test is a measurable function ϕ n : X n → [ 0 , 1 ] \phi_n:\mathcal{X}^n\to[0,1] ϕ n : X n → [ 0 , 1 ] , where ϕ n ( x 1 , … , x n ) \phi_n(x_1,\ldots,x_n) ϕ n ( x 1 , … , x n ) denotes the probability of rejecting the null hypothesis after observing ( x 1 , … , x n ) (x_1,\ldots,x_n) ( x 1 , … , x n ) . The performance of a test is summarized by its power function
π ϕ n ( θ ) : = E θ [ ϕ n ( X 1 , … , X n ) ] . \pi_{\phi_n}(\theta):=\mathbb{E}_\theta[\phi_n(X_1,\ldots,X_n)]. π ϕ n ( θ ) := E θ [ ϕ n ( X 1 , … , X n )] .
A test is said to have level α \alpha α if
sup θ : ψ ( θ ) ∈ C π ϕ n ( θ ) ≤ α . \sup_{\theta:\psi(\theta)\in C}\pi_{\phi_n}(\theta)\leq\alpha. θ : ψ ( θ ) ∈ C sup π ϕ n ( θ ) ≤ α .
Let T α ( n ) \mathbb{T}_{\alpha}^{(n)} T α ( n ) denote the collection of all level-α \alpha α tests. A level-α \alpha α test ϕ n ⋆ ∈ T α ( n ) \phi_n^\star\in\mathbb{T}_{\alpha}^{(n)} ϕ n ⋆ ∈ T α ( n ) is called uniformly most powerful (UMP) if
π ϕ n ⋆ ( θ ) ≥ π ϕ n ( θ ) , ∀ ϕ n ∈ T α ( n ) , ∀ θ : ψ ( θ ) ∉ C . \pi_{\phi_n^\star}(\theta) \ge\pi_{\phi_n}(\theta), \qquad\forall\phi_n \in\mathbb{T}_{\alpha}^{(n)}, \quad\forall\theta: \psi(\theta) \notin\mathcal{C}. π ϕ n ⋆ ( θ ) ≥ π ϕ n ( θ ) , ∀ ϕ n ∈ T α ( n ) , ∀ θ : ψ ( θ ) ∈ / C .
Uniformly most powerful tests need not exist, particularly for two-sided testing problems. In such situations one often restricts attention to the class of unbiased tests. A level-α \alpha α test is called unbiased if
sup θ : ψ ( θ ) ∈ C π ϕ n ( θ ) ≤ α , inf θ : ψ ( θ ) ∉ C π ϕ n ( θ ) ≥ α . \sup_{\theta:\psi(\theta)\in\mathcal{C}} \pi_{\phi_n}(\theta) \le\alpha, \qquad\inf_{\theta:\psi(\theta)\notin\mathcal{C}} \pi_{\phi_n}(\theta) \ge\alpha. θ : ψ ( θ ) ∈ C sup π ϕ n ( θ ) ≤ α , θ : ψ ( θ ) ∈ / C inf π ϕ n ( θ ) ≥ α .
Let T α , u ( n ) \mathbb{T}_{\alpha,u}^{(n)} T α , u ( n ) denote the collection of all level-α \alpha α unbiased tests. A test ϕ n ⋆ ∈ T α , u ( n ) \phi_n^\star\in\mathbb{T}_{\alpha,u}^{(n)} ϕ n ⋆ ∈ T α , u ( n ) is called uniformly most powerful unbiased (UMPU) if
π ϕ n ⋆ ( θ ) ≥ π ϕ n ( θ ) , ∀ ϕ n ∈ T α , u ( n ) , ∀ θ : ψ ( θ ) ∉ C . \pi_{\phi_n^\star}(\theta) \ge\pi_{\phi_n}(\theta), \qquad\forall\phi_n \in\mathbb{T}_{\alpha,u}^{(n)}, \quad\forall\theta: \psi(\theta) \notin\mathcal{C}. π ϕ n ⋆ ( θ ) ≥ π ϕ n ( θ ) , ∀ ϕ n ∈ T α , u ( n ) , ∀ θ : ψ ( θ ) ∈ / C .
Although introduced here for an i.i.d. sample, these definitions apply equally to a general sequence of classical experiments { P θ ( n ) : θ ∈ Θ } \{P_\theta^{(n)} : \theta\in\Theta\} { P θ ( n ) : θ ∈ Θ } : a test is a measurable [ 0 , 1 ] [0,1] [ 0 , 1 ] -valued function of the observation in the n n n th experiment, with power E θ P θ ( n ) [ ϕ n ] \mathbb{E}_{\theta}^{P_\theta^{(n)}}[\phi_n] E θ P θ ( n ) [ ϕ n ] .
Hypothesis testing with nuisance parameters
Many hypothesis testing problems involve nuisance parameters that are not of direct inferential interest but nevertheless affect the distribution of the observations. Assume in this subsection that the parameter space has the product form Θ = Θ η × Λ \Theta= \Theta_\eta\times\Lambda Θ = Θ η × Λ , and write θ = ( η , λ ) \theta= (\eta,\lambda) θ = ( η , λ ) , where η ∈ Θ η \eta\in\Theta_\eta η ∈ Θ η is the parameter of interest and λ ∈ Λ \lambda\in\Lambda λ ∈ Λ is a nuisance parameter. Let Θ 0 , Θ 1 ⊆ Θ η \Theta_0,\Theta_1 \subseteq\Theta_\eta Θ 0 , Θ 1 ⊆ Θ η be disjoint sets. The testing problem takes the form
H 0 : η ∈ Θ 0 , H 1 : η ∈ Θ 1 . (19) H_0 : \eta\in\Theta_0, \qquad H_1 : \eta\in\Theta_1. \tag*{(19)} H 0 : η ∈ Θ 0 , H 1 : η ∈ Θ 1 . (19)
The power function now depends on both the parameter of interest and the nuisance parameter,
π ϕ n ( η , λ ) = E η , λ [ ϕ n ( X 1 , … , X n ) ] , \pi_{\phi_n}(\eta,\lambda) = \mathbb{E}_{\eta,\lambda}[\phi_n(X_1,\ldots,X_n)], π ϕ n ( η , λ ) = E η , λ [ ϕ n ( X 1 , … , X n )] ,
and the level-α \alpha α condition becomes
sup η ∈ Θ 0 , λ ∈ Λ π ϕ n ( η , λ ) ≤ α . \sup_{\eta\in\Theta_0,\ \lambda\in\Lambda} \pi_{\phi_n}(\eta,\lambda) \le\alpha. η ∈ Θ 0 , λ ∈ Λ sup π ϕ n ( η , λ ) ≤ α .
Uniformly most powerful tests may also fail to exist in the presence of nuisance parameters. A natural decision-theoretic criterion is therefore to optimize the worst-case power over all possible nuisance parameters. Accordingly, for every η ∈ Θ 1 \eta\in\Theta_1 η ∈ Θ 1 , define the minimum (or worst-case) power function
π ‾ ϕ n ( η ) : = inf λ ∈ Λ π ϕ n ( η , λ ) . \underline{\pi}_{\phi_n}(\eta) := \inf_{\lambda\in\Lambda} \pi_{\phi_n}(\eta,\lambda). π ϕ n ( η ) := λ ∈ Λ inf π ϕ n ( η , λ ) .
A level-α \alpha α test ϕ n ⋆ \phi_n^\star ϕ n ⋆ is called maximin if
π ‾ ϕ n ⋆ ( η ) ≥ π ‾ ϕ n ( η ) , ∀ ϕ n ∈ T α ( n ) , ∀ η ∈ Θ 1 . \underline{\pi}_{\phi_n^\star}(\eta) \ge\underline{\pi}_{\phi_n}(\eta), \qquad\forall\phi_n \in\mathbb{T}_{\alpha}^{(n)}, \quad\forall\eta\in\Theta_1. π ϕ n ⋆ ( η ) ≥ π ϕ n ( η ) , ∀ ϕ n ∈ T α ( n ) , ∀ η ∈ Θ 1 .
Similarly, a level-α \alpha α unbiased test ϕ n ⋆ \phi_n^\star ϕ n ⋆ is called unbiased maximin if
π ‾ ϕ n ⋆ ( η ) ≥ π ‾ ϕ n ( η ) , ∀ ϕ n ∈ T α , u ( n ) , ∀ η ∈ Θ 1 . \underline{\pi}_{\phi_n^\star}(\eta) \ge\underline{\pi}_{\phi_n}(\eta), \qquad\forall\phi_n \in\mathbb{T}_{\alpha,u}^{(n)}, \quad\forall\eta\in\Theta_1. π ϕ n ⋆ ( η ) ≥ π ϕ n ( η ) , ∀ ϕ n ∈ T α , u ( n ) , ∀ η ∈ Θ 1 .
Here maximin is used in the uniform nuisance sense: the same test maximizes the minimum power over λ \lambda λ for every η ∈ Θ 1 \eta\in\Theta_1 η ∈ Θ 1 . This is the UMP min-max criterion in the terminology of [29 Section 3] ; it is stronger than optimizing a single minimum over both η \eta η and λ \lambda λ . We are now ready to discuss the quantum analogues of the notions introduced above.
Quantum hypothesis testing
The quantum analogue of the classical testing problem replaces probability distributions with quantum states and randomized test functions with binary positive operator-valued measures.
Let
S ( n ) = { ρ θ ( n ) : θ ∈ Θ } \mathcal{S}^{(n)}=\{\rho_{\theta}^{(n)}:\theta\in\Theta\} S ( n ) = { ρ θ ( n ) : θ ∈ Θ }
be a sequence of quantum statistical models on Hilbert spaces H ( n ) \mathcal{H}^{(n)} H ( n ) . A quantum test is an effect 0 ≤ T n ≤ I 0\leq T_n\leq I 0 ≤ T n ≤ I on H ( n ) \mathcal{H}^{(n)} H ( n ) , or equivalently the binary POVM { T n , I − T n } \{T_n,I-T_n\} { T n , I − T n } , where the outcome associated with T n T_n T n means rejection of the null. Its power function is
π T n ( θ ) : = Tr ( ρ θ ( n ) T n ) . \pi_{T_n}(\theta):=\operatorname{Tr}\left(\rho_{\theta}^{(n)}T_n\right). π T n ( θ ) := Tr ( ρ θ ( n ) T n ) .
The i.i.d. experiment is the special case
ρ θ ( n ) = ρ θ ⊗ n , H ( n ) = H ⊗ n . \rho_{\theta}^{(n)}=\rho_{\theta}^{\otimes n},\qquad\mathcal{H}^{(n)}=\mathcal{H}^{\otimes n}. ρ θ ( n ) = ρ θ ⊗ n , H ( n ) = H ⊗ n .
Unless stated otherwise, no tensor-product structure is assumed. This formulation includes the general q-LAN model sequences considered in § 4 .
The same definitions apply to an experiment represented by normal states on a von Neumann algebra (see Appendix § 12 for details): a test is an effect in that algebra, and its rejection probability is obtained by evaluating the state at the effect. In particular, for the Gaussian state ϕ μ , Σ \phi_{\mu,\Sigma} ϕ μ , Σ , tests belong to M = CCR ( Im Σ ) \mathcal{M}=\operatorname{CCR}(\operatorname{Im}\Sigma) M = CCR ( Im Σ ) and have rejection probability ϕ μ , Σ ( T ) \phi_{\mu,\Sigma}(T) ϕ μ , Σ ( T ) . For the q-LAN limiting state indexed by the local parameter h h h , we use the notation ϱ h ( T ) \varrho_h(T) ϱ h ( T ) .
The notions of level-α \alpha α , uniformly most powerful (UMP), and uniformly most powerful unbiased (UMPU) tests are defined exactly as in the classical setting by replacing expectations with quantum expectations. Thus a quantum test T n T_n T n has level α \alpha α if
sup θ : ψ ( θ ) ∈ C π T n ( θ ) ≤ α . \sup_{\theta:\psi(\theta)\in\mathcal{C}}\pi_{T_n}(\theta)\leq\alpha. θ : ψ ( θ ) ∈ C sup π T n ( θ ) ≤ α .
and the corresponding classes of level-α \alpha α and level-α \alpha α unbiased quantum tests are denoted by T α ( n ) \mathcal{T}_{\alpha}^{(n)} T α ( n ) and T α , u ( n ) \mathcal{T}_{\alpha,u}^{(n)} T α , u ( n ) , respectively. A test T n ⋆ ∈ T α ( n ) T_n^\star\in\mathcal{T}_{\alpha}^{(n)} T n ⋆ ∈ T α ( n ) is called uniformly most powerful if
π T n ⋆ ( θ ) ≥ π T n ( θ ) , ∀ T n ∈ T α ( n ) , ∀ θ : ψ ( θ ) ∉ C , \pi_{T_n^\star}(\theta)\geq\pi_{T_n}(\theta),\qquad\forall T_n\in\mathcal{T}_{\alpha}^{(n)},\quad\forall\theta:\psi(\theta)\notin\mathcal{C}, π T n ⋆ ( θ ) ≥ π T n ( θ ) , ∀ T n ∈ T α ( n ) , ∀ θ : ψ ( θ ) ∈ / C ,
while T n ⋆ ∈ T α , u ( n ) T_n^\star\in\mathcal{T}_{\alpha,u}^{(n)} T n ⋆ ∈ T α , u ( n ) is uniformly most powerful unbiased if the same inequality holds over the smaller class T α , u ( n ) \mathcal{T}_{\alpha,u}^{(n)} T α , u ( n ) .
The formulation with nuisance parameters is likewise identical to the classical setting. Writing θ = ( η , λ ) \theta=(\eta,\lambda) θ = ( η , λ ) , the testing problem is again given by (19) , and the power function becomes
π T n ( η , λ ) = Tr ( ρ η , λ ( n ) T n ) . \pi_{T_n}(\eta,\lambda)=\operatorname{Tr}\left(\rho_{\eta,\lambda}^{(n)}T_n\right). π T n ( η , λ ) = Tr ( ρ η , λ ( n ) T n ) .
For each η ∈ Θ 1 \eta\in\Theta_1 η ∈ Θ 1 , define the minimum power function
π ‾ T n ( η ) : = inf λ ∈ Λ π T n ( η , λ ) . \underline{\pi}_{T_n}(\eta):=\inf_{\lambda\in\Lambda}\pi_{T_n}(\eta,\lambda). π T n ( η ) := λ ∈ Λ inf π T n ( η , λ ) .
A level-α \alpha α quantum test is called maximin if its minimum power dominates that of every competing level-α \alpha α test, namely,
π ‾ T n ⋆ ( η ) ≥ π ‾ T n ( η ) , ∀ T n ∈ T α ( n ) , ∀ η ∈ Θ 1 . \underline{\pi}_{T_n^\star}(\eta)\geq\underline{\pi}_{T_n}(\eta),\qquad\forall T_n\in\mathcal{T}_{\alpha}^{(n)},\quad\forall\eta\in\Theta_1. π T n ⋆ ( η ) ≥ π T n ( η ) , ∀ T n ∈ T α ( n ) , ∀ η ∈ Θ 1 .
Similarly, a test T n ∗ ∈ T α , u ( n ) T_n^* \in\mathcal{T}_{\alpha,u}^{(n)} T n ∗ ∈ T α , u ( n ) is called unbiased maximin if the same inequality holds for every T n ∈ T α , u ( n ) T_n \in\mathcal{T}_{\alpha,u}^{(n)} T n ∈ T α , u ( n ) and every η ∈ Θ 1 \eta\in\Theta_1 η ∈ Θ 1 .
The decision-theoretic notions introduced above provide the framework for studying optimal hypothesis tests in both classical and quantum statistical models. In the absence of nuisance parameters, one seeks uniformly most powerful or uniformly most powerful unbiased tests whenever they exist. In contrast, the presence of nuisance parameters naturally leads to the maximin criterion, which optimizes the minimum power over all possible values of the nuisance parameter.
Unfortunately, the resulting maximin optimization problem is typically intractable. A powerful approach to overcoming this difficulty is to exploit symmetries of the statistical model. Suppose that the nuisance parameter space admits a group action under which the statistical model is covariant. One may then restrict attention to tests that are invariant under the corresponding group action. Since the class of invariant tests is generally much smaller than the class of all tests, constructing optimal invariant tests is often considerably simpler than solving the original maximin problem.
The Hunt–Stein principle justifies a reduction to invariant tests under suitable covariance and averaging assumptions. In the nuisance-maximin form used here, a uniformly most powerful invariant test attains the best minimum power over nuisance parameters. Quantum versions of this reduction are discussed in [29 Section 3] . In Appendix § 13 , we prove the version needed here for Euclidean nuisance groups acting on a von Neumann algebra. Its normal-state formulation covers both quantum and classical–quantum Gaussian experiments and supplies the symmetry reduction used in § 3 .
Hypothesis testing in a quantum Gaussian model with nuisance parameters
In this section we establish upper bounds for power functions in quantum Gaussian models and describe tests achieving them. Fix 0 < α < 1 0 < \alpha< 1 0 < α < 1 and write Φ ( z α ) = 1 − α \Phi(z_\alpha) = 1 - \alpha Φ ( z α ) = 1 − α , where Φ \Phi Φ is the standard normal c.d.f. Throughout this section, Σ = V + i S ⪰ 0 \Sigma= V + iS \succeq0 Σ = V + i S ⪰ 0 , with V ≻ 0 V \succ0 V ≻ 0 and S S S real skew-symmetric. Throughout this section, the Gaussian shift experiment is
{ ϕ h , Σ : h ∈ R d } , ϕ h , Σ ∼ N ( h , Σ ) , \{\phi_{h,\Sigma}: h \in\mathbb{R}^{d}\}, \qquad\phi_{h,\Sigma} \sim N(h,\Sigma), { ϕ h , Σ : h ∈ R d } , ϕ h , Σ ∼ N ( h , Σ ) ,
where ϕ μ , Σ \phi_{\mu,\Sigma} ϕ μ , Σ is the normal Gaussian state defined in Appendix § 12 . All tests are effects in M = CCR ( Im Σ ) \mathcal{M} = \operatorname{CCR}(\operatorname{Im}\Sigma) M = CCR ( Im Σ ) . When writing h = ( y , z ) h = (y,z) h = ( y , z ) , we use the shorthand ϕ y , z : = ϕ ( y , z ) , Σ \phi_{y,z} := \phi_{(y,z),\Sigma} ϕ y , z := ϕ ( y , z ) , Σ . For every effect T ∈ M T \in\mathcal{M} T ∈ M , its rejection probability under mean h h h is ϕ h , Σ ( T ) \phi_{h,\Sigma}(T) ϕ h , Σ ( T ) .
One-sided tests
Write h = ( y , z ) h = (y,z) h = ( y , z ) , with y ∈ R y \in\mathbb{R} y ∈ R and z ∈ R d − 1 z \in\mathbb{R}^{d-1} z ∈ R d − 1 , and write the Gaussian coordinates as X = ( Y , Z 1 , … , Z d − 1 ) T X = (Y,Z_1,\ldots,Z_{d-1})^{T} X = ( Y , Z 1 , … , Z d − 1 ) T , so Y = X 1 Y = X_1 Y = X 1 . The hypotheses are H 0 : y ≤ 0 H_0 : y \le0 H 0 : y ≤ 0 versus H 1 : y > 0 H_1 : y > 0 H 1 : y > 0 , with z z z nuisance. A level-α \alpha α test satisfies ϕ y , z ( T ) ≤ α \phi_{y,z}(T) \le\alpha ϕ y , z ( T ) ≤ α for every y ≤ 0 y \le0 y ≤ 0 and every z z z . Unlike the classical Gaussian problem, pointwise uniformly most powerful quantum tests need not exist. We therefore compare inf z ϕ y , z ( T ) \inf_z \phi_{y,z}(T) inf z ϕ y , z ( T ) at each y > 0 y > 0 y > 0 .
Theorem 3.1. Every level-α \alpha α effect T ∈ M T \in\mathcal{M} T ∈ M satisfies
inf z ∈ R d − 1 ϕ y , z ( T ) ≤ 1 − Φ ( z α − y V 11 ) , y > 0. \inf_{z \in\mathbb{R}^{d-1}} \phi_{y,z}(T) \le1 - \Phi\left(z_\alpha- \frac{y}{\sqrt{V_{11}}}\right), \qquad y > 0. z ∈ R d − 1 inf ϕ y , z ( T ) ≤ 1 − Φ ( z α − V 11 y ) , y > 0.
Equality is attained by the spectral projection T α = 1 ( z α V 11 , ∞ ) ( Y ) T_\alpha=\mathbf{1}_{(z_\alpha\sqrt{V_{11}},\infty)}(Y) T α = 1 ( z α V 11 , ∞ ) ( Y ) , which does not depend on y y y or z z z .
The following formulation involving testing of a linear functional of h h h will be used in the q-LAN model.
Theorem 3.2. For c ∈ R d ∖ { 0 } c\in\mathbb{R}^{d}\setminus\{0\} c ∈ R d ∖ { 0 } , consider H 0 : c T h ≤ 0 H_0:c^{T}h\leq0 H 0 : c T h ≤ 0 versus H 1 : c T h > 0 H_1:c^{T}h>0 H 1 : c T h > 0 in the model { ϕ h , Σ : h ∈ R d } \{\phi_{h,\Sigma}:h\in\mathbb{R}^{d}\} { ϕ h , Σ : h ∈ R d } . Every level-α \alpha α effect T ∈ M T\in\mathcal{M} T ∈ M satisfies
inf c T h ≥ s ϕ h , Σ ( T ) ≤ 1 − Φ ( z α − s c T V c ) , s > 0. \inf_{c^{T}h\geq s}\phi_{h,\Sigma}(T)\leq1-\Phi\left(z_\alpha-\frac{s}{\sqrt{c^{T}Vc}}\right),\qquad s>0. c T h ≥ s inf ϕ h , Σ ( T ) ≤ 1 − Φ ( z α − c T V c s ) , s > 0. Equality is attained by T α = 1 ( z α c T V c , ∞ ) ( c T X ) T_\alpha=\mathbf{1}_{(z_\alpha\sqrt{c^{T}Vc},\infty)}(c^{T}X) T α = 1 ( z α c T V c , ∞ ) ( c T X ) , simultaneously for all s > 0 s>0 s > 0 .
Proof. Choose an invertible real matrix L L L with first row c T c^{T} c T . The coordinates X ′ = L X X'=LX X ′ = L X have mean L h Lh L h and covariance Σ ′ = L Σ L T \Sigma'=L\Sigma L^{T} Σ ′ = L Σ L T , with Re Σ ′ ≻ 0 \operatorname{Re}\Sigma'\succ0 Re Σ ′ ≻ 0 and V 11 ′ = c T V c V'_{11}=c^{T}Vc V 11 ′ = c T V c . Apply Theorem 3.1 at first mean coordinate s s s :
inf c T h ≥ s ϕ h , Σ ( T ) ≤ inf c T h = s ϕ h , Σ ( T ) ≤ 1 − Φ ( z α − s c T V c ) . \inf_{c^{T}h\geq s}\phi_{h,\Sigma}(T)\leq\inf_{c^{T}h=s}\phi_{h,\Sigma}(T)\leq1-\Phi\left(z_\alpha-\frac{s}{\sqrt{c^{T}Vc}}\right). c T h ≥ s inf ϕ h , Σ ( T ) ≤ c T h = s inf ϕ h , Σ ( T ) ≤ 1 − Φ ( z α − c T V c s ) .
The proposed test has power depending only on c T h c^{T}h c T h , increasing in that argument, so its infimum is its value at s s s . □
Two-sided tests
For the same model and nonzero c c c , consider H 0 : c T h = 0 H_0:c^{T}h=0 H 0 : c T h = 0 versus H 1 : c T h ≠ 0 H_1:c^{T}h\neq0 H 1 : c T h = 0 . As in the classical case UMP test does not exist in the two-sided problem, so we restrict ourselves to unbiased tests. A level-α \alpha α unbiased effect satisfies ϕ h , Σ ( T ) ≤ α \phi_{h,\Sigma}(T)\leq\alpha ϕ h , Σ ( T ) ≤ α on the null and ϕ h , Σ ( T ) ≥ α \phi_{h,\Sigma}(T)\geq\alpha ϕ h , Σ ( T ) ≥ α on the alternative. The Hunt-Stein reduction is now used within this unbiased class.
Theorem 3.3. Every level-α \alpha α unbiased effect T ∈ M T\in\mathcal{M} T ∈ M satisfies
inf ∣ c T h ∣ ≥ s ϕ h , Σ ( T ) ≤ 1 − Φ ( z α / 2 − s c T V c ) + Φ ( − z α / 2 − s c T V c ) , s > 0. \inf_{\lvert c^{T}h\rvert\geq s}\phi_{h,\Sigma}(T)\leq1-\Phi\left(z_{\alpha/2}-\frac{s}{\sqrt{c^{T}Vc}}\right)+\Phi\left(-z_{\alpha/2}-\frac{s}{\sqrt{c^{T}Vc}}\right),\qquad s>0. ∣ c T h ∣ ≥ s inf ϕ h , Σ ( T ) ≤ 1 − Φ ( z α /2 − c T V c s ) + Φ ( − z α /2 − c T V c s ) , s > 0. Equality is attained by T α = 1 { ∣ x ∣ > z α / 2 c T V c } ( c T X ) T_\alpha=\mathbf{1}_{\{\lvert x\rvert>z_{\alpha/2}\sqrt{c^{T}Vc}\}}(c^{T}X) T α = 1 {∣ x ∣ > z α /2 c T V c } ( c T X ) , simultaneously for all s > 0 s>0 s > 0 .
Efficient testing in q-LAN models
In this section, we study efficient testing in q-LAN models. The upper bounds on power follow from Theorem 2.2 and the Gaussian testing results of § 3 . For motivation and for completeness, the corresponding classical theory is recalled in Appendix § 11 .
Let S ( n ) = { ρ θ ( n ) : θ ∈ Θ } \mathcal{S}^{(n)}=\{\rho_{\theta}^{(n)}:\theta\in\Theta\} S ( n ) = { ρ θ ( n ) : θ ∈ Θ } be a q-LAN model sequence whose central sequence is asymptotically D \mathcal{D} D -invariant at θ 0 ∈ Θ \theta_0\in\Theta θ 0 ∈ Θ , where Θ ⊂ R d \Theta\subset\mathbb{R}^{d} Θ ⊂ R d is open. Write
Δ θ 0 ( n ) ⟶ D N ( 0 , J ) , G : = Re J ≻ 0 , \Delta_{\theta_0}^{(n)}\overset{\mathcal{D}}{\longrightarrow}N(0,J),\qquad G:=\operatorname{Re}J\succ0, Δ θ 0 ( n ) ⟶ D N ( 0 , J ) , G := Re J ≻ 0 ,
and define, for this general model sequence,
X ( n ) : = G − 1 Δ ( n ) , Σ : = G − 1 J G − 1 . X^{(n)}:=G^{-1}\Delta^{(n)},\qquad\Sigma:=G^{-1}JG^{-1}. X ( n ) := G − 1 Δ ( n ) , Σ := G − 1 J G − 1 .
In the i.i.d. case ρ θ ( n ) = ρ θ ⊗ n \rho_{\theta}^{(n)}=\rho_{\theta}^{\otimes n} ρ θ ( n ) = ρ θ ⊗ n , let L θ 0 , j L_{\theta_0,j} L θ 0 , j be the single-copy SLDs from (12) , and write
L θ 0 , j [ a ] : = I ⊗ ( a − 1 ) ⊗ L θ 0 , j ⊗ I ⊗ ( n − a ) . L_{\theta_0,j}^{[a]}:=I^{\otimes(a-1)}\otimes L_{\theta_0,j}\otimes I^{\otimes(n-a)}. L θ 0 , j [ a ] := I ⊗ ( a − 1 ) ⊗ L θ 0 , j ⊗ I ⊗ ( n − a ) .
The SLD central sequence is
Δ j ( n ) : = 1 n ∑ a = 1 n L θ 0 , j [ a ] , j = 1 , … , d , (20) \Delta_j^{(n)}:=\frac{1}{\sqrt{n}}\sum_{a=1}^{n}L_{\theta_0,j}^{[a]},\qquad j=1,\ldots,d, \tag*{(20)} Δ j ( n ) := n 1 a = 1 ∑ n L θ 0 , j [ a ] , j = 1 , … , d , (20)
with J = J θ 0 J=J_{\theta_0} J = J θ 0 and G = G θ 0 G=G_{\theta_0} G = G θ 0 as in (14) .
Under the assumptions of this section, the sandwiched Le Cam third lemma (Corollary 3.7 of [14 ] ) gives, for every fixed h ∈ R d h\in\mathbb{R}^{d} h ∈ R d ,
( X ( n ) , ρ θ 0 + h / n ( n ) ) ⇝ N ( h , Σ ) , Σ = G − 1 J G − 1 . (21) \left(X^{(n)},\rho_{\theta_0+h/\sqrt{n}}^{(n)}\right)\rightsquigarrow N(h,\Sigma),\qquad\Sigma=G^{-1}JG^{-1}. \tag*{(21)} ( X ( n ) , ρ θ 0 + h / n ( n ) ) ⇝ N ( h , Σ ) , Σ = G − 1 J G − 1 . (21)
In the asymptotic representation theorem, this normalized extension has F = G F=G F = G and ( Re Σ ) F = I (\operatorname{Re}\Sigma)F=I ( Re Σ ) F = I . Its limiting states are therefore ϱ h = ϕ h , Σ \varrho_h=\phi_{h,\Sigma} ϱ h = ϕ h , Σ , so the Gaussian testing results of § 3 apply directly.
For finite n n n , write
Θ n : = { h ∈ R d : θ 0 + h / n ∈ Θ } . \Theta_n:=\{h\in\mathbb{R}^{d}:\theta_0+h/\sqrt{n}\in\Theta\}. Θ n := { h ∈ R d : θ 0 + h / n ∈ Θ } .
All finite-sample local probabilities below are understood on Θ n \Theta_n Θ n . Every fixed finite set of local parameters is contained in Θ n \Theta_n Θ n for all sufficiently large n n n . In the finite-submodel power criteria below, we adopt the convention inf ∅ = + ∞ \inf\varnothing=+\infty inf ∅ = + ∞ .
One-sided tests
By the quantum asymptotic representation theorem, limiting measurement outcome laws in the q-LAN model admit a representation in the Gaussian shift experiment S ( ∞ ) = { ϱ h = ϕ h , Σ : h ∈ R d } S^{(\infty)}=\{\varrho_h=\phi_{h,\Sigma}:h\in\mathbb{R}^{d}\} S ( ∞ ) = { ϱ h = ϕ h , Σ : h ∈ R d } . Together with Theorem 3.2 , this yields the following local asymptotic power bound.
Theorem 4.1. Let Θ ⊂ R d \Theta\subset\mathbb{R}^{d} Θ ⊂ R d be open and let ψ : Θ → R \psi:\Theta\to\mathbb{R} ψ : Θ → R be differentiable at θ 0 \theta_0 θ 0 , with nonzero gradient ψ ˙ θ 0 \dot{\psi}_{\theta_0} ψ ˙ θ 0 and ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 . Consider the testing problem
{ H 0 : ψ ( θ ) ≤ 0 , H 1 : ψ ( θ ) > 0. (22) \begin{cases}
H_0:\ \psi(\theta)\leq0,\\
H_1:\ \psi(\theta)>0. \tag*{(22)}
\end{cases} { H 0 : ψ ( θ ) ≤ 0 , H 1 : ψ ( θ ) > 0. (22) Every sequence of level-α \alpha α tests T ( n ) T^{(n)} T ( n ) for the original null, i.e.,
sup θ : ψ ( θ ) ≤ 0 Tr ( ρ θ ( n ) T ( n ) ) ≤ α for every n , \sup_{\theta:\psi(\theta)\leq0}\operatorname{Tr}\left(\rho_{\theta}^{(n)}T^{(n)}\right)\leq\alpha\qquad\text{for every }n, θ : ψ ( θ ) ≤ 0 sup Tr ( ρ θ ( n ) T ( n ) ) ≤ α for every n , satisfies, for every s > 0 s>0 s > 0 ,
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) Tr ( ρ θ 0 + h / n ( n ) T ( n ) ) ≤ 1 − Φ ( z α − s ψ ˙ θ 0 Σ ψ ˙ θ 0 T ) , (23) \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\operatorname{Tr}\left(\rho_{\theta_0+h/\sqrt{n}}^{(n)}T^{(n)}\right)
\leq1-\Phi\left(z_{\alpha}-\frac{s}{\sqrt{\dot{\psi}_{\theta_0}\Sigma\dot{\psi}_{\theta_0}^{\mathsf T}}}\right), \tag*{(23)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf Tr ( ρ θ 0 + h / n ( n ) T ( n ) ) ≤ 1 − Φ z α − ψ ˙ θ 0 Σ ψ ˙ θ 0 T s , (23)
where D D D ranges over nonempty finite subsets of R d \mathbb{R}^{d} R d and
H 1 , s ( n ) : = { h ∈ R d : n ψ ( θ 0 + h n ) ≥ s } . H_{1,s}^{(n)} := \left\{h \in\mathbb{R}^{d} : \sqrt{n}\,\psi\left(\theta_{0}+\frac{h}{\sqrt{n}}\right) \ge s\right\}. H 1 , s ( n ) := { h ∈ R d : n ψ ( θ 0 + n h ) ≥ s } .
Moreover, the scalar tests
T α ( n ) = { ψ ˙ θ 0 T X ( n ) > z α ψ ˙ θ 0 Σ ψ ˙ θ 0 T } T_{\alpha}^{(n)}=\left\{\dot{\psi}_{\theta_{0}}^{T}X^{(n)}>z_{\alpha}\sqrt{\dot{\psi}_{\theta_{0}}\Sigma\dot{\psi}_{\theta_{0}}^{T}}\right\} T α ( n ) = { ψ ˙ θ 0 T X ( n ) > z α ψ ˙ θ 0 Σ ψ ˙ θ 0 T }
attain equality in the local asymptotic power criterion (23) for every s > 0 s>0 s > 0 , and satisfy
lim n → ∞ Tr ( ρ θ 0 ( n ) T α ( n ) ) = α . \lim_{n\to\infty}\operatorname{Tr}\left(\rho_{\theta_{0}}^{(n)}T_{\alpha}^{(n)}\right)=\alpha. n → ∞ lim Tr ( ρ θ 0 ( n ) T α ( n ) ) = α .
Remark 2. The upper bounds in Theorems 4.1 , 4.2, and 4.3 apply to tests satisfying exact finite-sample level, with unbiasedness additionally required in the two-sided cases. The proposed scalar tests attain the corresponding Gaussian power bounds in the stated local asymptotic criteria, but are not asserted to satisfy these exact finite-sample constraints. All claims of asymptotic optimality, including those in the introduction, are understood in this sense.
The scalar test constructed in this theorem is an oracle procedure because its defining observable depends on θ 0 \theta_{0} θ 0 . For the smooth finite-dimensional i.i.d. models considered in § 5 , a two-stage adaptive construction attains the same local Gaussian power without knowledge of θ 0 \theta_{0} θ 0 . It uses single-copy measurements and may be regarded as a quantum analogue of a Wald procedure.
Two-sided tests
We consider two formulations of the two-sided testing problem: the original null hypothesis ψ ( θ ) = 0 \psi(\theta)=0 ψ ( θ ) = 0 and its linearized counterpart ψ ˙ θ 0 h = 0 \dot{\psi}_{\theta_{0}}h=0 ψ ˙ θ 0 h = 0 in the local parameter. The first requires an additional continuity assumption to transfer the level condition to the tangent null, whereas the second is treated directly under the q-LAN assumptions.
Theorem 4.2. Let Θ ⊂ R d \Theta\subset\mathbb{R}^{d} Θ ⊂ R d be open, and let ψ : Θ → R \psi:\Theta\to\mathbb{R} ψ : Θ → R be continuous in a neighborhood of θ 0 \theta_{0} θ 0 and differentiable at θ 0 \theta_{0} θ 0 , with nonzero gradient ψ ˙ θ 0 \dot{\psi}_{\theta_{0}} ψ ˙ θ 0 and ψ ( θ 0 ) = 0 \psi(\theta_{0})=0 ψ ( θ 0 ) = 0 . Consider the testing problem
{ H 0 : ψ ( θ ) = 0 , H 1 : ψ ( θ ) ≠ 0. (24) \begin{cases}
H_{0}: \psi(\theta)=0,\\
H_{1}: \psi(\theta)\ne0. \tag*{(24)}
\end{cases} { H 0 : ψ ( θ ) = 0 , H 1 : ψ ( θ ) = 0. (24) Suppose further that, for every sequence h n → h ∈ R d h_{n}\to h\in\mathbb{R}^{d} h n → h ∈ R d ,
∥ ρ θ 0 + h n / n ( n ) − ρ θ 0 + h / n ( n ) ∥ 1 → 0. (25) \left\|\rho_{\theta_{0}+h_{n}/\sqrt{n}}^{(n)}-\rho_{\theta_{0}+h/\sqrt{n}}^{(n)}\right\|_{1}\to0. \tag*{(25)} ρ θ 0 + h n / n ( n ) − ρ θ 0 + h / n ( n ) 1 → 0. (25) Under the asymptotically D \mathcal{D} D -invariant q-LAN assumptions of this section, every sequence of level-α \alpha α unbiased tests T ( n ) T^{(n)} T ( n ) for the original hypotheses satisfies, for every s > 0 s>0 s > 0 ,
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) Tr ( ρ θ 0 + h / n ( n ) T ( n ) ) ≤ 1 − Φ ( − z α / 2 − s ψ ˙ θ 0 Σ ψ ˙ θ 0 T , z α / 2 − s ψ ˙ θ 0 Σ ψ ˙ θ 0 T ) , (26) \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\operatorname{Tr}\left(\rho_{\theta_{0}+h/\sqrt{n}}^{(n)}T^{(n)}\right)
\le1-\Phi\left(-z_{\alpha/2}-\frac{s}{\sqrt{\dot{\psi}_{\theta_{0}}\Sigma\dot{\psi}_{\theta_{0}}^{T}}},z_{\alpha/2}-\frac{s}{\sqrt{\dot{\psi}_{\theta_{0}}\Sigma\dot{\psi}_{\theta_{0}}^{T}}}\right), \tag*{(26)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf Tr ( ρ θ 0 + h / n ( n ) T ( n ) ) ≤ 1 − Φ − z α /2 − ψ ˙ θ 0 Σ ψ ˙ θ 0 T s , z α /2 − ψ ˙ θ 0 Σ ψ ˙ θ 0 T s , (26)
where D D D ranges over nonempty finite subsets of R d \mathbb{R}^{d} R d and
H 1 , s ( n ) : = { h ∈ R d : ∣ n ψ ( θ 0 + h n ) ∣ ≥ s } . H_{1,s}^{(n)} := \left\{h \in\mathbb{R}^{d} : \left|\sqrt{n}\,\psi\left(\theta_{0}+\frac{h}{\sqrt{n}}\right)\right| \ge s\right\}. H 1 , s ( n ) := { h ∈ R d : n ψ ( θ 0 + n h ) ≥ s } .
Moreover, the scalar tests
T α ( n ) = { ∣ ψ ˙ θ 0 X ( n ) ∣ > z α / 2 ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T } T_{\alpha}^{(n)}=\left\{\left|\dot{\psi}_{\theta_{0}}X^{(n)}\right|>z_{\alpha/2}\sqrt{\dot{\psi}_{\theta_{0}}\Sigma_{\theta_{0}}\dot{\psi}_{\theta_{0}}^{T}}\right\} T α ( n ) = { ψ ˙ θ 0 X ( n ) > z α /2 ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T }
attain equality in the local asymptotic power criterion (26) for every s > 0 s>0 s > 0 , and satisfy
lim n → ∞ Tr ( ρ θ 0 ( n ) T α ( n ) ) = α . \lim_{n\to\infty}\operatorname{Tr}\left(\rho_{\theta_{0}}^{(n)}T_{\alpha}^{(n)}\right)=\alpha. n → ∞ lim Tr ( ρ θ 0 ( n ) T α ( n ) ) = α .
When the hypotheses are formulated directly in the local tangent parameter, the corresponding result holds without assumption (25) , as stated next.
Theorem 4.3. Let Θ ⊂ R d \Theta\subset\mathbb{R}^{d} Θ ⊂ R d be open and let ψ : Θ → R \psi:\Theta\to\mathbb{R} ψ : Θ → R be differentiable at θ 0 \theta_{0} θ 0 , with ψ ( θ 0 ) = 0 \psi(\theta_{0})=0 ψ ( θ 0 ) = 0 and nonzero gradient ψ ˙ θ 0 \dot{\psi}_{\theta_{0}} ψ ˙ θ 0 . Under the asymptotically D D D -invariant q-LAN assumptions of this section, put c : = ψ ˙ θ 0 T c:=\dot{\psi}_{\theta_{0}}^{T} c := ψ ˙ θ 0 T and consider the tangent testing problem
{ H 0 : c T h = 0 , H 1 : c T h ≠ 0. (27) \begin{cases}
H_{0}: \ c^{T}h=0,\\
H_{1}: \ c^{T}h\ne0. \tag*{(27)}
\end{cases} { H 0 : c T h = 0 , H 1 : c T h = 0. (27) For each n n n , let Θ n : = { h ∈ R d : θ 0 + h / n ∈ Θ } \Theta_{n}:=\{h\in\mathbb{R}^{d}:\theta_{0}+h/\sqrt{n}\in\Theta\} Θ n := { h ∈ R d : θ 0 + h / n ∈ Θ } . Every sequence of level-α \alpha α unbiased tests T ( n ) T^{(n)} T ( n ) for these local hypotheses satisfies, for each s > 0 s>0 s > 0 ,
inf ∅ ≠ D ⊂ H 1 , s ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D Tr ( ρ θ 0 + h / n ( n ) T ( n ) ) ≤ 1 − Φ ( − z α / 2 − s ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T , z α / 2 − s ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T ) . (28) \inf_{\substack{\varnothing\ne D\subset H_{1,s}\\ |D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D}\operatorname{Tr}\left(\rho_{\theta_{0}+h/\sqrt{n}}^{(n)}T^{(n)}\right)
\le1-\Phi\left(-z_{\alpha/2}-\frac{s}{\sqrt{\dot{\psi}_{\theta_{0}}\Sigma_{\theta_{0}}\dot{\psi}_{\theta_{0}}^{T}}},z_{\alpha/2}-\frac{s}{\sqrt{\dot{\psi}_{\theta_{0}}\Sigma_{\theta_{0}}\dot{\psi}_{\theta_{0}}^{T}}}\right). \tag*{(28)} ∅ = D ⊂ H 1 , s ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D inf Tr ( ρ θ 0 + h / n ( n ) T ( n ) ) ≤ 1 − Φ − z α /2 − ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T s , z α /2 − ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T s . (28) where
H 1 , s : = { h ∈ R d : ∣ c T h ∣ ≥ s } . H_{1,s}:=\{h\in\mathbb{R}^{d}:|c^{T}h|\ge s\}. H 1 , s := { h ∈ R d : ∣ c T h ∣ ≥ s } . Here level and unbiasedness are understood in the finite-sample sense for the local experiment indexed by Θ n \Theta_{n} Θ n .
Moreover, the scalar tests
T α ( n ) = { ∣ ψ ˙ θ 0 X ( n ) ∣ > z α / 2 ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T } T_{\alpha}^{(n)}=\left\{\left|\dot{\psi}_{\theta_{0}}X^{(n)}\right|>z_{\alpha/2}\sqrt{\dot{\psi}_{\theta_{0}}\Sigma_{\theta_{0}}\dot{\psi}_{\theta_{0}}^{T}}\right\} T α ( n ) = { ψ ˙ θ 0 X ( n ) > z α /2 ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T } attain equality in (28) for every s > 0 s>0 s > 0 . They have limiting rejection probability α \alpha α at every fixed h h h satisfying c T h = 0 c^{T}h=0 c T h = 0 , and limiting rejection probability at least α \alpha α at every fixed alternative.
Quantum adaptive Wald tests
The scalar SLD tests depend on the unknown parameter through the measurement itself. We implement them using a preliminary estimate on a vanishing fraction of the copies, followed by a scalar measurement on the remaining copies. A one-step correction removes the first-order error of the preliminary estimate. The resulting procedure has the same limiting local power as the oracle scalar SLD test, while using no knowledge of the true parameter. Fujiwara’s sequential adaptive method remains an alternative of practical interest.
Model and scalar influence observable
Let Θ ⊂ R d \Theta\subset\mathbb{R}^{d} Θ ⊂ R d be open, and let { ρ θ : θ ∈ Θ } \{\rho_{\theta}:\theta\in\Theta\} { ρ θ : θ ∈ Θ } be a model on H = C r \mathcal{H} = \mathbb{C}^{r} H = C r , where r ≥ 2 r \ge2 r ≥ 2 is fixed. Assume that θ ↦ ρ θ \theta\mapsto\rho_{\theta} θ ↦ ρ θ is twice continuously differentiable in trace norm and that rank ρ θ = k \operatorname{rank}\rho_{\theta} = k rank ρ θ = k for every θ ∈ Θ \theta\in\Theta θ ∈ Θ , where 1 ≤ k ≤ r 1 \le k \le r 1 ≤ k ≤ r is fixed. Let L θ , j L_{\theta,j} L θ , j , j = 1 , … , d j = 1,\ldots,d j = 1 , … , d , denote any operator-norm continuous choice of SLDs satisfying (12) . Such a choice exists under the constant-rank assumption; see Appendix § 17 .§ 1 . Let J θ J_{\theta} J θ and G θ = Re J θ G_{\theta} = \operatorname{Re} J_{\theta} G θ = Re J θ be as in (14) . Assume G θ ≻ 0 G_{\theta} \succ0 G θ ≻ 0 on Θ \Theta Θ . Let ψ : Θ → R \psi:\Theta\to\mathbb{R} ψ : Θ → R be C 2 C^{2} C 2 , and write c θ = ∇ ψ ( θ ) = ψ ˙ θ T c_{\theta} = \nabla\psi(\theta) = \dot{\psi}_{\theta}^{T} c θ = ∇ ψ ( θ ) = ψ ˙ θ T as a column vector. The hypotheses are
H 0 + : ψ ( θ ) ≤ 0 , H 1 + : ψ ( θ ) > 0 , H 0 ± : ψ ( θ ) = 0 , H 1 ± : ψ ( θ ) ≠ 0. (29) \begin{aligned}
H_{0}^{+}&:\ \psi(\theta) \le0, & H_{1}^{+}&:\ \psi(\theta) > 0,\\
H_{0}^{\pm}&:\ \psi(\theta) = 0, & H_{1}^{\pm}&:\ \psi(\theta) \ne0. \tag*{(29)}
\end{aligned} H 0 + H 0 ± : ψ ( θ ) ≤ 0 , : ψ ( θ ) = 0 , H 1 + H 1 ± : ψ ( θ ) > 0 , : ψ ( θ ) = 0. (29)
Define
v θ : = G θ − 1 c θ , Y θ : = ∑ j = 1 d ( v θ ) j L θ , j , V e f f ( θ ) : = c θ T G θ − 1 c θ . (30) v_{\theta} := G_{\theta}^{-1}c_{\theta}, \qquad
Y_{\theta} := \sum_{j=1}^{d}(v_{\theta})_{j}L_{\theta,j}, \qquad
V_{\mathrm{eff}}(\theta) := c_{\theta}^{T}G_{\theta}^{-1}c_{\theta}. \tag*{(30)} v θ := G θ − 1 c θ , Y θ := j = 1 ∑ d ( v θ ) j L θ , j , V eff ( θ ) := c θ T G θ − 1 c θ . (30)
The assumptions imply that Y θ Y_{\theta} Y θ is operator-norm continuous and locally bounded. Its spectral measure is M η ( B ) : = 1 B ( Y η ) M_{\eta}(B) := \mathbf{1}_{B}(Y_{\eta}) M η ( B ) := 1 B ( Y η ) for Borel sets B ⊂ R B \subset\mathbb{R} B ⊂ R .
We distinguish the true-state parameter ϑ \vartheta ϑ from the measurement-design parameter η \eta η . For fixed η \eta η , the outcome law is P ϑ ; η ( B ) = Tr { ρ ϑ M η ( B ) } P_{\vartheta;\eta}(B) = \operatorname{Tr}\{\rho_{\vartheta}M_{\eta}(B)\} P ϑ ; η ( B ) = Tr { ρ ϑ M η ( B )} , with mean and variance
μ ( ϑ , η ) : = Tr ( ρ ϑ Y η ) , ν ( ϑ , η ) : = Tr ( ρ ϑ Y η 2 ) − μ ( ϑ , η ) 2 . (31) \mu(\vartheta,\eta) := \operatorname{Tr}(\rho_{\vartheta}Y_{\eta}), \qquad
\nu(\vartheta,\eta) := \operatorname{Tr}(\rho_{\vartheta}Y_{\eta}^{2})-\mu(\vartheta,\eta)^{2}. \tag*{(31)} μ ( ϑ , η ) := Tr ( ρ ϑ Y η ) , ν ( ϑ , η ) := Tr ( ρ ϑ Y η 2 ) − μ ( ϑ , η ) 2 . (31)
All derivatives of this outcome law used below are with respect to ϑ \vartheta ϑ , holding η \eta η fixed. The SLD equations give
μ ( η , η ) = 0 , ∇ ϑ μ ( ϑ , η ) ∣ ϑ = η = c η , ν ( η , η ) = V e f f ( η ) . (32) \mu(\eta,\eta)=0,\qquad
\nabla_{\vartheta}\mu(\vartheta,\eta)\big|_{\vartheta=\eta}=c_{\eta},\qquad
\nu(\eta,\eta)=V_{\mathrm{eff}}(\eta). \tag*{(32)} μ ( η , η ) = 0 , ∇ ϑ μ ( ϑ , η ) ϑ = η = c η , ν ( η , η ) = V eff ( η ) . (32)
Consequently, on a sufficiently small neighborhood of any fixed point,
∣ ψ ( η ) + μ ( ϑ , η ) − ψ ( ϑ ) ∣ ≤ C ∥ ϑ − η ∥ 2 . (33) \left|\psi(\eta)+\mu(\vartheta,\eta)-\psi(\vartheta)\right| \le C\|\vartheta-\eta\|^{2}. \tag*{(33)} ∣ ψ ( η ) + μ ( ϑ , η ) − ψ ( ϑ ) ∣ ≤ C ∥ ϑ − η ∥ 2 . (33)
The constant is uniform for both parameters in that neighborhood. The identities and the bound are proved in Appendix § 17 .
Preliminary localization
Fix a compact set K ⊂ Θ K \subset\Theta K ⊂ Θ with nonempty interior, specified independently of the unknown parameter. Choose integers m n < n m_{n}<n m n < n such that
m n → ∞ , m n n → 0 , n m n → 0 , N n : = n − m n . (34) m_{n}\to\infty,\qquad\frac{m_{n}}{n}\to0,\qquad\frac{\sqrt{n}}{m_{n}}\to0,\qquad N_{n}:=n-m_{n}. \tag*{(34)} m n → ∞ , n m n → 0 , m n n → 0 , N n := n − m n . (34)
For example, m n = ⌊ n 2 / 3 ⌋ m_{n}=\lfloor n^{2/3}\rfloor m n = ⌊ n 2/3 ⌋ satisfies these conditions. Use the first m n m_{n} m n copies to construct a measurable K K K -valued pilot estimate θ ~ n \widetilde{\theta}_{n} θ n . The pilot measurement and its processing rule may be adaptive but must not depend on the unknown true parameter. The required property is
m n ( θ ~ n − θ ) = O P θ ( n ) ( 1 ) (35) \sqrt{m_{n}}(\widetilde{\theta}_{n}-\theta)=O_{\mathbb{P}_{\theta}^{(n)}}(1) \tag*{(35)} m n ( θ n − θ ) = O P θ ( n ) ( 1 ) (35)
for each fixed θ ∈ int K \theta\in\operatorname{int}K θ ∈ int K .
Here P θ ( n ) \mathbb{P}_{\theta}^{(n)} P θ ( n ) denotes the classical law of all recorded outcomes of the procedure on ρ θ ⊗ n \rho_{\theta}^{\otimes n} ρ θ ⊗ n ; a statement about the pilot uses only its marginal law. In particular, no uniform local-alternative rate is assumed in (35) .
Fujiwara’s sequential-design consistency and asymptotic-normality results [11 Theorems 8–9] provide such a pilot when their positivity, identifiability, smoothness, and design-selection hypotheses are satisfied; see also [12 ] .
The following fixed-measurement construction supplies a concrete alternative and also fits the fixed-design specialization of Fujiwara’s framework under his smoothness hypotheses.
Proposition 5.1 (An informationally complete pilot). In addition to the assumptions of Subsection 5.1, assume that θ ↦ ρ θ \theta\mapsto\rho_\theta θ ↦ ρ θ is injective on K K K . Put s = r 2 − 1 s=r^2-1 s = r 2 − 1 , choose a real basis H 1 , … , H s H_1,\ldots,H_s H 1 , … , H s of the traceless Hermitian matrices with ∥ H j ∥ o p ≤ 1 \lVert H_j\rVert_{\mathrm{op}}\le1 ∥ H j ∥ op ≤ 1 , and fix ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) . The effects
M j , ± p i l : = I ± ε H j 2 s , j = 1 , … , s , (36) M^{\mathrm{pil}}_{j,\pm}:=\frac{I\pm\varepsilon H_j}{2s},\qquad j=1,\ldots,s, \tag*{(36)} M j , ± pil := 2 s I ± ε H j , j = 1 , … , s , (36) form a strictly positive informationally complete POVM. Let X = { 1 , … , s } × { + , − } \mathcal{X}=\{1,\ldots,s\}\times\{+,-\} X = { 1 , … , s } × { + , − } and p θ ( x ) = Tr ( ρ θ M x p i l ) p_\theta(x)=\operatorname{Tr}(\rho_\theta M_x^{\mathrm{pil}}) p θ ( x ) = Tr ( ρ θ M x pil ) . Measure this POVM separately on the pilot copies and, for outcomes W 1 , … , W m n W_1,\ldots,W_{m_n} W 1 , … , W m n , choose
θ ~ n ∈ argmax η ∈ K ∑ a = 1 m n log p η ( W a ) , (37) \widetilde{\theta}_n\in\underset{\eta\in K}{\operatorname{argmax}}\sum_{a=1}^{m_n}\log p_\eta(W_a), \tag*{(37)} θ n ∈ η ∈ K argmax a = 1 ∑ m n log p η ( W a ) , (37) using a fixed choice for each possible outcome-count vector. For every θ ∈ int K \theta\in\operatorname{int}K θ ∈ int K , this estimator is consistent and satisfies
m n ( θ ~ n − θ ) ⇒ N d ( 0 , ( I θ p i l ) − 1 ) , I θ p i l : = ∑ x ∈ X ∇ p θ ( x ) ∇ p θ ( x ) T p θ ( x ) ≻ 0. (38) \sqrt{m_n}(\widetilde{\theta}_n-\theta)\Rightarrow N_d\left(0,(I_\theta^{\mathrm{pil}})^{-1}\right),\qquad I_\theta^{\mathrm{pil}}:=\sum_{x\in\mathcal{X}}\frac{\nabla p_\theta(x)\nabla p_\theta(x)^{\mathsf T}}{p_\theta(x)}\succ0. \tag*{(38)} m n ( θ n − θ ) ⇒ N d ( 0 , ( I θ pil ) − 1 ) , I θ pil := x ∈ X ∑ p θ ( x ) ∇ p θ ( x ) ∇ p θ ( x ) T ≻ 0. (38) In particular, (35) holds.
The proof of Proposition 5.1 is given in Appendix § 17 .
The two-stage estimator and tests
Let F n p i l \mathcal{F}_n^{\mathrm{pil}} F n pil be the sigma-field generated by the pilot data. Recall that M η ( B ) = 1 B ( Y η ) M_\eta(B)=\mathbf{1}_B(Y_\eta) M η ( B ) = 1 B ( Y η ) , with Y η Y_\eta Y η defined in (30) . Now measure M θ ~ n M_{\widetilde{\theta}_n} M θ n separately on each of the remaining N n N_n N n copies, and record the numerical outcomes Z n , 1 , … , Z n , N n Z_{n,1},\ldots,Z_{n,N_n} Z n , 1 , … , Z n , N n . Under the true parameter ϑ \vartheta ϑ , these outcomes are conditionally independent with common law P ϑ ; θ ~ n P_{\vartheta;\widetilde{\theta}_n} P ϑ ; θ n . Define
ψ ^ n : = ψ ( θ ~ n ) + Z ‾ n , Z ‾ n : = 1 N n ∑ a = 1 N n Z n , a . (39) \widehat{\psi}_n:=\psi(\widetilde{\theta}_n)+\overline{Z}_n,\qquad\overline{Z}_n:=\frac{1}{N_n}\sum_{a=1}^{N_n}Z_{n,a}. \tag*{(39)} ψ n := ψ ( θ n ) + Z n , Z n := N n 1 a = 1 ∑ N n Z n , a . (39)
The correction has conditional mean μ ( ϑ , θ ~ n ) \mu(\vartheta,\widetilde{\theta}_n) μ ( ϑ , θ n ) , so (33) makes the residual conditional bias quadratic in the pilot error. This explains the requirement n / m n → 0 \sqrt{n}/m_n\to0 n / m n → 0 . The condition m n / n → 0 m_n/n\to0 m n / n → 0 ensures that the preliminary stage has no first-order cost in the final variance.
Now set
V ^ n : = max { V e f f ( θ ~ n ) , n − 1 } , W n : = n ψ ^ n V ^ n . (40) \widehat{V}_n:=\max\{V_{\mathrm{eff}}(\widetilde{\theta}_n),n^{-1}\},\qquad W_n:=\frac{\sqrt{n}\widehat{\psi}_n}{\sqrt{\widehat{V}_n}}. \tag*{(40)} V n := max { V eff ( θ n ) , n − 1 } , W n := V n n ψ n . (40)
The deterministic lower bound makes the statistic well-defined even if the pilot falls at a critical point of ψ \psi ψ ; it is asymptotically inactive at the regular reference points considered below. For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) let z α = Φ − 1 ( 1 − α ) z_\alpha=\Phi^{-1}(1-\alpha) z α = Φ − 1 ( 1 − α ) , where Φ \Phi Φ is the standard normal distribution function. The one-sided test rejects when W n > z α W_n > z_\alpha W n > z α , and the two-sided test rejects when ∣ W n ∣ > z α / 2 |W_n| > z_{\alpha/2} ∣ W n ∣ > z α /2 . Denote the corresponding quantum effects on H ⊗ n \mathcal{H}^{\otimes n} H ⊗ n by T n , α + T^+_{n,\alpha} T n , α + and T n , α ± T^\pm_{n,\alpha} T n , α ± . Both effects incorporate the pilot measurement, the pilot-dependent second-stage measurement, and the final classical decision. We note that with the pilot in Proposition 5.1 , every quantum measurement acts on a single copy.
Local limit and asymptotic power
Theorem 5.2 (Local limit of the adaptive scalar estimator). Under the assumptions of Subsection 5.1, suppose K K K , the pilot, and the sample split satisfy (34)–(35). Fix θ 0 ∈ int K \theta_0 \in\operatorname{int} K θ 0 ∈ int K with c 0 : = c θ 0 ≠ 0 c_0 := c_{\theta_0} \ne0 c 0 := c θ 0 = 0 , and set V 0 : = V e f f ( θ 0 ) > 0 V_0 := V_{\mathrm{eff}}(\theta_0) > 0 V 0 := V eff ( θ 0 ) > 0 . For every sequence h n → h ∈ R d h_n \to h \in\mathbb{R}^d h n → h ∈ R d , put θ n = θ 0 + h n / n \theta_n = \theta_0 + h_n/\sqrt{n} θ n = θ 0 + h n / n . Under P θ n ( n ) P_{\theta_n}^{(n)} P θ n ( n ) ,
m n ( θ ~ n − θ n ) = O P θ n ( n ) ( 1 ) , V ^ n → V 0 in probability , \sqrt{m_n}(\widetilde{\theta}_n-\theta_n)=O_{P_{\theta_n}^{(n)}}(1), \qquad\widehat{V}_n \to V_0 \quad\text{in probability}, m n ( θ n − θ n ) = O P θ n ( n ) ( 1 ) , V n → V 0 in probability , n { ψ ^ n − ψ ( θ n ) } ⇒ N ( 0 , V 0 ) . (41) \sqrt{n}\{\widehat{\psi}_n-\psi(\theta_n)\} \Rightarrow N(0,V_0). \tag*{(41)} n { ψ n − ψ ( θ n )} ⇒ N ( 0 , V 0 ) . (41) If also ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 , then
n ψ ^ n ⇒ N ( c 0 T h , V 0 ) , W n ⇒ N ( c 0 T h / V 0 , 1 ) . (42) \sqrt{n}\widehat{\psi}_n \Rightarrow N(c_0^{T}h,V_0), \qquad W_n \Rightarrow N(c_0^{T}h/\sqrt{V_0},1). \tag*{(42)} n ψ n ⇒ N ( c 0 T h , V 0 ) , W n ⇒ N ( c 0 T h / V 0 , 1 ) . (42)
Corollary 5.3 (Local power and local asymptotic level). Suppose the assumptions of Theorem 5.2 hold and ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 . Define
β α + ( t ; V ) : = 1 − Φ ( z α − t / V ) , \beta_\alpha^+(t;V):=1-\Phi(z_\alpha-t/\sqrt{V}), β α + ( t ; V ) := 1 − Φ ( z α − t / V ) , β α ± ( t ; V ) : = 1 − Φ ( z α / 2 − t / V ) + Φ ( − z α / 2 − t / V ) . (43) \beta_\alpha^\pm(t;V):=1-\Phi(z_{\alpha/2}-t/\sqrt{V})+\Phi(-z_{\alpha/2}-t/\sqrt{V}). \tag*{(43)} β α ± ( t ; V ) := 1 − Φ ( z α /2 − t / V ) + Φ ( − z α /2 − t / V ) . (43) For ∙ ∈ { + , ± } \bullet\in\{+,\pm\} ∙ ∈ { + , ± } , set π n ∙ ( h ) = Tr ( ρ θ 0 + h / n ⊗ n T n , α ∙ ) \pi_n^\bullet(h)=\operatorname{Tr}(\rho_{\theta_0+h/\sqrt{n}}^{\otimes n}T_{n,\alpha}^\bullet) π n ∙ ( h ) = Tr ( ρ θ 0 + h / n ⊗ n T n , α ∙ ) . For every compact H ⊂ R d H\subset\mathbb{R}^d H ⊂ R d ,
sup h ∈ H ∣ π n ∙ ( h ) − β α ∙ ( c 0 T h ; V 0 ) ∣ → 0. (44) \sup_{h\in H}\left|\pi_n^\bullet(h)-\beta_\alpha^\bullet(c_0^{T}h;V_0)\right|\to0. \tag*{(44)} h ∈ H sup π n ∙ ( h ) − β α ∙ ( c 0 T h ; V 0 ) → 0. (44) In particular, the tests have local uniform asymptotic level at most α \alpha α for their respective actual nonlinear nulls:
lim sup n sup h ∈ H : ψ ( θ 0 + h / n ) ≤ 0 π n + ( h ) ≤ α , \limsup_{n}\ \sup_{h\in H:\psi(\theta_0+h/\sqrt{n})\le0}\pi_n^+(h)\le\alpha, n lim sup h ∈ H : ψ ( θ 0 + h / n ) ≤ 0 sup π n + ( h ) ≤ α , lim sup n sup h ∈ H : ψ ( θ 0 + h / n ) = 0 π n ± ( h ) ≤ α . (45) \limsup_{n}\ \sup_{h\in H:\psi(\theta_0+h/\sqrt{n})=0}\pi_n^\pm(h)\le\alpha. \tag*{(45)} n lim sup h ∈ H : ψ ( θ 0 + h / n ) = 0 sup π n ± ( h ) ≤ α . (45) An empty supremum in these level statements is taken to be zero. At h = 0 h=0 h = 0 , both rejection probabilities converge to α \alpha α .
The proofs are given in Appendix § 17 . The limiting powers in (44) coincide with those of the scalar SLD tests in § 4 , with variance V 0 = c 0 T G θ 0 − 1 c 0 V_0=c_0^T G_{\theta_0}^{-1}c_0 V 0 = c 0 T G θ 0 − 1 c 0 . The guarantees established here are local uniform asymptotic level and convergence to these power functions. Since β α ± ( t ; V 0 ) ≥ α \beta_\alpha^\pm(t;V_0)\ge\alpha β α ± ( t ; V 0 ) ≥ α , with equality at t = 0 t=0 t = 0 , the two-sided tests are also locally asymptotically unbiased for the tangent hypotheses c 0 T h = 0 c_0^T h=0 c 0 T h = 0 versus c 0 T h ≠ 0 c_0^T h\ne0 c 0 T h = 0 .
Sequential adaptive estimation
Fujiwara’s sequential adaptive maximum-likelihood method [11 , 12 ] provides another route to adaptive Wald testing. Starting from an initial estimate θ ^ 0 ∈ K \hat{\theta}_0 \in K θ ^ 0 ∈ K , at step k k k one measures M θ ^ k − 1 M_{\hat{\theta}_{k-1}} M θ ^ k − 1 , records the outcome x k x_k x k , and updates
θ ^ k ∈ argmax ϑ ∈ K ℓ k ( ϑ ) , \hat{\theta}_k \in\underset{\vartheta\in K}{\operatorname{argmax}}\,\ell_k(\vartheta), θ ^ k ∈ ϑ ∈ K argmax ℓ k ( ϑ ) ,
ℓ k ( ϑ ) : = ∑ j = 1 k log Tr { ρ ϑ M θ ^ j − 1 ( { x j } ) } . \ell_k(\vartheta) := \sum_{j=1}^{k}\log\operatorname{Tr}\left\{\rho_\vartheta M_{\hat{\theta}_{j-1}}(\{x_j\})\right\}. ℓ k ( ϑ ) := j = 1 ∑ k log Tr { ρ ϑ M θ ^ j − 1 ({ x j }) } .
The realized measurement settings are held fixed when maximizing the likelihood. Both the estimate and the measurement are thus updated throughout the experiment. Our numerical experience suggests that this sequential scheme may approach its limiting performance at smaller sample sizes than a two-stage procedure.
Let A θ A_\theta A θ denote the classical Fisher information of P ϑ ; θ P_{\vartheta;\theta} P ϑ ; θ in ϑ \vartheta ϑ , evaluated at ϑ = θ \vartheta=\theta ϑ = θ . For the scalar influence observable used here,
A θ v θ = c θ . A_\theta v_\theta=c_\theta. A θ v θ = c θ .
Indeed, the SLD equations imply that the score in direction v θ v_\theta v θ equals the observed eigenvalue of Y θ Y_\theta Y θ almost surely; multiplying by the score vector and taking expectations gives the identity. Consequently, whenever A θ ≻ 0 A_\theta\succ0 A θ ≻ 0 ,
c θ T A θ − 1 c θ = c θ T G θ − 1 c θ = V e f f ( θ ) . c_\theta^{\mathsf T}A_\theta^{-1}c_\theta=c_\theta^{\mathsf T}G_\theta^{-1}c_\theta=V_{\mathrm{eff}}(\theta). c θ T A θ − 1 c θ = c θ T G θ − 1 c θ = V eff ( θ ) .
Under Fujiwara’s positivity, identifiability, smoothness, and design-selection hypotheses, the adaptive MLE is strongly consistent and asymptotically normal with covariance A θ − 1 A_\theta^{-1} A θ − 1 [11 Theorems 8–9] . At a regular null point θ 0 ∈ int K \theta_0\in\operatorname{int}K θ 0 ∈ int K , with ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 and c θ 0 ≠ 0 c_{\theta_0}\ne0 c θ 0 = 0 , suppose additionally that
n ( θ ^ n − θ n ) ⇒ N d ( 0 , A θ 0 − 1 ) under θ n = θ 0 + h n n \sqrt{n}(\hat{\theta}_n-\theta_n)\Rightarrow N_d(0,A_{\theta_0}^{-1})\quad\text{under }\theta_n=\theta_0+\frac{h_n}{\sqrt{n}} n ( θ ^ n − θ n ) ⇒ N d ( 0 , A θ 0 − 1 ) under θ n = θ 0 + n h n
for every h n → h h_n\to h h n → h . The delta method then gives the same local powers as in (43) for the Wald statistic
W n s e q : = n ψ ( θ ^ n ) max { V e f f ( θ ^ n ) , n − 1 } . W_n^{\mathrm{seq}}:=\frac{\sqrt{n}\,\psi(\hat{\theta}_n)}{\sqrt{\max\{V_{\mathrm{eff}}(\hat{\theta}_n),n^{-1}\}}}. W n seq := max { V eff ( θ ^ n ) , n − 1 } n ψ ( θ ^ n ) .
This local regularity is an additional requirement beyond fixed-parameter asymptotic normality. The constant-rank assumption can be dropped for this sequential approach whenever a scalar-optimal measurement can be chosen continuously and the induced classical model satisfies the likelihood and local regularity conditions above.
If A θ A_\theta A θ is singular, one can use the fixed informationally complete POVM of Proposition 5.1 with a fixed probability ϵ ∈ ( 0 , 1 ) \epsilon\in(0,1) ϵ ∈ ( 0 , 1 ) , and M θ ^ k − 1 M_{\hat{\theta}_{k-1}} M θ ^ k − 1 otherwise, recording which measurement was used and updating the likelihood accordingly. The information at the true design is then
A ϵ , θ = ( 1 − ϵ ) A θ + ϵ I θ p i l ≻ 0. A_{\epsilon,\theta}=(1-\epsilon)A_\theta+\epsilon I_\theta^{\mathrm{pil}}\succ0. A ϵ , θ = ( 1 − ϵ ) A θ + ϵ I θ pil ≻ 0.
Under the corresponding likelihood and design regularity conditions, the scalar asymptotic variance is V ϵ ( θ ) : = c θ T A ϵ , θ − 1 c θ V_{\epsilon}(\theta) := c_{\theta}^{T}A_{\epsilon,\theta}^{-1}c_{\theta} V ϵ ( θ ) := c θ T A ϵ , θ − 1 c θ . The information inequalities A θ , I θ p i l ⪯ G θ A_{\theta}, I_{\theta}^{\mathrm{pil}} \preceq G_{\theta} A θ , I θ pil ⪯ G θ and A θ v θ = c θ A_{\theta}v_{\theta}=c_{\theta} A θ v θ = c θ give
V e f f ( θ ) ≤ V ϵ ( θ ) ≤ V e f f ( θ ) 1 − ϵ . V_{\mathrm{eff}}(\theta) \le V_{\epsilon}(\theta) \le\frac{V_{\mathrm{eff}}(\theta)}{1-\epsilon}. V eff ( θ ) ≤ V ϵ ( θ ) ≤ 1 − ϵ V eff ( θ ) .
Thus a sufficiently small fixed ϵ \epsilon ϵ gives an arbitrarily small asymptotic variance loss; the corresponding Wald statistic uses V ϵ V_{\epsilon} V ϵ for studentization. We use the two-stage construction for its direct proof of the local limit, while sequential adaptation remains of interest for its performance at moderate sample sizes.
Application to spectral functional testing
In this section, we apply the preceding results to testing smooth spectral functionals. We compare the local asymptotic powers of scalar SLD tests and tests based on weak Schur sampling (WSS), distinguishing the effects of classical post-processing and measurement choice. The comparison identifies conditions under which the WSS procedures attain the same limiting power as the scalar SLD test, which achieves the local maximin bound under the assumptions of § 4 .
Let S = { ρ θ : θ ∈ Θ ⊂ R d } \mathcal{S}=\{\rho_{\theta}:\theta\in\Theta\subset\mathbb{R}^{d}\} S = { ρ θ : θ ∈ Θ ⊂ R d } be a C 1 C^{1} C 1 parametric family of full-rank states on C r \mathbb{C}^{r} C r , where Θ \Theta Θ is open. Fix θ 0 ∈ Θ \theta_{0}\in\Theta θ 0 ∈ Θ and write
ρ 0 : = ρ θ 0 = U 0 diag ( α 0 ) U 0 ∗ , α 0 , 1 > ⋯ > α 0 , r > 0. (46) \rho_{0}:=\rho_{\theta_{0}}=U_{0}\operatorname{diag}(\alpha_{0})U_{0}^{*},\qquad\alpha_{0,1}>\cdots>\alpha_{0,r}>0. \tag*{(46)} ρ 0 := ρ θ 0 = U 0 diag ( α 0 ) U 0 ∗ , α 0 , 1 > ⋯ > α 0 , r > 0. (46)
Denote the decreasingly ordered eigenvalue vector of ρ θ \rho_{\theta} ρ θ by α θ = ( α θ , 1 , … , α θ , r ) T ∈ Δ r ∘ \alpha_{\theta}=(\alpha_{\theta,1},\ldots,\alpha_{\theta,r})^{T}\in\Delta_{r}^{\circ} α θ = ( α θ , 1 , … , α θ , r ) T ∈ Δ r ∘ , where
Δ r : = { x ∈ [ 0 , 1 ] r : ∑ i = 1 r x i = 1 } , Δ r ∘ : = { x ∈ ( 0 , 1 ) r : ∑ i = 1 r x i = 1 } . \Delta_{r}:=\left\{x\in[0,1]^{r}:\sum_{i=1}^{r}x_{i}=1\right\},\qquad\Delta_{r}^{\circ}:=\left\{x\in(0,1)^{r}:\sum_{i=1}^{r}x_{i}=1\right\}. Δ r := { x ∈ [ 0 , 1 ] r : i = 1 ∑ r x i = 1 } , Δ r ∘ := { x ∈ ( 0 , 1 ) r : i = 1 ∑ r x i = 1 } .
Assume that θ ↦ α θ \theta\mapsto\alpha_{\theta} θ ↦ α θ is differentiable at a fixed point θ 0 ∈ Θ \theta_{0}\in\Theta θ 0 ∈ Θ .
Let ψ \psi ψ be a spectral functional. Thus there exists a symmetric function
φ : Δ r → R \varphi:\Delta_{r}\to\mathbb{R} φ : Δ r → R
such that ψ ( θ ) = φ ( α θ ) \psi(\theta)=\varphi(\alpha_{\theta}) ψ ( θ ) = φ ( α θ ) . Assume that φ \varphi φ is continuously differentiable in a neighborhood of α 0 : = α θ 0 \alpha_{0}:=\alpha_{\theta_{0}} α 0 := α θ 0 , and that ψ ( θ 0 ) = 0 \psi(\theta_{0})=0 ψ ( θ 0 ) = 0 , ψ ˙ θ 0 ≠ 0 \dot{\psi}_{\theta_{0}}\ne0 ψ ˙ θ 0 = 0 .
We consider the one-sided hypotheses
H 0 : ψ ( θ ) ≤ 0 , H 1 : ψ ( θ ) > 0. H_{0}:\psi(\theta)\le0,\qquad H_{1}:\psi(\theta)>0. H 0 : ψ ( θ ) ≤ 0 , H 1 : ψ ( θ ) > 0.
Next define
A 0 : = D θ α θ 0 ∈ R r × d , b 0 : = ∇ φ ( α 0 ) ∈ R r , A_{0}:=D_{\theta}\alpha_{\theta_{0}}\in\mathbb{R}^{r\times d},\qquad b_{0}:=\nabla\varphi(\alpha_{0})\in\mathbb{R}^{r}, A 0 := D θ α θ 0 ∈ R r × d , b 0 := ∇ φ ( α 0 ) ∈ R r ,
where gradients are represented as column vectors. By the chain rule,
c 0 : = ψ ˙ θ 0 T = A 0 T b 0 ∈ R d . c_{0}:=\dot{\psi}_{\theta_{0}}^{T}=A_{0}^{T}b_{0}\in\mathbb{R}^{d}. c 0 := ψ ˙ θ 0 T = A 0 T b 0 ∈ R d .
Let
C 0 : = diag ( α 0 ) − α 0 α 0 T . C_{0}:=\operatorname{diag}(\alpha_{0})-\alpha_{0}\alpha_{0}^{T}. C 0 := diag ( α 0 ) − α 0 α 0 T .
For b 0 b_{0} b 0 , choose any local C 1 C^{1} C 1 extension of φ \varphi φ from the simplex to a neighborhood in R r \mathbb{R}^{r} R r . Two such extensions have gradients differing by a multiple of 1 1 1 . Since 1 T A 0 = 0 1^{T}A_{0}=0 1 T A 0 = 0 and C 0 1 = 0 C_{0}1=0 C 0 1 = 0 , this choice does not affect c 0 c_{0} c 0 , C 0 b 0 C_{0}b_{0} C 0 b 0 , or the variance formulas below. We next introduce two tests: one based on the scalar SLD statistic from § 4 , and the other based on weak Schur sampling.
Tests based on SLD statistics and weak Schur sampling
Test based on the SLD statistic
Let L 0 , j : = L θ 0 , j L_{0,j}:=L_{\theta_0,j} L 0 , j := L θ 0 , j , j = 1 , … , d j=1,\ldots,d j = 1 , … , d , be the SLDs at θ 0 \theta_0 θ 0 , and put
J 0 : = J θ 0 , G 0 : = Re J 0 . J_0:=J_{\theta_0},\qquad G_0:=\operatorname{Re}J_0. J 0 := J θ 0 , G 0 := Re J 0 .
Assume that G 0 ≻ 0 G_0\succ0 G 0 ≻ 0 . Recall from § 4 , that
Δ n , j : = 1 n ∑ a = 1 n L 0 , j [ a ] , X ( n ) : = G 0 − 1 Δ n , \Delta_{n,j}:=\frac{1}{\sqrt{n}}\sum_{a=1}^{n}L_{0,j}^{[a]},\qquad X^{(n)}:=G_0^{-1}\Delta_n, Δ n , j := n 1 a = 1 ∑ n L 0 , j [ a ] , X ( n ) := G 0 − 1 Δ n ,
and
Σ θ 0 : = G 0 − 1 J 0 G 0 − 1 . \Sigma_{\theta_0}:=G_0^{-1}J_0G_0^{-1}. Σ θ 0 := G 0 − 1 J 0 G 0 − 1 .
Put v 0 : = G 0 − 1 c 0 v_0:=G_0^{-1}c_0 v 0 := G 0 − 1 c 0 and define
Y 0 : = ∑ j = 1 d ( v 0 ) j L 0 , j . Y_0:=\sum_{j=1}^{d}(v_0)_jL_{0,j}. Y 0 := j = 1 ∑ d ( v 0 ) j L 0 , j .
Then
c 0 T X ( n ) = 1 n ∑ a = 1 n Y 0 [ a ] . c_0^{T}X^{(n)}=\frac{1}{\sqrt{n}}\sum_{a=1}^{n}Y_0^{[a]}. c 0 T X ( n ) = n 1 a = 1 ∑ n Y 0 [ a ] .
Thus the scalar statistic can be obtained by measuring Y 0 Y_0 Y 0 separately on each copy and summing the outcomes. This is the observable Y θ 0 Y_{\theta_0} Y θ 0 defined in § 5 . The scalar SLD variance for ψ \psi ψ is
V e f f : = ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T = c 0 T Σ θ 0 c 0 . (47) V_{\mathrm{eff}}:=\dot{\psi}_{\theta_0}\Sigma_{\theta_0}\dot{\psi}_{\theta_0}^{T}=c_0^{T}\Sigma_{\theta_0}c_0. \tag*{(47)} V eff := ψ ˙ θ 0 Σ θ 0 ψ ˙ θ 0 T = c 0 T Σ θ 0 c 0 . (47)
For the canonical SLD covariance Σ θ 0 = G 0 − 1 J 0 G 0 − 1 \Sigma_{\theta_0}=G_0^{-1}J_0G_0^{-1} Σ θ 0 = G 0 − 1 J 0 G 0 − 1 , we have
V e f f = c 0 T G 0 − 1 J 0 G 0 − 1 c 0 = c 0 T G 0 − 1 c 0 , V_{\mathrm{eff}}=c_0^{T}G_0^{-1}J_0G_0^{-1}c_0=c_0^{T}G_0^{-1}c_0, V eff = c 0 T G 0 − 1 J 0 G 0 − 1 c 0 = c 0 T G 0 − 1 c 0 ,
since Im J 0 \operatorname{Im}J_0 Im J 0 is real antisymmetric. Fix a significance level γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) , and let z γ : = Φ − 1 ( 1 − γ ) z_\gamma:=\Phi^{-1}(1-\gamma) z γ := Φ − 1 ( 1 − γ ) . The procedures below are local oracle procedures at θ 0 \theta_0 θ 0 : their defining observables, weights, and variance calibrations may depend on θ 0 \theta_0 θ 0 . Consider the local parameter sequence
θ n = θ 0 + h n , h ∈ R d . \theta_n=\theta_0+\frac{h}{\sqrt{n}},\qquad h\in\mathbb{R}^{d}. θ n = θ 0 + n h , h ∈ R d .
Define the q-LAN estimator of the functional by
ψ ^ n , 1 : = ψ ( θ 0 ) + 1 n ψ ˙ θ 0 X ( n ) = 1 n c 0 T X ( n ) . (48) \widehat{\psi}_{n,1}:=\psi(\theta_0)+\frac{1}{\sqrt{n}}\dot{\psi}_{\theta_0}X^{(n)}=\frac{1}{\sqrt{n}}c_0^{T}X^{(n)}. \tag*{(48)} ψ n , 1 := ψ ( θ 0 ) + n 1 ψ ˙ θ 0 X ( n ) = n 1 c 0 T X ( n ) . (48)
The corresponding test, calibrated to have local asymptotic level γ \gamma γ , is
T n , 1 : = { c 0 T X ( n ) > z γ V e f f } . (49) T_{n,1}:=\left\{c_0^{T}X^{(n)}>z_\gamma\sqrt{V_{\mathrm{eff}}}\right\}. \tag*{(49)} T n , 1 := { c 0 T X ( n ) > z γ V eff } . (49)
Here and below, { Y > a } \{Y>a\} { Y > a } denotes the spectral projection of the self-adjoint observable Y Y Y corresponding to the interval ( a , ∞ ) (a,\infty) ( a , ∞ ) . The test T n , 1 T_{n,1} T n , 1 in (49) is locally asymptotically optimal under the q-LAN and D \mathcal{D} D -invariance assumptions of Theorem 4.1 . Without D \mathcal{D} D -invariance, we do not assert such optimality. Nevertheless, we retain the notation V e f f V_{\mathrm{eff}} V eff in (47) for the scalar SLD variance, irrespective of whether the optimality conditions hold. We will, however, show that
V e f f ≤ V s p ≤ V W S S , V_{\mathrm{eff}} \le V_{\mathrm{sp}} \le V_{\mathrm{WSS}}, V eff ≤ V sp ≤ V WSS ,
where V s p V_{\mathrm{sp}} V sp is the minimum variance among linear statistics of the Gaussian WSS fluctuations having the correct local mean in every direction. Thus V e f f V_{\mathrm{eff}} V eff is also a lower bound for these variances, although it need not be attained within this class.
Test based on weak Schur Sampling
Under the Schur-Weyl decomposition (see Appendix § 10 for details),
( C r ) ⊗ n ≅ ⨁ λ ⊢ n ℓ ( λ ) ≤ r Q λ ⊗ P λ , (\mathbb{C}^{r})^{\otimes n} \cong\bigoplus_{\substack{\lambda\vdash n \\ \ell(\lambda) \le r}} \mathcal{Q}_{\lambda} \otimes\mathcal{P}_{\lambda}, ( C r ) ⊗ n ≅ λ ⊢ n ℓ ( λ ) ≤ r ⨁ Q λ ⊗ P λ ,
let Π λ \Pi_{\lambda} Π λ denote the orthogonal projection onto the λ \lambda λ -isotypic component. Weak Schur sampling is the projective measurement
{ Π λ : λ ⊢ n , ℓ ( λ ) ≤ r } . \{\Pi_{\lambda} : \lambda\vdash n,\ \ell(\lambda) \le r\}. { Π λ : λ ⊢ n , ℓ ( λ ) ≤ r } .
If the input state is ρ θ ⊗ n \rho_{\theta}^{\otimes n} ρ θ ⊗ n , its outcome λ ( n ) = ( λ 1 ( n ) , … , λ r ( n ) ) \lambda^{(n)} = \left(\lambda_{1}^{(n)},\ldots,\lambda_{r}^{(n)}\right) λ ( n ) = ( λ 1 ( n ) , … , λ r ( n ) ) has the Schur-Weyl distribution (see Appendix § 10 for details) denoted by:
λ ( n ) ∼ SW n ( α θ ) . \lambda^{(n)} \sim\operatorname{SW}^{n}(\alpha_{\theta}). λ ( n ) ∼ SW n ( α θ ) .
Define the empirical Young-diagram spectrum and the WSS plug-in estimator by
α ^ n : = λ ( n ) n , ψ ^ n , 2 : = φ ( α ^ n ) . \widehat{\alpha}_{n} := \frac{\lambda^{(n)}}{n}, \qquad\widehat{\psi}_{n,2} := \varphi\left(\widehat{\alpha}_{n}\right). α n := n λ ( n ) , ψ n , 2 := φ ( α n ) .
We define the weak-Schur-sampling variance by
V W S S : = b 0 T C 0 b 0 = ∑ i = 1 r α 0 , i b 0 , i 2 − ( ∑ i = 1 r α 0 , i b 0 , i ) 2 . (50) V_{\mathrm{WSS}} := b_{0}^{T}C_{0}b_{0} = \sum_{i=1}^{r}\alpha_{0,i}b_{0,i}^{2} - \left(\sum_{i=1}^{r}\alpha_{0,i}b_{0,i}\right)^{2}. \tag*{(50)} V WSS := b 0 T C 0 b 0 = i = 1 ∑ r α 0 , i b 0 , i 2 − ( i = 1 ∑ r α 0 , i b 0 , i ) 2 . (50)
The corresponding WSS test is
T n , 2 : = ∑ λ ⊢ n , ℓ ( λ ) ≤ r : n [ φ ( λ / n ) − φ ( α 0 ) ] > z γ V W S S Π λ . (51) T_{n,2} := \sum_{\substack{\lambda\vdash n,\ \ell(\lambda) \le r: \\ \sqrt{n}\,[\varphi(\lambda/n)-\varphi(\alpha_{0})] > z_{\gamma}\sqrt{V_{\mathrm{WSS}}}}} \Pi_{\lambda}. \tag*{(51)} T n , 2 := λ ⊢ n , ℓ ( λ ) ≤ r : n [ φ ( λ / n ) − φ ( α 0 )] > z γ V WSS ∑ Π λ . (51)
Equivalently, after observing λ ( n ) \lambda^{(n)} λ ( n ) , the test rejects H 0 H_{0} H 0 precisely when
n [ ψ ^ n , 2 − ψ ( θ 0 ) ] V W S S > z γ . \frac{\sqrt{n}\,[\widehat{\psi}_{n,2}-\psi(\theta_{0})]}{\sqrt{V_{\mathrm{WSS}}}} > z_{\gamma}. V WSS n [ ψ n , 2 − ψ ( θ 0 )] > z γ .
Weak Schur sampling under local alternatives
Lemma 6.1 (Local WSS central limit theorem and delta method). Suppose (46) holds. Let α n ∈ Δ r ∘ \alpha_n \in\Delta_r^\circ α n ∈ Δ r ∘ satisfy
n ( α n − α 0 ) → a , 1 T a = 0 , \sqrt{n}(\alpha_n-\alpha_0)\to a,\qquad\mathbf{1}^{T}a=0, n ( α n − α 0 ) → a , 1 T a = 0 , and let λ ( n ) ∼ S W n ( α n ) \lambda^{(n)}\sim\mathrm{SW}^{n}(\alpha_n) λ ( n ) ∼ SW n ( α n ) . Then
n ( α ^ n − α n ) ⇒ N r ( 0 , C 0 ) , (52) \sqrt{n}(\widehat{\alpha}_n-\alpha_n)\Rightarrow N_r(0,C_0), \tag*{(52)} n ( α n − α n ) ⇒ N r ( 0 , C 0 ) , (52) n ( α ^ n − α 0 ) ⇒ N r ( a , C 0 ) , (53) \sqrt{n}(\widehat{\alpha}_n-\alpha_0)\Rightarrow N_r(a,C_0), \tag*{(53)} n ( α n − α 0 ) ⇒ N r ( a , C 0 ) , (53) n { φ ( α ^ n ) − φ ( α n ) } ⇒ N ( 0 , V W S S ) . (54) \sqrt{n}\{\varphi(\widehat{\alpha}_n)-\varphi(\alpha_n)\}\Rightarrow N(0,V_{\mathrm{WSS}}). \tag*{(54)} n { φ ( α n ) − φ ( α n )} ⇒ N ( 0 , V WSS ) . (54) The vector Gaussian laws are supported on { x ∈ R r : 1 T x = 0 } \{x\in\mathbb{R}^r:\mathbf{1}^{T}x=0\} { x ∈ R r : 1 T x = 0 } .
Theorem 6.2 (Local comparison of the q-LAN and WSS tests). Assume that the model is q-LAN and D-extendible at θ 0 \theta_0 θ 0 . Then, under ρ θ 0 + h / n ⊗ n \rho_{\theta_0+h/\sqrt{n}}^{\otimes n} ρ θ 0 + h / n ⊗ n , the following assertions hold.
(i) The q-LAN estimator satisfies
n [ ψ ^ n , 1 − ψ ( θ 0 + h n ) ] ⇒ N ( 0 , V e f f ) . (55) \sqrt{n}\left[\widehat{\psi}_{n,1}-\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)\right]\Rightarrow N(0,V_{\mathrm{eff}}). \tag*{(55)} n [ ψ n , 1 − ψ ( θ 0 + n h ) ] ⇒ N ( 0 , V eff ) . (55) (ii) The WSS estimator satisfies
n [ ψ ^ n , 2 − ψ ( θ 0 + h n ) ] ⇒ N ( 0 , V W S S ) . (56) \sqrt{n}\left[\widehat{\psi}_{n,2}-\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)\right]\Rightarrow N(0,V_{\mathrm{WSS}}). \tag*{(56)} n [ ψ n , 2 − ψ ( θ 0 + n h ) ] ⇒ N ( 0 , V WSS ) . (56) (iii) Let κ ( h ) : = ψ ˙ θ 0 h = c 0 T h \kappa(h):=\dot{\psi}_{\theta_0}h=c_0^{T}h κ ( h ) := ψ ˙ θ 0 h = c 0 T h , then the limiting powers of the two tests T n , 1 T_{n,1} T n , 1 and T n , 2 T_{n,2} T n , 2 are
lim n → ∞ Tr [ ρ θ 0 + h / n ⊗ n T n , 1 ] = 1 − Φ ( z γ − κ ( h ) V e f f ) , (57) \lim_{n\to\infty}\operatorname{Tr}\left[\rho_{\theta_0+h/\sqrt{n}}^{\otimes n}T_{n,1}\right]
=1-\Phi\left(z_\gamma-\frac{\kappa(h)}{\sqrt{V_{\mathrm{eff}}}}\right), \tag*{(57)} n → ∞ lim Tr [ ρ θ 0 + h / n ⊗ n T n , 1 ] = 1 − Φ ( z γ − V eff κ ( h ) ) , (57) lim n → ∞ Tr [ ρ θ 0 + h / n ⊗ n T n , 2 ] = 1 − Φ ( z γ − κ ( h ) V W S S ) . (58) \lim_{n\to\infty}\operatorname{Tr}\left[\rho_{\theta_0+h/\sqrt{n}}^{\otimes n}T_{n,2}\right]
=1-\Phi\left(z_\gamma-\frac{\kappa(h)}{\sqrt{V_{\mathrm{WSS}}}}\right). \tag*{(58)} n → ∞ lim Tr [ ρ θ 0 + h / n ⊗ n T n , 2 ] = 1 − Φ ( z γ − V WSS κ ( h ) ) . (58) (iv) Along every one-dimensional submodel
t ⟼ θ 0 + t h t\longmapsto\theta_0+th t ⟼ θ 0 + t h for which κ ( h ) > 0 \kappa(h)>0 κ ( h ) > 0 , the Pitman asymptotic relative efficiency of T n , 1 T_{n,1} T n , 1 with respect to T n , 2 T_{n,2} T n , 2 is
ARE P ( T 1 , T 2 ) = V W S S V e f f . (59) \operatorname{ARE}_{\mathrm{P}}(T_1,T_2)=\frac{V_{\mathrm{WSS}}}{V_{\mathrm{eff}}}. \tag*{(59)} ARE P ( T 1 , T 2 ) = V eff V WSS . (59)
The variance V W S S V_{\mathrm{WSS}} V WSS is the limiting variance of the unrestricted WSS plug-in estimator, whereas V e f f V_{\mathrm{eff}} V eff is that of the scalar SLD estimator. A measurement and a decision rule based on its outcome are different statistical objects. We therefore introduce V s p V_{\mathrm{sp}} V sp : the minimum variance of a linear statistic in the Gaussian WSS limit whose mean equals c 0 T h c_0^{T}h c 0 T h for every local parameter h h h .
Put
Z n : = n ( α ^ n − α 0 ) , H : = { z ∈ R r : 1 T z = 0 } . Z_n:=\sqrt{n}(\widehat{\alpha}_n-\alpha_0),\qquad H:=\{z\in\mathbb{R}^r:\mathbf{1}^{T}z=0\}. Z n := n ( α n − α 0 ) , H := { z ∈ R r : 1 T z = 0 } .
Under θ n = θ 0 + h / n \theta_n = \theta_0 + h/\sqrt{n} θ n = θ 0 + h / n , Lemma 6.1 gives
Z n ⇒ Z ∼ N r ( A 0 h , C 0 ) . Z_n \Rightarrow Z \sim N_r(A_0h,C_0). Z n ⇒ Z ∼ N r ( A 0 h , C 0 ) .
The first-order fluctuation of the unrestricted plug-in is
n { φ ( α ^ n ) − φ ( α 0 ) } = b 0 T Z n + o P ( 1 ) . \sqrt{n}\{\varphi(\widehat{\alpha}_n)-\varphi(\alpha_0)\}=b_0^T Z_n+o_P(1). n { φ ( α n ) − φ ( α 0 )} = b 0 T Z n + o P ( 1 ) .
Its limiting mean is c 0 T h c_0^T h c 0 T h , because c 0 = A 0 T b 0 c_0=A_0^T b_0 c 0 = A 0 T b 0 , and its limiting variance is b 0 T C 0 b 0 = V W S S b_0^T C_0b_0=V_{\mathrm{WSS}} b 0 T C 0 b 0 = V WSS . We now ask whether a different linear statistic computed from the same WSS outcome can have the same limiting mean c 0 T h c_0^T h c 0 T h for every h h h , but smaller limiting variance. Since the limit Z ∼ N r ( A 0 h , C 0 ) Z\sim N_r(A_0h,C_0) Z ∼ N r ( A 0 h , C 0 ) , for a linear statistic w T Z w^T Z w T Z we have,
E h [ w T Z ] = w T A 0 h , Var h ( w T Z ) = w T C 0 w . \mathbb{E}_h[w^T Z]=w^T A_0h,\qquad\operatorname{Var}_h(w^T Z)=w^T C_0w. E h [ w T Z ] = w T A 0 h , Var h ( w T Z ) = w T C 0 w .
Thus w T Z w^T Z w T Z has mean c 0 T h c_0^T h c 0 T h for every h h h precisely when
A 0 T w = c 0 . (60) A_0^T w=c_0. \tag*{(60)} A 0 T w = c 0 . (60)
Among all such w w w , we therefore seek one minimizing the limiting variance w T C 0 w w^T C_0w w T C 0 w . Adding a multiple of 1 1 1 to w w w changes neither the statistic on H H H nor its variance, so we choose the representative satisfying α 0 T w = 0 \alpha_0^T w=0 α 0 T w = 0 . For this representative,
w T C 0 w = w T diag ( α 0 ) w , w^T C_0w=w^T\operatorname{diag}(\alpha_0)w, w T C 0 w = w T diag ( α 0 ) w ,
giving the optimization problem
min w ∈ R r { w T diag ( α 0 ) w : A 0 T w = c 0 , α 0 T w = 0 } . (61) \min_{w\in\mathbb{R}^r}\left\{w^T\operatorname{diag}(\alpha_0)w:A_0^T w=c_0,\ \alpha_0^T w=0\right\}. \tag*{(61)} w ∈ R r min { w T diag ( α 0 ) w : A 0 T w = c 0 , α 0 T w = 0 } . (61)
Put I s p , 0 : = A 0 T diag ( α 0 ) − 1 A 0 I_{\mathrm{sp},0}:=A_0^T\operatorname{diag}(\alpha_0)^{-1}A_0 I sp , 0 := A 0 T diag ( α 0 ) − 1 A 0 and I s p , 0 + I_{\mathrm{sp},0}^{+} I sp , 0 + denote its Moore–Penrose inverse.
Remark 3. We record the observation that v T I s p , 0 v = ( A 0 v ) T diag ( α 0 ) − 1 ( A 0 v ) v^T I_{\mathrm{sp},0}v=(A_0v)^T\operatorname{diag}(\alpha_0)^{-1}(A_0v) v T I sp , 0 v = ( A 0 v ) T diag ( α 0 ) − 1 ( A 0 v ) for all v ∈ R d v\in\mathbb{R}^d v ∈ R d , and hence we have ker ( I s p , 0 ) = ker ( A 0 ) \operatorname{ker}(I_{\mathrm{sp},0})=\operatorname{ker}(A_0) ker ( I sp , 0 ) = ker ( A 0 ) . Also, as I s p , 0 I_{\mathrm{sp},0} I sp , 0 is symmetric, Ran ( I s p , 0 ) = ker ( I s p , 0 ) ⊥ = ker ( A 0 ) ⊥ = Ran ( A 0 T ) \operatorname{Ran}(I_{\mathrm{sp},0})=\operatorname{ker}(I_{\mathrm{sp},0})^\perp=\operatorname{ker}(A_0)^\perp=\operatorname{Ran}(A_0^T) Ran ( I sp , 0 ) = ker ( I sp , 0 ) ⊥ = ker ( A 0 ) ⊥ = Ran ( A 0 T ) . In particular, c 0 = A 0 T b 0 c_0=A_0^T b_0 c 0 = A 0 T b 0 belongs to Ran ( I s p , 0 ) \operatorname{Ran}(I_{\mathrm{sp},0}) Ran ( I sp , 0 ) , so I s p , 0 I s p , 0 + c 0 = c 0 I_{\mathrm{sp},0}I_{\mathrm{sp},0}^{+}c_0=c_0 I sp , 0 I sp , 0 + c 0 = c 0 .
The unique minimizer of (61) is
w 0 : = diag ( α 0 ) − 1 A 0 I s p , 0 + c 0 , w_0:=\operatorname{diag}(\alpha_0)^{-1}A_0I_{\mathrm{sp},0}^{+}c_0, w 0 := diag ( α 0 ) − 1 A 0 I sp , 0 + c 0 ,
and its minimum value is
V s p : = c 0 T I s p , 0 + c 0 . V_{\mathrm{sp}}:=c_0^T I_{\mathrm{sp},0}^{+}c_0. V sp := c 0 T I sp , 0 + c 0 .
When A 0 A_0 A 0 has full column rank, the Moore–Penrose inverse is the ordinary inverse. The corresponding linearly post-processed WSS test rejects when w 0 T Z n > z γ V s p w_0^T Z_n>z_\gamma\sqrt{V_{\mathrm{sp}}} w 0 T Z n > z γ V sp . By Lemma 6.1 ,
w 0 T Z n ⇒ N ( c 0 T h , V s p ) w_0^T Z_n\Rightarrow N(c_0^T h,V_{\mathrm{sp}}) w 0 T Z n ⇒ N ( c 0 T h , V sp )
under θ n = θ 0 + h / n \theta_n=\theta_0+h/\sqrt{n} θ n = θ 0 + h / n . Its limiting power is therefore
1 − Φ ( z γ − c 0 T h V s p ) . 1-\Phi\left(z_\gamma-\frac{c_0^T h}{\sqrt{V_{\mathrm{sp}}}}\right). 1 − Φ ( z γ − V sp c 0 T h ) .
For c 0 T h > 0 c_0^T h>0 c 0 T h > 0 , this power is maximal among the corresponding upper-tail tests based on linear statistics w T Z n w^T Z_n w T Z n satisfying A 0 T w = c 0 A_0^T w=c_0 A 0 T w = c 0 .
Corollary 6.3 (Comparison of the three asymptotic variances). Let Θ ⊂ R d \Theta\subset\mathbb{R}^{d} Θ ⊂ R d be open, and suppose that θ ↦ ρ θ \theta\mapsto\rho_{\theta} θ ↦ ρ θ is a C 1 C^{1} C 1 family of strictly positive density matrices on C r \mathbb{C}^{r} C r in a neighborhood of θ 0 \theta_{0} θ 0 . Assume that ρ θ 0 \rho_{\theta_{0}} ρ θ 0 has simple spectrum, that φ \varphi φ is symmetric and C 1 C^{1} C 1 in a neighborhood of α 0 \alpha_{0} α 0 , and that
c 0 = A 0 T b 0 ≠ 0 , Re J θ 0 ≻ 0. c_{0}=A_{0}^{T}b_{0}\ne0,\qquad\operatorname{Re}J_{\theta_{0}}\succ0. c 0 = A 0 T b 0 = 0 , Re J θ 0 ≻ 0. For V e f f V_{\mathrm{eff}} V eff as defined in (47) , take the canonical SLD covariance
Σ θ 0 = ( Re J θ 0 ) − 1 J θ 0 ( Re J θ 0 ) − 1 . \Sigma_{\theta_{0}}=(\operatorname{Re}J_{\theta_{0}})^{-1}J_{\theta_{0}}(\operatorname{Re}J_{\theta_{0}})^{-1}. Σ θ 0 = ( Re J θ 0 ) − 1 J θ 0 ( Re J θ 0 ) − 1 . Then the three variances satisfy
0 < V e f f ≤ V s p ≤ V W S S . (62) 0<V_{\mathrm{eff}}\le V_{\mathrm{sp}}\le V_{\mathrm{WSS}}. \tag*{(62)} 0 < V eff ≤ V sp ≤ V WSS . (62) Moreover, the following equality criteria hold:
V W S S = V s p ⟺ C 0 b 0 ∈ Ran ( A 0 ) , (63) V_{\mathrm{WSS}}=V_{\mathrm{sp}}\quad\Longleftrightarrow\quad C_{0}b_{0}\in\operatorname{Ran}(A_{0}), \tag*{(63)} V WSS = V sp ⟺ C 0 b 0 ∈ Ran ( A 0 ) , (63) V s p = V e f f ⟺ [ Y θ 0 , ρ θ 0 ] = 0 , (64) V_{\mathrm{sp}}=V_{\mathrm{eff}}\quad\Longleftrightarrow\quad[Y_{\theta_{0}},\rho_{\theta_{0}}]=0, \tag*{(64)} V sp = V eff ⟺ [ Y θ 0 , ρ θ 0 ] = 0 , (64) where Y θ 0 Y_{\theta_{0}} Y θ 0 is the scalar SLD observable defined in § 5 . When A 0 A_{0} A 0 has full column rank, the second criterion is equivalently
( Re J θ 0 − I s p , 0 ) I s p , 0 − 1 c 0 = 0. (\operatorname{Re}J_{\theta_{0}}-I_{\mathrm{sp},0})I_{\mathrm{sp},0}^{-1}c_{0}=0. ( Re J θ 0 − I sp , 0 ) I sp , 0 − 1 c 0 = 0. If the SLD space at θ 0 \theta_{0} θ 0 is D ρ 0 D_{\rho_{0}} D ρ 0 -invariant, then V s p = V e f f V_{\mathrm{sp}}=V_{\mathrm{eff}} V sp = V eff , but equality with V W S S V_{\mathrm{WSS}} V WSS additionally requires C 0 b 0 ∈ Ran ( A 0 ) C_{0}b_{0}\in\operatorname{Ran}(A_{0}) C 0 b 0 ∈ Ran ( A 0 ) . In particular, if the SLD space is D ρ 0 D_{\rho_{0}} D ρ 0 -invariant and rank ( A 0 ) = r − 1 \operatorname{rank}(A_{0})=r-1 rank ( A 0 ) = r − 1 , then
V W S S = V s p = V e f f . V_{\mathrm{WSS}}=V_{\mathrm{sp}}=V_{\mathrm{eff}}. V WSS = V sp = V eff . The Pitman asymptotic relative efficiency of the corresponding scalar SLD and WSS plug-in tests is
ARE P ( T 1 , T 2 ) = V W S S V e f f = V W S S V s p V s p V e f f ≥ 1 , \operatorname{ARE}_{\mathrm{P}}(T_{1},T_{2})=\frac{V_{\mathrm{WSS}}}{V_{\mathrm{eff}}}=\frac{V_{\mathrm{WSS}}}{V_{\mathrm{sp}}}\frac{V_{\mathrm{sp}}}{V_{\mathrm{eff}}}\ge1, ARE P ( T 1 , T 2 ) = V eff V WSS = V sp V WSS V eff V sp ≥ 1 , with equality if and only if both equality conditions hold.
Remark 4. The optimization defining V s p V_{\mathrm{sp}} V sp concerns linear statistics of the Gaussian WSS limit having mean c 0 T h c_{0}^{T}h c 0 T h for every h h h . It also applies to estimators with an expansion
n { ψ ^ n − ψ ( θ 0 ) } = w T Z n + o P ( 1 ) , A 0 T w = c 0 , \sqrt{n}\{\widehat{\psi}_{n}-\psi(\theta_{0})\}=w^{T}Z_{n}+o_{\mathrm{P}}(1),\qquad A_{0}^{T}w=c_{0}, n { ψ n − ψ ( θ 0 )} = w T Z n + o P ( 1 ) , A 0 T w = c 0 ,
under every fixed local alternative. Such an estimator need not be linear in the finite-sample WSS outcome, as the next example illustrates.
Two explicit examples
Example 6.4 (A removable post-processing loss). For ∣ θ ∣ < 1 / 20 |\theta|<1/20 ∣ θ ∣ < 1/20 , let
ρ θ = diag ( 1 2 + θ , 3 10 − θ , 1 5 ) , ψ ( θ ) = Tr ( ρ θ 2 ) − 19 50 = 2 5 θ + 2 θ 2 . \rho_{\theta}=\operatorname{diag}\left(\frac{1}{2}+\theta,\frac{3}{10}-\theta,\frac{1}{5}\right),\qquad\psi(\theta)=\operatorname{Tr}(\rho_{\theta}^{2})-\frac{19}{50}=\frac{2}{5}\theta+2\theta^{2}. ρ θ = diag ( 2 1 + θ , 10 3 − θ , 5 1 ) , ψ ( θ ) = Tr ( ρ θ 2 ) − 50 19 = 5 2 θ + 2 θ 2 .
At θ 0 = 0 \theta_0 = 0 θ 0 = 0 , we have
α 0 = ( 1 / 2 3 / 10 1 / 5 ) , A 0 = d α θ d θ ∣ θ = 0 = ( 1 − 1 0 ) , \boldsymbol{\alpha}_0 = \begin{pmatrix} 1/2 \\ 3/10 \\ 1/5 \end{pmatrix}, \qquad A_0 = \left.\frac{\mathrm{d}\boldsymbol{\alpha}_\theta}{\mathrm{d}\theta}\right|_{\theta=0} = \begin{pmatrix} 1 \\ -1 \\ 0 \end{pmatrix}, α 0 = 1/2 3/10 1/5 , A 0 = d θ d α θ θ = 0 = 1 − 1 0 ,
b 0 = 2 α 0 = ( 1 3 / 5 2 / 5 ) , c 0 = A 0 T b 0 = 2 5 . b_0 = 2\boldsymbol{\alpha}_0 = \begin{pmatrix} 1 \\ 3/5 \\ 2/5 \end{pmatrix}, \qquad c_0 = A_0^{T}b_0 = \frac{2}{5}. b 0 = 2 α 0 = 1 3/5 2/5 , c 0 = A 0 T b 0 = 5 2 .
Writing D 0 = diag ( α 0 ) D_0 = \operatorname{diag}(\boldsymbol{\alpha}_0) D 0 = diag ( α 0 ) , the spectral information is I s p , 0 = A 0 T D 0 − 1 A 0 = 2 + 10 / 3 = 16 / 3 I_{\mathrm{sp},0} = A_0^{T}D_0^{-1}A_0 = 2 + 10/3 = 16/3 I sp , 0 = A 0 T D 0 − 1 A 0 = 2 + 10/3 = 16/3 . Since the model is diagonal, its SLD at zero is L 0 , 1 = diag ( 2 , − 10 / 3 , 0 ) L_{0,1} = \operatorname{diag}(2,-10/3,0) L 0 , 1 = diag ( 2 , − 10/3 , 0 ) , and hence
G 0 = Tr ( ρ 0 L 0 , 1 2 ) = 1 2 ⋅ 2 2 + 3 10 ( − 10 3 ) 2 = 16 3 = I s p , 0 . G_0 = \operatorname{Tr}(\rho_0L_{0,1}^{2}) = \frac{1}{2}\cdot2^2 + \frac{3}{10}\left(-\frac{10}{3}\right)^2 = \frac{16}{3} = I_{\mathrm{sp},0}. G 0 = Tr ( ρ 0 L 0 , 1 2 ) = 2 1 ⋅ 2 2 + 10 3 ( − 3 10 ) 2 = 3 16 = I sp , 0 .
Consequently, V e f f = c 0 2 / G 0 = 3 / 100 V_{\mathrm{eff}} = c_0^2/G_0 = 3/100 V eff = c 0 2 / G 0 = 3/100 and V s p = c 0 2 / I s p , 0 = 3 / 100 V_{\mathrm{sp}} = c_0^2/I_{\mathrm{sp},0} = 3/100 V sp = c 0 2 / I sp , 0 = 3/100 . The efficient spectral weight is
w 0 = D 0 − 1 A 0 I s p , 0 − 1 c 0 = ( 2 − 10 / 3 0 ) 3 16 2 5 = ( 3 / 20 − 1 / 4 0 ) . w_0 = D_0^{-1}A_0I_{\mathrm{sp},0}^{-1}c_0 = \begin{pmatrix} 2 \\ -10/3 \\ 0 \end{pmatrix}\frac{3}{16}\frac{2}{5} = \begin{pmatrix} 3/20 \\ -1/4 \\ 0 \end{pmatrix}. w 0 = D 0 − 1 A 0 I sp , 0 − 1 c 0 = 2 − 10/3 0 16 3 5 2 = 3/20 − 1/4 0 .
For the unrestricted plug-in, α 0 T b 0 = 19 / 25 \boldsymbol{\alpha}_0^{T}b_0 = 19/25 α 0 T b 0 = 19/25 , so
b ˉ 0 = b 0 − ( α 0 T b 0 ) 1 = ( 6 / 25 − 4 / 25 − 9 / 25 ) . \bar{b}_0 = b_0 - (\boldsymbol{\alpha}_0^{T}b_0)\boldsymbol{1} = \begin{pmatrix} 6/25 \\ -4/25 \\ -9/25 \end{pmatrix}. b ˉ 0 = b 0 − ( α 0 T b 0 ) 1 = 6/25 − 4/25 − 9/25 .
Its asymptotic variance is
V W S S = b ˉ 0 T C 0 b ˉ 0 = b 0 T D 0 b 0 − ( α 0 T b 0 ) 2 = 1 2 + 3 10 ( 3 5 ) 2 + 1 5 ( 2 5 ) 2 − ( 19 25 ) 2 = 39 625 . \begin{aligned}
V_{\mathrm{WSS}} = \bar{b}_0^{T}C_0\bar{b}_0 = b_0^{T}D_0b_0 - (\boldsymbol{\alpha}_0^{T}b_0)^2 \\
&= \frac{1}{2} + \frac{3}{10}\left(\frac{3}{5}\right)^2 + \frac{1}{5}\left(\frac{2}{5}\right)^2 - \left(\frac{19}{25}\right)^2 = \frac{39}{625}.
\end{aligned} V WSS = b ˉ 0 T C 0 b ˉ 0 = b 0 T D 0 b 0 − ( α 0 T b 0 ) 2 = 2 1 + 10 3 ( 5 3 ) 2 + 5 1 ( 5 2 ) 2 − ( 25 19 ) 2 = 625 39 .
Thus
V e f f = V s p = 3 100 , V W S S = 39 625 , V W S S − V s p = 81 2500 . (65) V_{\mathrm{eff}} = V_{\mathrm{sp}} = \frac{3}{100}, \qquad V_{\mathrm{WSS}} = \frac{39}{625}, \qquad V_{\mathrm{WSS}} - V_{\mathrm{sp}} = \frac{81}{2500}. \tag*{(65)} V eff = V sp = 100 3 , V WSS = 625 39 , V WSS − V sp = 2500 81 . (65)
We now identify the additional fluctuation of the unrestricted plug-in. Let h n → h h_n \to h h n → h , put θ n = h n / n \theta_n = h_n/\sqrt{n} θ n = h n / n , and define Z n : = n ( α ^ n − α 0 ) = ( Z n , 1 , Z n , 2 , Z n , 3 ) T Z_n := \sqrt{n}(\hat{\boldsymbol{\alpha}}_n - \boldsymbol{\alpha}_0) = (Z_{n,1}, Z_{n,2}, Z_{n,3})^{T} Z n := n ( α ^ n − α 0 ) = ( Z n , 1 , Z n , 2 , Z n , 3 ) T . By Lemma 6.1 , under P θ n ⊗ n \mathbb{P}_{\theta_n}^{\otimes n} P θ n ⊗ n ,
Z n ⇒ Z = ( Z 1 , Z 2 , Z 3 ) T ∼ N 3 ( A 0 h , C 0 ) . Z_n \Rightarrow Z = (Z_1,Z_2,Z_3)^{T} \sim N_3(A_0h,C_0). Z n ⇒ Z = ( Z 1 , Z 2 , Z 3 ) T ∼ N 3 ( A 0 h , C 0 ) .
In particular, 1 T Z = 0 \boldsymbol{1}^{T}Z = 0 1 T Z = 0 almost surely. Now observe that
b ˉ 0 − w 0 = ( 9 / 100 9 / 100 − 9 / 25 ) = 9 100 1 − 9 20 ( 0 0 1 ) . \bar{b}_0 - w_0 = \begin{pmatrix} 9/100 \\ 9/100 \\ -9/25 \end{pmatrix} = \frac{9}{100}\boldsymbol{1} - \frac{9}{20}\begin{pmatrix} 0 \\ 0 \\ 1 \end{pmatrix}. b ˉ 0 − w 0 = 9/100 9/100 − 9/25 = 100 9 1 − 20 9 0 0 1 .
Since 1 T Z = 0 \boldsymbol{1}^{T}Z = 0 1 T Z = 0 , we obtain
b ˉ 0 T Z = b ˉ 0 T Z = w 0 T Z − 9 20 Z 3 . (66) \bar{b}_0^{T}Z = \bar{b}_0^{T}Z = w_0^{T}Z - \frac{9}{20}Z_3. \tag*{(66)} b ˉ 0 T Z = b ˉ 0 T Z = w 0 T Z − 20 9 Z 3 . (66)
Moreover, Cov ( w 0 T Z , Z 3 ) = α 0 , 3 ( w 0 ) 3 − ( α 0 T w 0 ) α 0 , 3 = 0 \operatorname{Cov}(w_0^{T}Z,Z_3) = \alpha_{0,3}(w_0)_3 - (\boldsymbol{\alpha}_0^{T}w_0)\alpha_{0,3} = 0 Cov ( w 0 T Z , Z 3 ) = α 0 , 3 ( w 0 ) 3 − ( α 0 T w 0 ) α 0 , 3 = 0 , and Var ( Z 3 ) = α 0 , 3 ( 1 − α 0 , 3 ) = 4 / 25 \operatorname{Var}(Z_3) = \alpha_{0,3}(1-\alpha_{0,3}) = 4/25 Var ( Z 3 ) = α 0 , 3 ( 1 − α 0 , 3 ) = 4/25 . Since Z Z Z is Gaussian, Z 3 ∼ N ( 0 , 4 / 25 ) Z_3 \sim N(0,4/25) Z 3 ∼ N ( 0 , 4/25 ) is independent of w 0 T Z w_0^{T}Z w 0 T Z . The model fixes the third eigenvalue, and the unrestricted plug-in adds the independent noise term in (66) , whose variance is ( 9 / 20 ) 2 ( 4 / 25 ) = 81 / 2500 = V W S S − V s p (9/20)^2(4/25)=81/2500=V_{\mathrm{WSS}}-V_{\mathrm{sp}} ( 9/20 ) 2 ( 4/25 ) = 81/2500 = V WSS − V sp .
An explicit nonlinear post-processing with the efficient limit is
θ ^ n s p = 4 5 α ^ n , 1 α ^ n , 1 + α ^ n , 2 − 1 2 , ψ ^ n s p = 2 5 θ ^ n s p + 2 ( θ ^ n s p ) 2 . \widehat{\theta}^{\mathrm{sp}}_n=\frac{4}{5}\frac{\widehat{\alpha}_{n,1}}{\widehat{\alpha}_{n,1}+\widehat{\alpha}_{n,2}}-\frac{1}{2},\qquad\widehat{\psi}^{\mathrm{sp}}_n=\frac{2}{5}\widehat{\theta}^{\mathrm{sp}}_n+2(\widehat{\theta}^{\mathrm{sp}}_n)^2. θ n sp = 5 4 α n , 1 + α n , 2 α n , 1 − 2 1 , ψ n sp = 5 2 θ n sp + 2 ( θ n sp ) 2 .
The denominator never vanishes for n ≥ 1 n\ge1 n ≥ 1 , since α ^ n = λ ( n ) / n \widehat{\alpha}_n=\lambda^{(n)}/n α n = λ ( n ) / n is a decreasingly ordered probability vector and α ^ n , 1 + α ^ n , 2 ≥ α ^ n , 1 ≥ 1 / 3 \widehat{\alpha}_{n,1}+\widehat{\alpha}_{n,2}\ge\widehat{\alpha}_{n,1}\ge1/3 α n , 1 + α n , 2 ≥ α n , 1 ≥ 1/3 .
To compute the local limit, the gradient of ( x 1 , x 2 ) ↦ ( 4 / 5 ) x 1 / ( x 1 + x 2 ) − 1 / 2 (x_1,x_2)\mapsto(4/5)x_1/(x_1+x_2)-1/2 ( x 1 , x 2 ) ↦ ( 4/5 ) x 1 / ( x 1 + x 2 ) − 1/2 at ( 1 / 2 , 3 / 10 ) (1/2,3/10) ( 1/2 , 3/10 ) is ( 3 / 8 , − 5 / 8 ) T (3/8,-5/8)^{T} ( 3/8 , − 5/8 ) T . Since ψ ′ ( 0 ) = 2 / 5 \psi'(0)=2/5 ψ ′ ( 0 ) = 2/5 , the chain rule gives
n ψ ^ n s p = 2 5 ( 3 8 Z n , 1 − 5 8 Z n , 2 ) + o P ( 1 ) = 3 20 Z n , 1 − 1 4 Z n , 2 + o P ( 1 ) = w 0 T Z n + o P ( 1 ) . \begin{aligned}
\sqrt{n}\widehat{\psi}^{\mathrm{sp}}_n=\frac{2}{5}\left(\frac{3}{8}Z_{n,1}-\frac{5}{8}Z_{n,2}\right)+o_{\mathbb{P}}(1) \\
&=\frac{3}{20}Z_{n,1}-\frac{1}{4}Z_{n,2}+o_{\mathbb{P}}(1)=w_0^{T}Z_n+o_{\mathbb{P}}(1).
\end{aligned} n ψ n sp = 5 2 ( 8 3 Z n , 1 − 8 5 Z n , 2 ) + o P ( 1 ) = 20 3 Z n , 1 − 4 1 Z n , 2 + o P ( 1 ) = w 0 T Z n + o P ( 1 ) .
Consequently, under ρ θ n ⊗ n \rho_{\theta_n}^{\otimes n} ρ θ n ⊗ n ,
n ψ ^ n s p ⇒ N ( c 0 h , V s p ) = N ( 2 5 h , 3 100 ) . \sqrt{n}\widehat{\psi}^{\mathrm{sp}}_n\Rightarrow N(c_0h,V_{\mathrm{sp}})=N\left(\frac{2}{5}h,\frac{3}{100}\right). n ψ n sp ⇒ N ( c 0 h , V sp ) = N ( 5 2 h , 100 3 ) .
Thus efficient post-processing of the same Schur outcome achieves V s p = V e f f V_{\mathrm{sp}}=V_{\mathrm{eff}} V sp = V eff , whereas the unrestricted plug-in has variance V W S S V_{\mathrm{WSS}} V WSS . The corresponding variance ratio is V W S S / V s p = 52 / 25 V_{\mathrm{WSS}}/V_{\mathrm{sp}}=52/25 V WSS / V sp = 52/25 .
Example 6.5 (A gap beyond optimal linear WSS post-processing). For ∣ θ ∣ < 1 / 20 |\theta|<1/20 ∣ θ ∣ < 1/20 , let
D θ = diag ( 1 2 + θ , 3 10 − θ , 1 5 ) , H = ( 0 1 0 1 0 0 0 0 0 ) . D_\theta=\operatorname{diag}\left(\frac{1}{2}+\theta,\frac{3}{10}-\theta,\frac{1}{5}\right),\qquad H=\begin{pmatrix}0&1&0\\1&0&0\\0&0&0\end{pmatrix}. D θ = diag ( 2 1 + θ , 10 3 − θ , 5 1 ) , H = 0 1 0 1 0 0 0 0 0 . and put U θ = exp ( − i ω θ H ) U_\theta=\exp(-\mathrm{i}\omega\theta H) U θ = exp ( − i ω θ H ) and ρ θ ( ω ) = U θ D θ U θ ∗ \rho_\theta^{(\omega)}=U_\theta D_\theta U_\theta^* ρ θ ( ω ) = U θ D θ U θ ∗ , where ω \omega ω is a known fixed real constant. We consider the same functional ψ ( θ ) = Tr ( ( ρ θ ( ω ) ) 2 ) − 19 / 50 = 2 θ / 5 + 2 θ 2 \psi(\theta)=\operatorname{Tr}((\rho_\theta^{(\omega)})^2)-19/50=2\theta/5+2\theta^2 ψ ( θ ) = Tr (( ρ θ ( ω ) ) 2 ) − 19/50 = 2 θ /5 + 2 θ 2 .
Since unitary conjugation preserves the spectrum,
α θ = ( 1 / 2 + θ 3 / 10 − θ 1 / 5 ) , A 0 = d α θ d θ ∣ θ = 0 = ( 1 − 1 0 ) . \alpha_\theta=\begin{pmatrix}1/2+\theta\\3/10-\theta\\1/5\end{pmatrix},\qquad A_0=\left.\frac{\mathrm{d}\alpha_\theta}{\mathrm{d}\theta}\right|_{\theta=0}=\begin{pmatrix}1\\-1\\0\end{pmatrix}. α θ = 1/2 + θ 3/10 − θ 1/5 , A 0 = d θ d α θ θ = 0 = 1 − 1 0 . As in Example 6.4, α 0 = ( 1 / 2 , 3 / 10 , 1 / 5 ) T \alpha_0=(1/2,3/10,1/5)^T α 0 = ( 1/2 , 3/10 , 1/5 ) T , b 0 = 2 α 0 = ( 1 , 3 / 5 , 2 / 5 ) T b_0=2\alpha_0=(1,3/5,2/5)^T b 0 = 2 α 0 = ( 1 , 3/5 , 2/5 ) T and c 0 = A 0 T b 0 = 2 / 5 c_0=A_0^Tb_0=2/5 c 0 = A 0 T b 0 = 2/5 . Writing D 0 = diag ( α 0 ) D_0=\operatorname{diag}(\alpha_0) D 0 = diag ( α 0 ) , we obtain
I s p , 0 = A 0 T D 0 − 1 A 0 = 16 3 , I_{\mathrm{sp},0}=A_0^TD_0^{-1}A_0=\frac{16}{3}, I sp , 0 = A 0 T D 0 − 1 A 0 = 3 16 , w 0 = D 0 − 1 A 0 I s p , 0 − 1 c 0 = ( 3 / 20 − 1 / 4 0 ) . w_0=D_0^{-1}A_0I_{\mathrm{sp},0}^{-1}c_0=\begin{pmatrix}3/20\\-1/4\\0\end{pmatrix}. w 0 = D 0 − 1 A 0 I sp , 0 − 1 c 0 = 3/20 − 1/4 0 . Hence V s p = c 0 2 / I s p , 0 = 3 / 100 V_{\mathrm{sp}}=c_0^2/I_{\mathrm{sp},0}=3/100 V sp = c 0 2 / I sp , 0 = 3/100 and V W S S = b 0 T C 0 b 0 = 39 / 625 V_{\mathrm{WSS}}=b_0^TC_0b_0=39/625 V WSS = b 0 T C 0 b 0 = 39/625 . These quantities and the WSS law are unchanged for every ω \omega ω .
We now compute the SLD information. Since U 0 = I 3 U_0=I_3 U 0 = I 3 , we have ρ 0 = D 0 \rho_0=D_0 ρ 0 = D 0 , and
∂ θ ρ θ ( ω ) ∣ θ = 0 = diag ( 1 , − 1 , 0 ) − i ω [ H , D 0 ] = ( 1 i ω / 5 0 − i ω / 5 − 1 0 0 0 0 ) . \left.\partial_\theta\rho_\theta^{(\omega)}\right|_{\theta=0}=\operatorname{diag}(1,-1,0)-i\omega[H,D_0]=\begin{pmatrix}1&i\omega/5&0\\-i\omega/5&-1&0\\0&0&0\end{pmatrix}. ∂ θ ρ θ ( ω ) θ = 0 = diag ( 1 , − 1 , 0 ) − iω [ H , D 0 ] = 1 − iω /5 0 iω /5 − 1 0 0 0 0 .
The SLD at zero, denoted by L 0 ( ω ) L_0^{(\omega)} L 0 ( ω ) , satisfies 2 ∂ θ ρ θ ( ω ) ∣ θ = 0 = ρ 0 L 0 ( ω ) + L 0 ( ω ) ρ 0 2\left.\partial_\theta\rho_\theta^{(\omega)}\right|_{\theta=0}=\rho_0L_0^{(\omega)}+L_0^{(\omega)}\rho_0 2 ∂ θ ρ θ ( ω ) θ = 0 = ρ 0 L 0 ( ω ) + L 0 ( ω ) ρ 0 . Thus
L 0 ( ω ) = ( 2 i ω / 2 0 − i ω / 2 − 10 / 3 0 0 0 0 ) . L_0^{(\omega)}=\begin{pmatrix}2&i\omega/2&0\\-i\omega/2&-10/3&0\\0&0&0\end{pmatrix}. L 0 ( ω ) = 2 − iω /2 0 iω /2 − 10/3 0 0 0 0 .
The diagonal contribution to the SLD information is 1 2 ⋅ 2 2 + 3 10 ( − 10 / 3 ) 2 = 16 / 3 = I s p , 0 \frac{1}{2}\cdot2^2+\frac{3}{10}(-10/3)^2=16/3=I_{\mathrm{sp},0} 2 1 ⋅ 2 2 + 10 3 ( − 10/3 ) 2 = 16/3 = I sp , 0 . The off-diagonal contribution, denoted by K 0 K_0 K 0 , is
K 0 = ( α 0 , 1 + α 0 , 2 ) ∣ ( L 0 ( ω ) ) 12 ∣ 2 = ( 1 2 + 3 10 ) ω 2 4 = ω 2 5 . K_0=(\alpha_{0,1}+\alpha_{0,2})\left|(L_0^{(\omega)})_{12}\right|^2=\left(\frac{1}{2}+\frac{3}{10}\right)\frac{\omega^2}{4}=\frac{\omega^2}{5}. K 0 = ( α 0 , 1 + α 0 , 2 ) ( L 0 ( ω ) ) 12 2 = ( 2 1 + 10 3 ) 4 ω 2 = 5 ω 2 .
Consequently, G 0 = Tr ( ρ 0 ( L 0 ( ω ) ) 2 ) = I s p , 0 + K 0 = 16 / 3 + ω 2 / 5 G_0=\operatorname{Tr}(\rho_0(L_0^{(\omega)})^2)=I_{\mathrm{sp},0}+K_0=16/3+\omega^2/5 G 0 = Tr ( ρ 0 ( L 0 ( ω ) ) 2 ) = I sp , 0 + K 0 = 16/3 + ω 2 /5 . It follows that
V e f f = c 0 2 G 0 = ( 2 / 5 ) 2 16 / 3 + ω 2 / 5 = 12 400 + 15 ω 2 . (67) V_{\mathrm{eff}}=\frac{c_0^2}{G_0}=\frac{(2/5)^2}{16/3+\omega^2/5}=\frac{12}{400+15\omega^2}. \tag*{(67)} V eff = G 0 c 0 2 = 16/3 + ω 2 /5 ( 2/5 ) 2 = 400 + 15 ω 2 12 . (67)
For example, ω = 4 \omega=4 ω = 4 gives
V e f f = 3 160 < V s p = 3 100 < V W S S = 39 625 . V_{\mathrm{eff}}=\frac{3}{160}<V_{\mathrm{sp}}=\frac{3}{100}<V_{\mathrm{WSS}}=\frac{39}{625}. V eff = 160 3 < V sp = 100 3 < V WSS = 625 39 .
Thus the additional information is supplied by the parameter-dependent rotation, which WSS does not record.
Finally, for ω ≠ 0 \omega\ne0 ω = 0 , D ρ 0 L 0 ( ω ) D_{\rho_0}L_0^{(\omega)} D ρ 0 L 0 ( ω ) is a nonzero matrix with zero diagonal, whereas every nonzero real multiple of L 0 ( ω ) L_0^{(\omega)} L 0 ( ω ) has nonzero diagonal. Hence the one-dimensional SLD space is not D ρ 0 D_{\rho_0} D ρ 0 -invariant.
Connection with quantum-to-classical sufficiency
Our results are directly related to quantum-to-classical sufficiency, which separates selection of a measurement from optimization of its classical decision rule. In [30 Theorem 5.1] , WSS is shown to be sufficient relative to unitarily invariant POVMs: for every such measurement N N N , a parameter-independent Markov kernel K N K_N K N satisfies
P ρ N = K N P ρ W S S for every state ρ . (68) P_\rho^N=K_NP_\rho^{\mathrm{WSS}}\quad\text{for every state }\rho. \tag*{(68)} P ρ N = K N P ρ WSS for every state ρ . (68)
Together with appropriate symmetrization, this yields exact Bayes and minimax reductions to WSS for the invariant decision problems treated in [30 Theorems 5.4 and 5.6] .
Equation (68) continues to hold when the state model is restricted to any subfamily. What may fail on a restricted family is the separate justification for restricting all quantum procedures to invariant ones. For a family ρ θ = U θ diag ( α θ ) U θ ∗ \rho_\theta=U_\theta\operatorname{diag}(\alpha_\theta)U_\theta^* ρ θ = U θ diag ( α θ ) U θ ∗ , arbitrary conjugation can break the defining relation between spectrum and eigenbasis. Symmetrization then need not preserve the decision problem. Example 6.5 exhibits precisely this effect. In contrast, when all conjugates of every admissible spectrum belong to the parameter space and the eigenbasis is treated as nuisance, the relevant symmetry reduction is compatible with the model. Finally, equality of the variances for one functional is much weaker than experiment-level sufficiency. It concerns one scalar derivative at one reference point. Quantum-to-classical sufficiency concerns all decision rules induced by the prescribed measurement class and, in (68) , holds at every sample size. Conversely, first-order inefficiency of one plug-in does not contradict sufficiency of the observation from which it is computed.
Discussion
In this paper, we have developed a local asymptotic maximin theory for quantum testing using the asymptotic representation theorem of [14 ] . Under the stated asymptotic D D D -invariance assumptions, the Gaussian bounds constrain the local power of the admissible tests, and scalar SLD procedures attain these envelopes in local asymptotic power. We also construct adaptive tests which have the same local powers under their own smoothness assumptions; identifying those powers with the general testing bound additionally requires the hypotheses of § 4 . For spectral functionals, we compare the plug-in rule, the optimal linear first-order WSS statistic, and the scalar SLD statistic and characterize conditions under which the gaps V W S S − V s p V_{\mathrm{WSS}}-V_{\mathrm{sp}} V WSS − V sp and V s p − V e f f V_{\mathrm{sp}}-V_{\mathrm{eff}} V sp − V eff vanish.
Several limitations of the present work suggest natural directions for future research. First, our analysis relies crucially on the smoothness of the functional being tested. Extending the theory to nonsmooth functionals will likely require substantially different techniques. Second, although our general testing results apply to smooth functionals that may depend on both eigenvalues and eigenvectors, our detailed measurement comparison concerns spectral functionals. It would be valuable to develop corresponding comparisons for eigenvector-dependent functionals, particularly in models outside the asymptotic D D D -invariance framework, and to identify practically implementable procedures with sharp local power guarantees.
A further direction concerns the assumption of a simple spectrum. Our theory of relative efficiency relies critically on this assumption because the limiting power is derived using the Gaussian approximation of [27 ] . When the spectrum contains repeated eigenvalues, this approximation no longer applies. Indeed, as shown in [35 ] , the relevant limiting distribution is then non-Gaussian, and its analysis requires fundamentally different techniques. Developing an efficiency theory for quantum models with repeated eigenvalues lies beyond the scope of the present paper and remains an important problem for future work.
Finally, although the underlying quantum state may be high-dimensional, the functionals of interest are often low-dimensional, suggesting that suitably targeted inferential procedures may still be possible. However, as the literature on high-dimensional classical statistics demonstrates, conventional LAN theory generally cannot be applied directly in such regimes. Addressing high-dimensional quantum models will therefore require alternative notions of asymptotic optimality, together with inferential and computational methods that go beyond those developed here. We leave these questions for future investigation.
Acknowledgements
K.Y. was supported by JSPS KAKENHI Grant Numbers JP23H01090 and JP22K03466 and JST ERATO Grant Number JPMJER2402, Japan. R.N.K. was supported in part by JSPS KAKENHI Grant Numbers JP22H05105 and JP23K25774. R.N.K. also acknowledges financial support from the MEXT Scholarship for Research Students.
Disclosure of AI Use
GPT 5 and 6 were used to assist with reviewing the manuscript, checking mathematical arguments, and copyediting. The authors have independently verified the content and take full responsibility for its correctness.
Weak Schur sampling and the Schur-Weyl distribution
Let H d : = C d \mathcal{H}_d := \mathbb{C}^d H d := C d . We first define the representations of the group U ( d ) U(d) U ( d ) and S ( n ) S(n) S ( n ) (the group of unitary matrices and permutations respectively) on ( C d ) ⊗ n (\mathbb{C}^d)^{\otimes n} ( C d ) ⊗ n . Let f a = f a 1 ⊗ ⋯ ⊗ f a n f_{\mathbf{a}} = f_{a_1} \otimes\cdots\otimes f_{a_n} f a = f a 1 ⊗ ⋯ ⊗ f a n , where f 1 , … , f d f_1,\ldots,f_d f 1 , … , f d are basis elements of C d \mathbb{C}^d C d and a i ∈ { 1 , … , d } a_i \in\{1,\ldots,d\} a i ∈ { 1 , … , d } (note that f a f_{\mathbf{a}} f a 's span ( C d ) ⊗ n (\mathbb{C}^d)^{\otimes n} ( C d ) ⊗ n ) and consider the following actions:
π n ( T ) : f a 1 ⊗ ⋯ ⊗ f a n → T f a 1 ⊗ ⋯ ⊗ T f a n , T ∈ U ( d ) (69) \pi_n(T): f_{a_1} \otimes\cdots\otimes f_{a_n} \to T f_{a_1} \otimes\cdots\otimes T f_{a_n}, \qquad T \in U(d) \tag*{(69)} π n ( T ) : f a 1 ⊗ ⋯ ⊗ f a n → T f a 1 ⊗ ⋯ ⊗ T f a n , T ∈ U ( d ) (69)
π ~ n ( σ ) : f a 1 ⊗ ⋯ ⊗ f a n → f a σ − 1 ( 1 ) ⊗ ⋯ ⊗ f a σ − 1 ( n ) , σ ∈ S ( n ) . (70) \widetilde{\pi}_n(\sigma): f_{a_1} \otimes\cdots\otimes f_{a_n} \to f_{a_{\sigma^{-1}(1)}} \otimes\cdots\otimes f_{a_{\sigma^{-1}(n)}}, \qquad\sigma\in S(n). \tag*{(70)} π n ( σ ) : f a 1 ⊗ ⋯ ⊗ f a n → f a σ − 1 ( 1 ) ⊗ ⋯ ⊗ f a σ − 1 ( n ) , σ ∈ S ( n ) . (70)
It can be shown that the representation space can be decomposed into direct sum of subspaces each of which is a tensor product of irreducible representations of U ( d ) U(d) U ( d ) and S ( n ) S(n) S ( n ) . This decomposition is called Schur-Weyl duality in the literature (for a detailed account see [15 , 17 ] ).
Define a partition of n n n as a vector λ = ( λ 1 , … , λ d ) \lambda= (\lambda_1,\ldots,\lambda_d) λ = ( λ 1 , … , λ d ) with
λ 1 ≥ ⋯ ≥ λ d ≥ 0 , ∑ i = 1 d λ i = n . \lambda_1 \ge\cdots\ge\lambda_d \ge0,\qquad\sum_{i=1}^{d}\lambda_i = n. λ 1 ≥ ⋯ ≥ λ d ≥ 0 , i = 1 ∑ d λ i = n .
A Young diagram is defined by an ordered tuple of integers λ = ( λ 1 , … , λ d ) \lambda= (\lambda_1,\ldots,\lambda_d) λ = ( λ 1 , … , λ d ) with λ 1 ≥ ⋯ ≥ λ d ≥ 0 \lambda_1 \ge\cdots\ge\lambda_d \ge0 λ 1 ≥ ⋯ ≥ λ d ≥ 0 and can be represented graphically by a diagram with d d d lines each containing λ i \lambda_i λ i boxes. For example a typical Young diagram looks like:
Figure 1.
with λ = ( 3 , 2 , 1 ) \lambda= (3,2,1) λ = ( 3 , 2 , 1 ) .
Thus, a partition of n n n can be identified with a Young diagram with n n n boxes. Recall that
Y n , d = { λ ⊢ n : ℓ ( λ ) ≤ d } \mathbb{Y}_{n,d} = \{\lambda\vdash n : \ell(\lambda) \le d\} Y n , d = { λ ⊢ n : ℓ ( λ ) ≤ d }
is the set of partitions of n n n with at most d d d nonzero parts, equivalently Young diagrams with at most d d d rows.
Schur-Weyl duality gives a unitary decomposition
H d ⊗ n ≅ ⨁ λ ∈ Y n , d U λ ⊗ V λ . \mathcal{H}_d^{\otimes n} \cong\bigoplus_{\lambda\in\mathbb{Y}_{n,d}} \mathcal{U}_{\lambda} \otimes\mathcal{V}_{\lambda}. H d ⊗ n ≅ λ ∈ Y n , d ⨁ U λ ⊗ V λ .
where U λ \mathcal{U}_{\lambda} U λ carries an irreducible representation of U ( d ) U(d) U ( d ) and V λ \mathcal{V}_{\lambda} V λ carries an irreducible representation of S n S_n S n .
Weak Schur sampling and Schur-Weyl distribution
Let Π λ \Pi_{\lambda} Π λ denote the orthogonal projection onto the isotypic component U λ ⊗ V λ \mathcal{U}_{\lambda} \otimes\mathcal{V}_{\lambda} U λ ⊗ V λ .
Definition 8 (Weak Schur sampling). Weak Schur sampling is the projective measurement
W S S n : = { Π λ : λ ∈ Y n , d } . \mathrm{WSS}_n := \{\Pi_{\lambda} : \lambda\in\mathbb{Y}_{n,d}\}. WSS n := { Π λ : λ ∈ Y n , d } . Its outcome is the Young diagram λ \lambda λ .
The induced classical experiment is
F n : = { P ρ W S S n : ρ ∈ S ( C d ) } , \mathcal{F}_n := \left\{P_{\rho}^{\mathrm{WSS}_n} : \rho\in\mathcal{S}(\mathbb{C}^d)\right\}, F n := { P ρ WSS n : ρ ∈ S ( C d ) } ,
where
P ρ W S S n ( λ ) = Tr ( ρ ⊗ n Π λ ) . P_{\rho}^{\mathrm{WSS}_n}(\lambda) = \operatorname{Tr}\left(\rho^{\otimes n}\Pi_\lambda\right). P ρ WSS n ( λ ) = Tr ( ρ ⊗ n Π λ ) .
It can be shown that (cf. [30 ] ) the weak Schur-sampling distribution is constant on unitary orbits and therefore depends on ρ \rho ρ only through its spectrum. In particular the following holds.
Corollary A.1 (Unitary invariance of the weak Schur distribution). Let ρ , σ ∈ S ( C d ) \rho, \sigma\in\mathcal{S}(\mathbb{C}^d) ρ , σ ∈ S ( C d ) be unitarily conjugate:
σ = U ρ U † \sigma= U\rho U^\dagger σ = U ρ U † for some U ∈ U ( d ) U \in U(d) U ∈ U ( d ) . Then
Tr ( σ ⊗ n Π λ ) = Tr ( ρ ⊗ n Π λ ) \operatorname{Tr}\left(\sigma^{\otimes n}\Pi_\lambda\right) = \operatorname{Tr}\left(\rho^{\otimes n}\Pi_\lambda\right) Tr ( σ ⊗ n Π λ ) = Tr ( ρ ⊗ n Π λ ) for every λ ∈ Y n , d \lambda\in\mathbb{Y}_{n,d} λ ∈ Y n , d .
We may therefore write
P θ ( n ) ( λ ) : = P ρ W S S n ( λ ) P_{\theta}^{(n)}(\lambda) := P_{\rho}^{\mathrm{WSS}_n}(\lambda) P θ ( n ) ( λ ) := P ρ WSS n ( λ )
whenever spec ( ρ ) = θ \operatorname{spec}(\rho) = \theta spec ( ρ ) = θ .
Next, we describe the Schur-Weyl distribution using standard and semistandard Young tableaux. For a cell u = ( i , j ) u=(i,j) u = ( i , j ) of λ \lambda λ , let λ j ′ \lambda'_j λ j ′ be the number of rows containing at least j j j cells. Its hook length is
h λ ( i , j ) = λ i − j + λ j ′ − i + 1. h_\lambda(i,j) = \lambda_i-j+\lambda'_j-i+1. h λ ( i , j ) = λ i − j + λ j ′ − i + 1.
The number of standard Young tableaux of shape λ \lambda λ is
f λ = n ! ∏ u ∈ λ h λ ( u ) . f^\lambda= \frac{n!}{\prod_{u\in\lambda}h_\lambda(u)}. f λ = ∏ u ∈ λ h λ ( u ) n ! .
Let SSYT d ( λ ) \operatorname{SSYT}_d(\lambda) SSYT d ( λ ) denote the semistandard Young tableaux of shape λ \lambda λ with entries in { 1 , … , d } \{1,\ldots,d\} { 1 , … , d } . Entries increase weakly along rows and strictly down columns. If m i ( T ) m_i(T) m i ( T ) counts the entries equal to i i i , define the Schur polynomial by
s λ ( θ ) = ∑ T ∈ SSYT d ( λ ) ∏ i = 1 d θ i m i ( T ) . s_\lambda(\theta)=\sum_{T\in\operatorname{SSYT}_d(\lambda)}\prod_{i=1}^{d}\theta_i^{m_i(T)}. s λ ( θ ) = T ∈ SSYT d ( λ ) ∑ i = 1 ∏ d θ i m i ( T ) .
The Schur-Weyl distribution is
P θ ( n ) ( λ ) = f λ s λ ( θ ) , λ ∈ Y n , d . P_\theta^{(n)}(\lambda)=f^\lambda s_\lambda(\theta),\qquad\lambda\in\mathbb{Y}_{n,d}. P θ ( n ) ( λ ) = f λ s λ ( θ ) , λ ∈ Y n , d .
Local Asymptotic Normality and Asymptotic Representation Theorem in the classical case
In this section, we recall the classical local asymptotic normality framework and the associated Gaussian representation of limiting statistics. These results reduce the analysis of local testing power to a Gaussian shift experiment. Our presentation follows van der Vaart [44 Chapters 7 and 15] .
Local asymptotic normality
Consider a sequence of statistical experiments
E n = ( X n , A n , { P n , θ : θ ∈ Θ } ) , \mathcal{E}_n = (\mathcal{X}_n, \mathcal{A}_n, \{P_{n,\theta}: \theta\in\Theta\}), E n = ( X n , A n , { P n , θ : θ ∈ Θ }) ,
where Θ ⊂ R p \Theta\subset\mathbb{R}^p Θ ⊂ R p is open, and fix θ 0 ∈ Θ \theta_0 \in\Theta θ 0 ∈ Θ . We work with dominated experiments and write p n , θ p_{n,\theta} p n , θ for densities relative to a common measure μ n \mu_n μ n . Let R n R_n R n be nonsingular p × p p \times p p × p matrices satisfying ∥ R n − 1 ∥ → 0 \lVert R_n^{-1}\rVert\to0 ∥ R n − 1 ∥ → 0 , and set
P n , h : = P n , θ 0 + R n − 1 h , P n , 0 : = P n , θ 0 . P_{n,h} := P_{n,\theta_0+R_n^{-1}h}, \qquad P_{n,0} := P_{n,\theta_0}. P n , h := P n , θ 0 + R n − 1 h , P n , 0 := P n , θ 0 .
For every bounded sequence of local parameters, these distributions are defined for all sufficiently large n n n .
Definition 9 (Local asymptotic normality). The sequence ( E n ) (\mathcal{E}_n) ( E n ) is locally asymptotically normal at θ 0 \theta_0 θ 0 , with normalization R n R_n R n and information matrix I I I , if I I I is symmetric and nonnegative definite and there are statistics Δ n ∈ R p \Delta_n \in\mathbb{R}^p Δ n ∈ R p such that, under P n , 0 P_{n,0} P n , 0 ,
Δ n ⇒ N p ( 0 , I ) , (71) \Delta_n \Rightarrow\mathcal{N}_p(0,I), \tag*{(71)} Δ n ⇒ N p ( 0 , I ) , (71) and, for every deterministic sequence h n → h ∈ R p h_n \to h \in\mathbb{R}^p h n → h ∈ R p ,
log p n , θ 0 + R n − 1 h n p n , θ 0 = h T Δ n − 1 2 h T I h + o P n , 0 ( 1 ) . (72) \log\frac{p_{n,\theta_0+R_n^{-1}h_n}}{p_{n,\theta_0}} = h^\mathsf{T}\Delta_n - \frac{1}{2}h^\mathsf{T}Ih + o_{P_{n,0}}(1). \tag*{(72)} log p n , θ 0 p n , θ 0 + R n − 1 h n = h T Δ n − 2 1 h T I h + o P n , 0 ( 1 ) . (72) The likelihood ratio is evaluated at the observed data, and Δ n \Delta_n Δ n is called a central sequence.
This is the dominated-experiment formulation of [44 Definition 7.14, p. 104] . Tightness of Δ n \Delta_n Δ n allows h h h to be replaced by h n h_n h n on the right-hand side of (72) . In the scalar normalization used below, R n = r n I d p R_n = r_n I_{d_p} R n = r n I d p , where r n → ∞ r_n \to\infty r n → ∞ , so that local alternatives have the form θ 0 + h / r n \theta_0 + h/r_n θ 0 + h / r n .
For independent observations from a model with density p θ p_\theta p θ , a standard sufficient condition is differentiability in quadratic mean at θ 0 \theta_0 θ 0 : there exists ℓ ˙ θ 0 ∈ L 2 ( P θ 0 ) p \dot{\ell}_{\theta_0} \in L^2(P_{\theta_0})^p ℓ ˙ θ 0 ∈ L 2 ( P θ 0 ) p such that
∫ ( p θ 0 + t − p θ 0 − 1 2 t T ℓ ˙ θ 0 p θ 0 ) 2 d μ = o ( ∥ t ∥ 2 ) . (73) \int\left(\sqrt{p_{\theta_0+t}}-\sqrt{p_{\theta_0}}-\frac{1}{2}t^\mathsf{T}\dot{\ell}_{\theta_0}\sqrt{p_{\theta_0}}\right)^2\,d\mu= o(\lVert t\rVert^2). \tag*{(73)} ∫ ( p θ 0 + t − p θ 0 − 2 1 t T ℓ ˙ θ 0 p θ 0 ) 2 d μ = o (∥ t ∥ 2 ) . (73)
For P n , θ = P θ ⊗ n P_{n,\theta}=P_\theta^{\otimes n} P n , θ = P θ ⊗ n , this gives LAN with
R n = n I d p , Δ n = 1 n ∑ j = 1 n ℓ ˙ θ 0 ( X j ) , I = E θ 0 [ ℓ ˙ θ 0 ℓ ˙ θ 0 T ] , R_n=\sqrt{n}I_{d_p}, \qquad\Delta_n=\frac{1}{\sqrt{n}}\sum_{j=1}^{n}\dot{\ell}_{\theta_0}(X_j), \qquad I=E_{\theta_0}[\dot{\ell}_{\theta_0}\dot{\ell}_{\theta_0}^\mathsf{T}], R n = n I d p , Δ n = n 1 j = 1 ∑ n ℓ ˙ θ 0 ( X j ) , I = E θ 0 [ ℓ ˙ θ 0 ℓ ˙ θ 0 T ] ,
as in [44 Theorem 7.2 and Example 7.15] . The general definition also applies to experiments without a product structure.
Throughout the remainder of the appendix, assume that I I I is positive definite. The corresponding Gaussian shift experiment observes
Z ∼ Q h : = N p ( h , I − 1 ) , h ∈ R p . (74) Z \sim Q_h := \mathcal{N}_p(h,I^{-1}), \qquad h \in\mathbb{R}^p. \tag*{(74)} Z ∼ Q h := N p ( h , I − 1 ) , h ∈ R p . (74)
Its likelihood ratio satisfies
log d Q h d Q 0 ( z ) = h T I z − 1 2 h T I h . (75) \log\frac{dQ_h}{dQ_0}(z)=h^\mathsf{T}Iz-\frac{1}{2}h^\mathsf{T}Ih. \tag*{(75)} log d Q 0 d Q h ( z ) = h T I z − 2 1 h T I h . (75)
Comparing (72) and (75) identifies I − 1 Δ n I^{-1}\Delta_n I − 1 Δ n as the limiting Gaussian observation. More precisely, LAN implies mutual contiguity of P n , h n P_{n,h_n} P n , h n and P n , 0 P_{n,0} P n , 0 for every h n → h h_n \to h h n → h , and Le Cam’s third lemma gives
Δ n ⇒ N p ( I h , I ) , I − 1 Δ n ⇒ N p ( h , I − 1 ) under P n , h n . (76) \Delta_n \Rightarrow\mathcal{N}_p(Ih,I), \qquad I^{-1}\Delta_n \Rightarrow\mathcal{N}_p(h,I^{-1}) \quad\text{under } P_{n,h_n}. \tag*{(76)} Δ n ⇒ N p ( I h , I ) , I − 1 Δ n ⇒ N p ( h , I − 1 ) under P n , h n . (76)
Contiguity also transfers o P n , 0 ( 1 ) o_{P_{n,0}}(1) o P n , 0 ( 1 ) remainders to these local alternatives.
Asymptotic representation
The Gaussian experiment describes the possible limiting laws of statistics under local perturbations. The following formulation extends the independent-sample statement of [44 Theorem 7.10, p. 98] to LAN experiments, as explained in [44 Section 7.6, p. 104] .
Theorem B.1 (Asymptotic representation). Suppose ( E n ) (\mathcal{E}_n) ( E n ) is LAN at θ 0 \theta_0 θ 0 , with nonsingular information I I I . Let T n T_n T n be R q \mathbb{R}^q R q -valued statistics such that, for every fixed h ∈ R p h \in\mathbb{R}^p h ∈ R p ,
Law P n , h ( T n ) ⇒ L h . \operatorname{Law}_{P_{n,h}}(T_n) \Rightarrow L_h. Law P n , h ( T n ) ⇒ L h . There exists a Markov kernel K K K from R p \mathbb{R}^p R p to R q \mathbb{R}^q R q , independent of h h h , satisfying
L h ( B ) = ∫ R p K ( z , B ) Q h ( d z ) (77) L_h(B) = \int_{\mathbb{R}^p} K(z,B)\,Q_h(\mathrm{d}z) \tag*{(77)} L h ( B ) = ∫ R p K ( z , B ) Q h ( d z ) (77) for every Borel set B ⊂ R q B \subset\mathbb{R}^q B ⊂ R q and every h h h . Equivalently, there is a measurable map t t t , also independent of h h h , such that
L h = Law ( t ( Z , U ) ) , Z ∼ Q h , U ∼ Unif ( 0 , 1 ) , U ⊥ Z . L_h = \operatorname{Law}(t(Z,U)), \qquad Z \sim Q_h,\quad U \sim\operatorname{Unif}(0,1),\quad U \perp Z. L h = Law ( t ( Z , U )) , Z ∼ Q h , U ∼ Unif ( 0 , 1 ) , U ⊥ Z .
A local power bound
We now use the scalar normalization R n = r n I d R_n = r_n I_d R n = r n I d . Let ψ : Θ → R \psi: \Theta\to\mathbb{R} ψ : Θ → R be differentiable at θ 0 \theta_0 θ 0 , with
ψ ( θ 0 ) = 0 , c : = ∇ ψ ( θ 0 ) ≠ 0 , \psi(\theta_0)=0,\qquad c := \nabla\psi(\theta_0) \ne0, ψ ( θ 0 ) = 0 , c := ∇ ψ ( θ 0 ) = 0 ,
where c c c is a column vector. Consider the hypotheses
H 0 : ψ ( θ ) ≤ 0 and H 1 : ψ ( θ ) > 0. H_0 : \psi(\theta) \le0 \qquad\text{and}\qquad H_1 : \psi(\theta) > 0. H 0 : ψ ( θ ) ≤ 0 and H 1 : ψ ( θ ) > 0.
Because
r n ψ ( θ 0 + h / r n ) → c T h , r_n\psi(\theta_0+h/r_n) \to c^{\mathsf{T}}h, r n ψ ( θ 0 + h / r n ) → c T h ,
the local null and alternative are described by the half-spaces c T h ≤ 0 c^{\mathsf{T}}h \le0 c T h ≤ 0 and c T h > 0 c^{\mathsf{T}}h > 0 c T h > 0 . For a significance level 0 < η < 1 0 < \eta< 1 0 < η < 1 , write
σ 2 = c T I − 1 c , z η = Φ − 1 ( 1 − η ) , \sigma^2 = c^{\mathsf{T}}I^{-1}c,\qquad z_\eta= \Phi^{-1}(1-\eta), σ 2 = c T I − 1 c , z η = Φ − 1 ( 1 − η ) ,
where Φ \Phi Φ is the standard normal distribution function.
Theorem B.2 (Local asymptotic power bound). Suppose the experiments are LAN at θ 0 \theta_0 θ 0 , with scalar normalization r n → ∞ r_n \to\infty r n → ∞ and positive definite information matrix I I I . Under the preceding assumptions on ψ \psi ψ , every sequence of level-η \eta η tests, satisfying
sup θ ∈ Θ : ψ ( θ ) ≤ 0 E n , θ φ n ≤ η for each n , (78) \sup_{\theta\in\Theta:\,\psi(\theta)\le0}\mathbb{E}_{n,\theta}\varphi_n \le\eta\qquad\text{for each }n, \tag*{(78)} θ ∈ Θ : ψ ( θ ) ≤ 0 sup E n , θ φ n ≤ η for each n , (78) obeys, for every h h h with c T h > 0 c^{\mathsf{T}}h > 0 c T h > 0 ,
lim sup n → ∞ E n , θ 0 + h / r n φ n ≤ β η ∗ ( h ) : = 1 − Φ ( z η − c T h c T I − 1 c ) . (79) \limsup_{n\to\infty}\mathbb{E}_{n,\theta_0+h/r_n}\varphi_n
\le\beta_\eta^*(h) := 1-\Phi\left(z_\eta-\frac{c^{\mathsf{T}}h}{\sqrt{c^{\mathsf{T}}I^{-1}c}}\right). \tag*{(79)} n → ∞ lim sup E n , θ 0 + h / r n φ n ≤ β η ∗ ( h ) := 1 − Φ ( z η − c T I − 1 c c T h ) . (79)
This is the power bound of [44 Theorem 15.4, p. 219] , expressed in the present notation.
Attainment of the bound
The following result gives the attainment statement corresponding to [44 Addendum 15.5, p. 219] .
Proposition B.3 (Attainment of the local power envelope). Under the assumptions of Theorem B.2 , suppose that
S n = c T I − 1 Δ n σ + o P n , 0 ( 1 ) . (80) S_n = \frac{c^{\mathsf T}I^{-1}\Delta_n}{\sigma} + o_{P_{n,0}}(1). \tag*{(80)} S n = σ c T I − 1 Δ n + o P n , 0 ( 1 ) . (80) For every h n → h h_n \to h h n → h ,
S n ⇒ N ( c T h σ , 1 ) under P n , h n . S_n \Rightarrow\mathcal{N}\left(\frac{c^{\mathsf T}h}{\sigma},1\right) \qquad\text{under } P_{n,h_n}. S n ⇒ N ( σ c T h , 1 ) under P n , h n . Consequently, the tests ϕ n ⋆ = 1 { S n > z η } \phi_n^\star= 1\{S_n > z_\eta\} ϕ n ⋆ = 1 { S n > z η } have limiting power β η ⋆ ( h ) \beta_\eta^\star(h) β η ⋆ ( h ) for every fixed h h h . For every sequence h n → h h_n \to h h n → h satisfying ψ ( θ 0 + h n / r n ) ≤ 0 \psi(\theta_0 + h_n/r_n) \le0 ψ ( θ 0 + h n / r n ) ≤ 0 eventually,
lim sup n → ∞ E n , θ 0 + h n / r n ϕ n ⋆ ≤ η . \limsup_{n\to\infty} \mathbb{E}_{n,\theta_0+h_n/r_n}\phi_n^\star\le\eta. n → ∞ lim sup E n , θ 0 + h n / r n ϕ n ⋆ ≤ η .
An efficient estimator provides a standard implementation. Suppose
r n ( θ ^ n − θ 0 ) = I − 1 Δ n + o P n , 0 ( 1 ) , σ ^ n → σ in P n , 0 -probability , r_n(\widehat{\theta}_n-\theta_0) = I^{-1}\Delta_n + o_{P_{n,0}}(1), \qquad\widehat{\sigma}_n \to\sigma\text{ in } P_{n,0}\text{-probability}, r n ( θ n − θ 0 ) = I − 1 Δ n + o P n , 0 ( 1 ) , σ n → σ in P n , 0 -probability ,
where σ ^ n > 0 \widehat{\sigma}_n > 0 σ n > 0 . Differentiability of ψ \psi ψ and Slutsky’s lemma yield
W n : = r n ψ ( θ ^ n ) σ ^ n = c T I − 1 Δ n σ + o P n , 0 ( 1 ) . W_n := \frac{r_n\psi(\widehat{\theta}_n)}{\widehat{\sigma}_n} = \frac{c^{\mathsf T}I^{-1}\Delta_n}{\sigma} + o_{P_{n,0}}(1). W n := σ n r n ψ ( θ n ) = σ c T I − 1 Δ n + o P n , 0 ( 1 ) .
The Wald test rejecting when W n > z η W_n > z_\eta W n > z η therefore attains the local power envelope. In regular models, σ ^ n 2 \widehat{\sigma}_n^2 σ n 2 can be obtained by consistently estimating c T I − 1 c c^{\mathsf T}I^{-1}c c T I − 1 c ; see [44 Example 15.6, p. 220] .
Remark 5 (Dependence on the hypotheses). For the simple Gaussian hypotheses h = 0 h=0 h = 0 versus one specified alternative h h h , the Neyman-Pearson power is
1 − Φ ( z η − h T I h ) . 1-\Phi\left(z_\eta-\sqrt{h^{\mathsf T}Ih}\right). 1 − Φ ( z η − h T I h ) .
The half-space envelope (79) instead reflects size control along the entire local null, including nuisance directions tangent to its boundary.
Von Neumann algebra formulation of degenerate Gaussian models
In Subsection 2.2 we described Gaussian experiments using the usual Hilbert-space representation when the commutation matrix is nonsingular, and as classical-quantum Gaussian experiments when the commutation matrix is degenerate. In this appendix we give the corresponding von Neumann algebra formulation needed for the Gaussian testing results and for later quasi-characteristic-function calculations.
Weyl representations. Let S ∈ R d × d S \in\mathbb{R}^{d \times d} S ∈ R d × d be a real skew-symmetric matrix with d ≥ 1 d \ge1 d ≥ 1 . A regular Weyl representation associated with S S S is a strongly continuous family of unitaries { W S ( ξ ) : ξ ∈ R d } \{W_S(\xi): \xi\in\mathbb{R}^d\} { W S ( ξ ) : ξ ∈ R d } on a separable Hilbert space satisfying
W S ( ξ ) W S ( η ) = e − − 1 ξ T S η W S ( ξ + η ) , W S ( ξ ) ∗ = W S ( − ξ ) . (81) W_S(\xi)W_S(\eta)=e^{-\sqrt{-1}\xi^{T}S\eta}W_S(\xi+\eta), \qquad W_S(\xi)^*=W_S(-\xi). \tag*{(81)} W S ( ξ ) W S ( η ) = e − − 1 ξ T S η W S ( ξ + η ) , W S ( ξ ) ∗ = W S ( − ξ ) . (81)
We denote by
CCR ( S ) : = { W S ( ξ ) : ξ ∈ R d } ” \operatorname{CCR}(S):=\{W_S(\xi):\xi\in\mathbb{R}^d\}” CCR ( S ) := { W S ( ξ ) : ξ ∈ R d } ”
the von Neumann algebra generated by the Weyl operators. This notation depends on the representation; the concrete representation used throughout the remainder of this appendix is fixed below.
When S S S is nonsingular, d = 2 m d=2m d = 2 m and the Stone–von Neumann theorem implies that every regular irreducible representation is unitarily equivalent to the Schrödinger representation on L 2 ( R ) ⊗ m L^2(\mathbb{R})^{\otimes m} L 2 ( R ) ⊗ m . If
W S ( ξ ) = e − 1 ξ T X , X = ( X 1 , … , X 2 m ) T , W_S(\xi)=e^{\sqrt{-1}\xi^T X}, \qquad X=(X_1,\ldots,X_{2m})^T, W S ( ξ ) = e − 1 ξ T X , X = ( X 1 , … , X 2 m ) T ,
then, on the common invariant Schwartz core, (cf. [10 ] , Section 1.3) (81) gives
[ X i , X j ] = 2 − 1 S i j I . (82) [X_i,X_j]=2\sqrt{-1}S_{ij}I. \tag*{(82)} [ X i , X j ] = 2 − 1 S ij I . (82)
For the standard canonical observables ( Q 1 , P 1 , … , Q m , P m ) (Q_1,P_1,\ldots,Q_m,P_m) ( Q 1 , P 1 , … , Q m , P m ) ,
S = 1 2 ⨁ j = 1 m Ω , Ω = ( 0 1 − 1 0 ) . S=\frac{1}{2}\bigoplus_{j=1}^{m}\Omega,\qquad\Omega=\begin{pmatrix}0&1\\-1&0\end{pmatrix}. S = 2 1 j = 1 ⨁ m Ω , Ω = ( 0 − 1 1 0 ) .
In this representation CCR ( S ) = B ( L 2 ( R ) ⊗ m ) \operatorname{CCR}(S)=\mathcal{B}(L^2(\mathbb{R})^{\otimes m}) CCR ( S ) = B ( L 2 ( R ) ⊗ m ) .
The degenerate case. Suppose now that S S S is possibly singular, with
rank ( S ) = 2 k , r = d − 2 k . \operatorname{rank}(S)=2k,\qquad r=d-2k. rank ( S ) = 2 k , r = d − 2 k .
There exists an invertible real matrix T T T such that
T T S T = 0 r × r ⊕ 1 2 ⨁ j = 1 k Ω . (83) T^TST=0_{r\times r}\oplus\frac{1}{2}\bigoplus_{j=1}^{k}\Omega. \tag*{(83)} T T S T = 0 r × r ⊕ 2 1 j = 1 ⨁ k Ω. (83)
Let X j X_j X j be the self-adjoint generator determined by W S ( t e j ) = e − 1 t X j W_S(te_j)=e^{\sqrt{-1}tX_j} W S ( t e j ) = e − 1 t X j , and define the transformed observables by
X ~ = T T X , W X ~ ( η ) : = W S ( T η ) . \widetilde{X}=T^TX,\qquad W_{\widetilde{X}}(\eta):=W_S(T\eta). X = T T X , W X ( η ) := W S ( T η ) .
Their commutation matrix is S ~ = T T S T \widetilde{S}=T^TST S = T T S T . We realize them as
X ~ = ( Q ~ 1 , … , Q ~ r , Q 1 , P 1 , … , Q k , P k ) T , \widetilde{X}=(\widetilde{Q}_1,\ldots,\widetilde{Q}_r,Q_1,P_1,\ldots,Q_k,P_k)^T, X = ( Q 1 , … , Q r , Q 1 , P 1 , … , Q k , P k ) T ,
where the first r r r coordinates commute with all coordinates and ( Q j , P j ) (Q_j,P_j) ( Q j , P j ) are canonical pairs. The original coordinates are recovered as X = T − T X ~ X=T^{-T}\widetilde{X} X = T − T X ; in particular, X 1 X_1 X 1 will always refer to the first original coordinate.
To realize this system on an ordinary Hilbert space, introduce auxiliary conjugate observables P ~ 1 , … , P ~ r \widetilde{P}_1,\ldots,\widetilde{P}_r P 1 , … , P r and consider H ~ = L 2 ( R ) ⊗ ( r + k ) \widetilde{\mathcal{H}}=L^2(\mathbb{R})^{\otimes(r+k)} H = L 2 ( R ) ⊗ ( r + k ) . The enlarged collection
( Q ~ 1 , P ~ 1 , … , Q ~ r , P ~ r , Q 1 , P 1 , … , Q k , P k ) (\widetilde{Q}_1,\widetilde{P}_1,\ldots,\widetilde{Q}_r,\widetilde{P}_r,Q_1,P_1,\ldots,Q_k,P_k) ( Q 1 , P 1 , … , Q r , P r , Q 1 , P 1 , … , Q k , P k )
has a nondegenerate commutation matrix. We use the regular irreducible Schrödinger representation of this enlarged system on H ~ \widetilde{\mathcal{H}} H .
The algebra corresponding to the original degenerate system is the von Neumann subalgebra generated by
{ e − 1 t Q ~ j : t ∈ R , 1 ≤ j ≤ r } \{e^{\sqrt{-1}t\widetilde{Q}_j}:t\in\mathbb{R},\ 1\leq j\leq r\} { e − 1 t Q j : t ∈ R , 1 ≤ j ≤ r }
and by the Weyl operators of the remaining k k k canonical pairs. Consequently,
CCR ( S ) ≃ L ∞ ( R r ) ⊗ B ( L 2 ( R ) ⊗ k ) . (84) \operatorname{CCR}(S)\simeq L^\infty(\mathbb{R}^r)\otimes B\left(L^2(\mathbb{R})^{\otimes k}\right). \tag*{(84)} CCR ( S ) ≃ L ∞ ( R r ) ⊗ B ( L 2 ( R ) ⊗ k ) . (84)
Here L ∞ ( R r ) L^\infty(\mathbb{R}^r) L ∞ ( R r ) is taken with respect to Lebesgue measure. From now on, M = CCR ( S ) \mathcal{M}=\operatorname{CCR}(S) M = CCR ( S ) denotes this concrete von Neumann algebra, with the original Weyl operators W S ( ξ ) W_S(\xi) W S ( ξ ) . Zero-dimensional factors are omitted.
This is the von Neumann algebra counterpart of the classical–quantum representation used in Subsection 2.2. The auxiliary conjugate variables P ~ j \widetilde{P}_j P j do not belong to the original CCR ( S ) \operatorname{CCR}(S) CCR ( S ) ; they are introduced only to obtain a nondegenerate ambient representation. We refer the reader to [9 ] for further details on quantum-classical hybrid systems.
Gaussian states. As in the Gaussian testing model, assume that V V V is real symmetric and positive definite and that
Σ : = V + − 1 S ⪰ 0. \Sigma:=V+\sqrt{-1}S\succeq0. Σ := V + − 1 S ⪰ 0.
For every μ ∈ R d \mu\in\mathbb{R}^d μ ∈ R d , there is a unique normal state ϕ μ , Σ \phi_{\mu,\Sigma} ϕ μ , Σ on M \mathcal{M} M with the characteristic function below. We call it the Gaussian state with mean μ \mu μ and covariance Σ \Sigma Σ , and write ϕ = ϕ μ , Σ ∼ N ( μ , Σ ) \phi=\phi_{\mu,\Sigma}\sim N(\mu,\Sigma) ϕ = ϕ μ , Σ ∼ N ( μ , Σ ) .
ϕ ( W S ( ξ ) ) = exp ( − 1 ξ T μ − 1 2 ξ T V ξ ) , ξ ∈ R d . (85) \phi(W_S(\xi))=\exp\left(\sqrt{-1}\xi^{T}\mu-\frac{1}{2}\xi^{T}V\xi\right),\qquad\xi\in\mathbb{R}^d. \tag*{(85)} ϕ ( W S ( ξ )) = exp ( − 1 ξ T μ − 2 1 ξ T V ξ ) , ξ ∈ R d . (85)
The condition V ≻ 0 V\succ0 V ≻ 0 ensures a nondegenerate classical marginal and normality in the fixed Lebesgue representation. The matrix Σ \Sigma Σ itself is allowed to be singular. Uniqueness follows because the linear span of the Weyl operators is ultraweakly dense in M \mathcal{M} M .
When S S S is nonsingular this is the usual density-operator Gaussian state introduced in Subsection 2.2. When S S S is singular, the representation (84) identifies it with the corresponding classical–quantum Gaussian experiment.
For clarity, the tensor-product decomposition of M \mathcal{M} M does not require the state to be a product in the chosen coordinates. Under the change X ~ = T T X \widetilde{X}=T^{T}X X = T T X , the mean and covariance become
μ ~ = T T μ , V ~ = T T V T , Σ ~ = T T Σ T . \widetilde{\mu}=T^{T}\mu,\qquad\widetilde{V}=T^{T}VT,\qquad\widetilde{\Sigma}=T^{T}\Sigma T. μ = T T μ , V = T T V T , Σ = T T Σ T .
Write X ~ = ( C , R ) \widetilde{X}=(C,R) X = ( C , R ) , with classical coordinates C = ( Q ~ 1 , … , Q ~ r ) T C=(\widetilde{Q}_1,\ldots,\widetilde{Q}_r)^{T} C = ( Q 1 , … , Q r ) T and quantum coordinates R = ( Q 1 , P 1 , … , Q k , P k ) T R=(Q_1,P_1,\ldots,Q_k,P_k)^{T} R = ( Q 1 , P 1 , … , Q k , P k ) T . When both parts are present, partition
μ ~ = ( μ c μ q ) , Σ ~ = ( V c K K T V q + − 1 S q ) , S q = 1 2 ⨁ j = 1 k Ω . \widetilde{\mu}=\begin{pmatrix}\mu_c\\ \mu_q\end{pmatrix},\qquad
\widetilde{\Sigma}=
\begin{pmatrix}
V_c & K\\
K^T & V_q+\sqrt{-1}S_q
\end{pmatrix},\qquad
S_q=\frac{1}{2}\bigoplus_{j=1}^{k}\Omega. μ = ( μ c μ q ) , Σ = ( V c K T K V q + − 1 S q ) , S q = 2 1 j = 1 ⨁ k Ω.
Then V c ≻ 0 V_c \succ0 V c ≻ 0 and the Schur complement gives
Γ + − 1 S q ⪰ 0 , Γ : = V q − K T V c − 1 K . \Gamma+\sqrt{-1}S_q\succeq0,\qquad\Gamma:=V_q-K^TV_c^{-1}K. Γ + − 1 S q ⪰ 0 , Γ := V q − K T V c − 1 K .
Let f f f be the density of N ( μ c , V c ) N(\mu_c,V_c) N ( μ c , V c ) and let ρ c \rho_c ρ c be the usual quantum Gaussian density operator with mean μ q + K T V c − 1 ( c − μ c ) \mu_q+K^TV_c^{-1}(c-\mu_c) μ q + K T V c − 1 ( c − μ c ) and covariance Γ + − 1 S q \Gamma+\sqrt{-1}S_q Γ + − 1 S q . In the representation (84) , the state is
ϕ ( A ) = ∫ R r f ( c ) Tr ( ρ c A ( c ) ) d c , A ∈ M . \phi(A)=\int_{\mathbb{R}^r}f(c)\operatorname{Tr}(\rho_cA(c))\,\mathrm{d}c,\qquad A\in\mathcal{M}. ϕ ( A ) = ∫ R r f ( c ) Tr ( ρ c A ( c )) d c , A ∈ M .
Here A ( c ) A(c) A ( c ) is the operator-valued representative of A A A in the tensor-product algebra. This construction proves existence and normality and allows classical–quantum correlations. The further linear change
R ′ = R − K T V c − 1 C R'=R-K^TV_c^{-1}C R ′ = R − K T V c − 1 C
preserves the commutation relations and gives ( C , R ′ ) (C,R') ( C , R ′ ) the block-diagonal covariance V c ⊕ ( Γ + − 1 S q ) V_c\oplus(\Gamma+\sqrt{-1}S_q) V c ⊕ ( Γ + − 1 S q ) , hence a product Gaussian state. This is the product form used in Subsection 2.2. If r = 0 r=0 r = 0 or k = 0 k=0 k = 0 , the construction reduces to the usual quantum or classical Gaussian state, respectively.
For later use, the ordered products of Weyl operators satisfy
ϕ ( ∏ t = 1 q W S ( ξ t ) ) = exp [ ∑ t = 1 q ( − 1 ξ t T μ − 1 2 ξ t T Σ ξ t ) − ∑ 1 ≤ t < u ≤ q ξ t T Σ ξ u ] (86) \phi\left(\prod_{t=1}^{q}W_S(\xi_t)\right)
=\exp\left[\sum_{t=1}^{q}\left(\sqrt{-1}\xi_t^T\mu-\frac{1}{2}\xi_t^T\Sigma\xi_t\right)-\sum_{1\leq t<u\leq q}\xi_t^T\Sigma\xi_u\right] \tag*{(86)} ϕ ( t = 1 ∏ q W S ( ξ t ) ) = exp [ t = 1 ∑ q ( − 1 ξ t T μ − 2 1 ξ t T Σ ξ t ) − 1 ≤ t < u ≤ q ∑ ξ t T Σ ξ u ] (86)
Indeed, (81) gives
∏ t = 1 q W S ( ξ t ) = exp ( − − 1 ∑ t < u ξ t T S ξ u ) W S ( ∑ t = 1 q ξ t ) , \prod_{t=1}^{q}W_S(\xi_t)
=\exp\left(-\sqrt{-1}\sum_{t<u}\xi_t^TS\xi_u\right)W_S\left(\sum_{t=1}^{q}\xi_t\right), t = 1 ∏ q W S ( ξ t ) = exp ( − − 1 t < u ∑ ξ t T S ξ u ) W S ( t = 1 ∑ q ξ t ) ,
and (86) follows from (85) . Since ξ T S ξ = 0 \xi^TS\xi=0 ξ T S ξ = 0 , one may equivalently replace ξ t T Σ ξ t \xi_t^T\Sigma\xi_t ξ t T Σ ξ t in the first sum by ξ t T V ξ t \xi_t^TV\xi_t ξ t T V ξ t .
Translation of Gaussian mean
The enlarged representation also permits arbitrary translations of the Gaussian mean. For every a ∈ R d a\in\mathbb{R}^d a ∈ R d there exists a unitary U ( a ) U(a) U ( a ) on H ~ \widetilde{\mathcal{H}} H such that
U ( a ) ∗ W S ( ξ ) U ( a ) = e − 1 ξ T a W S ( ξ ) , ξ ∈ R d . (87) U(a)^*W_S(\xi)U(a)=e^{\sqrt{-1}\xi^Ta}W_S(\xi),\qquad\xi\in\mathbb{R}^d. \tag*{(87)} U ( a ) ∗ W S ( ξ ) U ( a ) = e − 1 ξ T a W S ( ξ ) , ξ ∈ R d . (87)
The unitaries U ( a ) U(a) U ( a ) can be chosen as a strongly continuous projective representation. Indeed, let ( E ~ , ω ~ ) (\widetilde{E},\widetilde{\omega}) ( E , ω ) be the ambient symplectic space and W ~ \widetilde{W} W its Weyl representation, with convention
W ~ ( v ) W ~ ( w ) = e − − 1 ω ~ ( v , w ) / 2 W ~ ( v + w ) . \widetilde{W}(v)\widetilde{W}(w)=e^{-\sqrt{-1}\widetilde{\omega}(v,w)/2}\widetilde{W}(v+w). W ( v ) W ( w ) = e − − 1 ω ( v , w ) /2 W ( v + w ) .
Identify the original coordinate space E = R d E=\mathbb{R}^d E = R d with its embedded subspace, so that W ~ ( ξ ) = W S ( ξ ) \widetilde{W}(\xi)=W_S(\xi) W ( ξ ) = W S ( ξ ) for ξ ∈ E \xi\in E ξ ∈ E . Nondegeneracy permits a linear choice of v ( a ) ∈ E ~ v(a)\in\widetilde{E} v ( a ) ∈ E such that
ω ~ ( v ( a ) , ξ ) = ξ T a , ξ ∈ E . \widetilde{\omega}(v(a),\xi)=\xi^Ta,\qquad\xi\in E. ω ( v ( a ) , ξ ) = ξ T a , ξ ∈ E .
Taking U ( a ) = W ~ ( v ( a ) ) U(a)=\widetilde{W}(v(a)) U ( a ) = W ( v ( a )) gives (87) . Since conjugation maps every Weyl generator to a scalar multiple of itself, it defines a ∗ * ∗ -automorphism
α a ( A ) : = U ( a ) ∗ A U ( a ) , A ∈ M . \alpha_a(A):=U(a)^*AU(a),\qquad A\in\mathcal{M}. α a ( A ) := U ( a ) ∗ A U ( a ) , A ∈ M .
These maps are ultraweakly continuous and satisfy
α 0 = id , α a ∘ α b = α a + b . \alpha_0=\operatorname{id},\qquad\alpha_a\circ\alpha_b=\alpha_{a+b}. α 0 = id , α a ∘ α b = α a + b .
The scalar phases in the projective representation disappear under conjugation. Strong continuity of U ( a ) U(a) U ( a ) also makes this action point-ultraweakly continuous. The implementing unitaries may lie outside M \mathcal{M} M .
Finally, we note that the characteristic function and normality give the covariance relation
ϕ μ , Σ ∘ α a = ϕ μ + a , Σ . (88) \phi_{\mu,\Sigma}\circ\alpha_a=\phi_{\mu+a,\Sigma}. \tag*{(88)} ϕ μ , Σ ∘ α a = ϕ μ + a , Σ . (88)
The following lemma identifies the invariant algebra used in the proof of Theorem 3.1 .
Lemma C.1 (Invariant algebra under nuisance translations). For the CCR representation M \mathcal{M} M and the displacement unitaries U ( a ) U(a) U ( a ) constructed above,
{ A ∈ M : U ( a ) ∗ A U ( a ) = A for all a ∈ R d with a 1 = 0 } = { e − − 1 t X 1 : t ∈ R } ′ ′ . \left\{A\in\mathcal{M}:U(a)^*AU(a)=A\text{ for all }a\in\mathbb{R}^d\text{ with }a_1=0\right\}
=\left\{e^{-\sqrt{-1}tX_1}:t\in\mathbb{R}\right\}^{\prime\prime}. { A ∈ M : U ( a ) ∗ A U ( a ) = A for all a ∈ R d with a 1 = 0 } = { e − − 1 t X 1 : t ∈ R } ′′ . In particular, a test T ∈ M T\in\mathcal{M} T ∈ M , 0 ≤ T ≤ I 0\leq T\leq I 0 ≤ T ≤ I , is invariant under these translations if and only if T = g ( X 1 ) T=g(X_1) T = g ( X 1 ) for some Borel function g : R → [ 0 , 1 ] g:\mathbb{R}\to[0,1] g : R → [ 0 , 1 ] .
Proof. Denote the algebra on the left by N \mathcal{N} N . In the ambient symplectic space introduced above, put
F : = { v ∈ E ~ : ω ~ ( v , e 1 ) = 0 } . F:=\{v\in\widetilde{E}:\widetilde{\omega}(v,e_1)=0\}. F := { v ∈ E : ω ( v , e 1 ) = 0 } .
For v ∈ F v\in F v ∈ F , define a ( v ) j : = ω ~ ( v , e j ) a(v)_j:=\widetilde{\omega}(v,e_j) a ( v ) j := ω ( v , e j ) . Then a ( v ) 1 = 0 a(v)_1=0 a ( v ) 1 = 0 , and the Weyl relations give
W ~ ( v ) ∗ W S ( ξ ) W ~ ( v ) = e − 1 ξ T a ( v ) W S ( ξ ) , ξ ∈ E . \widetilde{W}(v)^*W_S(\xi)\widetilde{W}(v)=e^{\sqrt{-1}\xi^{\mathsf T}a(v)}W_S(\xi),\qquad\xi\in E. W ( v ) ∗ W S ( ξ ) W ( v ) = e − 1 ξ T a ( v ) W S ( ξ ) , ξ ∈ E .
Thus conjugation by W ~ ( v ) \widetilde{W}(v) W ( v ) preserves M \mathcal{M} M and agrees there with conjugation by U ( a ( v ) ) U(a(v)) U ( a ( v )) : both maps are ultraweakly continuous and agree on the Weyl generators. Consequently,
N ⊆ { W ~ ( v ) : v ∈ F } ′ ′ . \mathcal{N}\subseteq\{\widetilde{W}(v):v\in F\}^{\prime\prime}. N ⊆ { W ( v ) : v ∈ F } ′′ .
Extend q 1 = e 1 q_1=e_1 q 1 = e 1 to a symplectic basis ( q 1 , p 1 , … , q n , p n ) (q_1,p_1,\ldots,q_n,p_n) ( q 1 , p 1 , … , q n , p n ) of E ~ \widetilde{E} E . This is possible because ω ~ \widetilde{\omega} ω is nondegenerate: choose p 1 p_1 p 1 with ω ~ ( e 1 , p 1 ) = 1 \widetilde{\omega}(e_1,p_1)=1 ω ( e 1 , p 1 ) = 1 and complete the basis in the symplectic orthogonal complement of their span. By the Stone–von Neumann theorem, we may work in the corresponding Schrödinger representation on L 2 ( R ) ⊗ H r e s t L^2(\mathbb{R})\otimes\mathcal{H}_{\mathrm{rest}} L 2 ( R ) ⊗ H rest , where H r e s t = L 2 ( R n − 1 ) \mathcal{H}_{\mathrm{rest}}=L^2(\mathbb{R}^{n-1}) H rest = L 2 ( R n − 1 ) and X 1 = Q 1 ⊗ I X_1=Q_1\otimes I X 1 = Q 1 ⊗ I . Since
F = span { q 1 , q 2 , p 2 , … , q n , p n } , F=\operatorname{span}\{q_1,q_2,p_2,\ldots,q_n,p_n\}, F = span { q 1 , q 2 , p 2 , … , q n , p n } ,
its Weyl operators generate
{ W ~ ( v ) : v ∈ F } ′ ′ = D ⊗ ‾ B ( H r e s t ) , \{\widetilde{W}(v):v\in F\}^{\prime\prime}=\mathcal{D}\,\overline{\otimes}\,\mathcal{B}(\mathcal{H}_{\mathrm{rest}}), { W ( v ) : v ∈ F } ′′ = D ⊗ B ( H rest ) ,
where D = L ∞ ( R ) \mathcal{D}=L^{\infty}(\mathbb{R}) D = L ∞ ( R ) acts by multiplication on L 2 ( R ) L^{2}(\mathbb{R}) L 2 ( R ) . The multiplication algebra is maximal abelian (cf. [25 ] Example 5.1.6,) so
{ W ~ ( v ) : v ∈ F } ′ = D ⊗ ‾ C I r e s t = { e − − 1 t X 1 : t ∈ R } ′ ′ . \{\widetilde{W}(v):v\in F\}'=\mathcal{D}\,\overline{\otimes}\,\mathbb{C}I_{\mathrm{rest}}=\{e^{-\sqrt{-1}tX_{1}}:t\in\mathbb{R}\}''. { W ( v ) : v ∈ F } ′ = D ⊗ C I rest = { e − − 1 t X 1 : t ∈ R } ′′ .
Here H r e s t = C \mathcal{H}_{\mathrm{rest}}=\mathbb{C} H rest = C when n = 1 n=1 n = 1 . This proves one inclusion. Conversely, if a 1 = 0 a_{1}=0 a 1 = 0 , conjugation by U ( a ) U(a) U ( a ) fixes every e − − 1 t X 1 e^{-\sqrt{-1}tX_{1}} e − − 1 t X 1 and hence their von Neumann algebra by ultraweak continuity. The assertion for tests follows from the spectral theorem, choosing a Borel representative with values in [ 0 , 1 ] [0,1] [ 0 , 1 ] .
The Quantum Hunt–Stein Theorem
This section establishes the nuisance-symmetry reduction used in § 3 . We treat Euclidean nuisance groups, which include all translations needed for the Gaussian testing problems. The formulation uses normal states on a von Neumann algebra and thus covers classical, quantum, and classical–quantum experiments. Kumagai and Hayashi [29 Theorem 3.2 and Remark 3.2] formulate a related noncompact reduction on B ( H ) \mathcal{B}(\mathcal{H}) B ( H ) . Since a hybrid Gaussian experiment may instead have a nontrivial classical center, we prove the required algebraic version directly.
Statistical model and tests. Let M \mathcal{M} M be a von Neumann algebra with identity I I I and separable predual M ∗ \mathcal{M}_{*} M ∗ . We use the duality M = ( M ∗ ) ∗ \mathcal{M}=(\mathcal{M}_{*})^{*} M = ( M ∗ ) ∗ ; its weak-star topology σ ( M , M ∗ ) \sigma(\mathcal{M},\mathcal{M}_{*}) σ ( M , M ∗ ) is the ultraweak topology. Thus A j → A A_{j}\to A A j → A ultraweakly means that ω ( A j ) → ω ( A ) \omega(A_{j})\to\omega(A) ω ( A j ) → ω ( A ) for every ω ∈ M ∗ \omega\in\mathcal{M}_{*} ω ∈ M ∗ . A normal state is a positive ultraweakly continuous linear functional φ \varphi φ with φ ( I ) = 1 \varphi(I)=1 φ ( I ) = 1 . Let
S = { φ θ , ξ : θ ∈ Θ , ξ ∈ Ξ } . \mathcal{S}=\{\varphi_{\theta,\xi}:\theta\in\Theta,\ \xi\in\Xi\}. S = { φ θ , ξ : θ ∈ Θ , ξ ∈ Ξ } .
be a family of normal states on M \mathcal{M} M , where θ \theta θ is the parameter of interest and ξ \xi ξ is nuisance. Suppose that Ξ \Xi Ξ is nonempty and that Θ 0 , Θ 1 \Theta_{0},\Theta_{1} Θ 0 , Θ 1 are disjoint nonempty sets with Θ = Θ 0 ∪ Θ 1 \Theta=\Theta_{0}\cup\Theta_{1} Θ = Θ 0 ∪ Θ 1 . Consider
H 0 : θ ∈ Θ 0 , H 1 : θ ∈ Θ 1 . (89) H_{0}:\theta\in\Theta_{0},\qquad H_{1}:\theta\in\Theta_{1}. \tag*{(89)} H 0 : θ ∈ Θ 0 , H 1 : θ ∈ Θ 1 . (89)
When M = B ( H ) \mathcal{M}=\mathcal{B}(\mathcal{H}) M = B ( H ) , this notation includes φ θ , ξ ( A ) = Tr ( ρ θ , ξ A ) \varphi_{\theta,\xi}(A)=\operatorname{Tr}(\rho_{\theta,\xi}A) φ θ , ξ ( A ) = Tr ( ρ θ , ξ A ) . For a hybrid experiment, φ θ , ξ \varphi_{\theta,\xi} φ θ , ξ is evaluated on its specified observable algebra.
A test is an effect T ∈ M T\in\mathcal{M} T ∈ M , meaning 0 ≤ T ≤ I 0\leq T\leq I 0 ≤ T ≤ I . Write π T ( θ , ξ ) : = φ θ , ξ ( T ) \pi_{T}(\theta,\xi):=\varphi_{\theta,\xi}(T) π T ( θ , ξ ) := φ θ , ξ ( T ) for its rejection probability. At an alternative θ ∈ Θ 1 \theta\in\Theta_{1} θ ∈ Θ 1 , its type II error probability is
β T ( φ θ , ξ ) : = 1 − φ θ , ξ ( T ) . \beta_{T}(\varphi_{\theta,\xi}):=1-\varphi_{\theta,\xi}(T). β T ( φ θ , ξ ) := 1 − φ θ , ξ ( T ) .
Fix α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . Define
T α : = { T ∈ M : 0 ≤ T ≤ I , φ θ , ξ ( T ) ≤ α for all ( θ , ξ ) ∈ Θ 0 × Ξ } , \mathcal{T}_{\alpha}:=\{T\in\mathcal{M}:0\leq T\leq I,\ \varphi_{\theta,\xi}(T)\leq\alpha\ \text{for all }(\theta,\xi)\in\Theta_{0}\times\Xi\}, T α := { T ∈ M : 0 ≤ T ≤ I , φ θ , ξ ( T ) ≤ α for all ( θ , ξ ) ∈ Θ 0 × Ξ } ,
T α , u : = { T ∈ T α : φ θ , ξ ( T ) ≥ α for all ( θ , ξ ) ∈ Θ 1 × Ξ } . \mathcal{T}_{\alpha,u}:=\{T\in\mathcal{T}_{\alpha}:\varphi_{\theta,\xi}(T)\geq\alpha\ \text{for all }(\theta,\xi)\in\Theta_{1}\times\Xi\}. T α , u := { T ∈ T α : φ θ , ξ ( T ) ≥ α for all ( θ , ξ ) ∈ Θ 1 × Ξ } .
Thus unbiasedness here means that power is at least the prescribed level α \alpha α throughout the alternative. For θ ∈ Θ 1 \theta\in\Theta_{1} θ ∈ Θ 1 , the minimum nuisance power is
π T ( θ ) : = inf ξ ∈ Ξ φ θ , ξ ( T ) . \pi_{T}(\theta):=\inf_{\xi\in\Xi}\varphi_{\theta,\xi}(T). π T ( θ ) := ξ ∈ Ξ inf φ θ , ξ ( T ) .
Maximizing this quantity is equivalent to minimizing sup ξ ∈ Ξ β T ( φ θ , ξ ) = 1 − π T ( θ ) \sup_{\xi\in\Xi}\beta_{T}(\varphi_{\theta,\xi})=1-\pi_{T}(\theta) sup ξ ∈ Ξ β T ( φ θ , ξ ) = 1 − π T ( θ ) .
Covariant nuisance action. Let G = ( R q , + ) G = (\mathbb{R}^{q},+) G = ( R q , + ) , where q ≥ 0 q \ge0 q ≥ 0 , act on Ξ \Xi Ξ , with action denoted by t ∘ ξ t \circ\xi t ∘ ξ . Suppose that { τ t : t ∈ R q } \{\tau_t:t \in\mathbb{R}^{q}\} { τ t : t ∈ R q } is an action on M \mathcal{M} M by normal ∗ * ∗ -automorphisms:
τ 0 = id , τ s ∘ τ t = τ s + t . \tau_0 = \operatorname{id}, \qquad\tau_s \circ\tau_t = \tau_{s+t}. τ 0 = id , τ s ∘ τ t = τ s + t .
Assume that the action is point-ultraweakly continuous: for every A ∈ M A \in\mathcal{M} A ∈ M and ω ∈ M ∗ \omega\in\mathcal{M}_* ω ∈ M ∗ , the function t ↦ ω ( τ t ( A ) ) t \mapsto\omega(\tau_t(A)) t ↦ ω ( τ t ( A )) is continuous. We assume that the action is covariant, i.e.
ϕ θ , ξ ( τ t ( A ) ) = ϕ θ , t ∘ ξ ( A ) , A ∈ M , θ ∈ Θ , ξ ∈ Ξ , t ∈ R q . (90) \phi_{\theta,\xi}(\tau_t(A)) = \phi_{\theta,t\circ\xi}(A), \qquad A \in\mathcal{M}, \quad\theta\in\Theta, \quad\xi\in\Xi, \quad t \in\mathbb{R}^{q}. \tag*{(90)} ϕ θ , ξ ( τ t ( A )) = ϕ θ , t ∘ ξ ( A ) , A ∈ M , θ ∈ Θ , ξ ∈ Ξ , t ∈ R q . (90)
Only the nuisance parameter is changed by this action. For example, if τ t ( A ) = V t ∗ A V t \tau_t(A) = V_t^* A V_t τ t ( A ) = V t ∗ A V t for a strongly continuous unitary or projective representation on a Hilbert space, then (90) is the usual relation V t ρ θ , ξ V t ∗ = ρ θ , t ∘ ξ V_t\rho_{\theta,\xi}V_t^* = \rho_{\theta,t\circ\xi} V t ρ θ , ξ V t ∗ = ρ θ , t ∘ ξ .
Next we define the class of invariant elements with respect to this action.
M G : = { A ∈ M : τ t ( A ) = A for every t ∈ R q } , \mathcal{M}^{G} := \{A \in\mathcal{M}:\tau_t(A) = A \text{ for every } t \in\mathbb{R}^{q}\}, M G := { A ∈ M : τ t ( A ) = A for every t ∈ R q } ,
This naturally gives us the definition of invariant tests, and invariant unbiased tests.
T α , G : = T α ∩ M G , T α , u , G : = T α , u ∩ M G . \mathcal{T}_{\alpha,G} := \mathcal{T}_{\alpha} \cap\mathcal{M}^{G}, \qquad\mathcal{T}_{\alpha,u,G} := \mathcal{T}_{\alpha,u} \cap\mathcal{M}^{G}. T α , G := T α ∩ M G , T α , u , G := T α , u ∩ M G .
Theorem D.1 (Quantum Hunt–Stein reduction for Euclidean nuisance groups). Under the preceding assumptions, for every effect T ∈ M T \in\mathcal{M} T ∈ M there exists an invariant effect T ‾ ∈ M G \overline{T} \in\mathcal{M}^{G} T ∈ M G such that
inf η ∈ Ξ ϕ θ , η ( T ) ≤ ϕ θ , ξ ( T ‾ ) ≤ sup η ∈ Ξ ϕ θ , η ( T ) , θ ∈ Θ , ξ ∈ Ξ . (91) \inf_{\eta\in\Xi}\phi_{\theta,\eta}(T) \le\phi_{\theta,\xi}(\overline{T}) \le\sup_{\eta\in\Xi}\phi_{\theta,\eta}(T), \qquad\theta\in\Theta,\quad\xi\in\Xi. \tag*{(91)} η ∈ Ξ inf ϕ θ , η ( T ) ≤ ϕ θ , ξ ( T ) ≤ η ∈ Ξ sup ϕ θ , η ( T ) , θ ∈ Θ , ξ ∈ Ξ. (91) In particular, symmetrization preserves level and level-α \alpha α unbiasedness, and
π T ‾ ( θ ) ≥ π T ( θ ) , θ ∈ Θ 1 . \pi_{\overline{T}}(\theta) \ge\pi_T(\theta), \qquad\theta\in\Theta_1. π T ( θ ) ≥ π T ( θ ) , θ ∈ Θ 1 . Consequently, for every θ ∈ Θ 1 \theta\in\Theta_1 θ ∈ Θ 1 ,
inf T ∈ T α , G sup ξ ∈ Ξ β T ( ϕ θ , ξ ) = inf T ∈ T α sup ξ ∈ Ξ β T ( ϕ θ , ξ ) . (92) \inf_{T \in\mathcal{T}_{\alpha,G}}\sup_{\xi\in\Xi}\beta_T(\phi_{\theta,\xi}) = \inf_{T \in\mathcal{T}_{\alpha}}\sup_{\xi\in\Xi}\beta_T(\phi_{\theta,\xi}). \tag*{(92)} T ∈ T α , G inf ξ ∈ Ξ sup β T ( ϕ θ , ξ ) = T ∈ T α inf ξ ∈ Ξ sup β T ( ϕ θ , ξ ) . (92) inf T ∈ T α , u , G sup ξ ∈ Ξ β T ( ϕ θ , ξ ) = inf T ∈ T α , u sup ξ ∈ Ξ β T ( ϕ θ , ξ ) . (93) \inf_{T \in\mathcal{T}_{\alpha,u,G}}\sup_{\xi\in\Xi}\beta_T(\phi_{\theta,\xi}) = \inf_{T \in\mathcal{T}_{\alpha,u}}\sup_{\xi\in\Xi}\beta_T(\phi_{\theta,\xi}). \tag*{(93)} T ∈ T α , u , G inf ξ ∈ Ξ sup β T ( ϕ θ , ξ ) = T ∈ T α , u inf ξ ∈ Ξ sup β T ( ϕ θ , ξ ) . (93)
Proof. If q = 0 q = 0 q = 0 , the group is trivial and we take T ‾ = T \overline{T} = T T = T . Assume henceforth that q ≥ 1 q \ge1 q ≥ 1 .
Construction of the averages. For R > 0 R > 0 R > 0 , let F R = [ − R , R ] q F_R = [-R,R]^q F R = [ − R , R ] q , and write ∣ E ∣ |E| ∣ E ∣ for the Lebesgue measure of a measurable set E ⊂ R q E \subset\mathbb{R}^{q} E ⊂ R q . Define
T R : = 1 ∣ F R ∣ ∫ F R τ t ( T ) d t , (94) T_R := \frac{1}{|F_R|}\int_{F_R}\tau_t(T)\,\mathrm{d}t, \tag*{(94)} T R := ∣ F R ∣ 1 ∫ F R τ t ( T ) d t , (94)
where the integral is defined in the ultraweak sense. More explicitly, T R T_R T R is determined by
ω ( T R ) = 1 ∣ F R ∣ ∫ F R ω ( τ t ( T ) ) d t , ω ∈ M ∗ . (95) \omega(T_R) = \frac{1}{|F_R|}\int_{F_R}\omega(\tau_t(T))\,\mathrm{d}t, \qquad\omega\in\mathcal{M}_*. \tag*{(95)} ω ( T R ) = ∣ F R ∣ 1 ∫ F R ω ( τ t ( T )) d t , ω ∈ M ∗ . (95)
The scalar integrand is continuous. Since ∗ * ∗ -automorphisms are isometric, the right-hand side is a linear functional of ω \omega ω with absolute value at most ∥ ω ∥ ∥ T ∥ \|\omega\|\,\|T\| ∥ ω ∥ ∥ T ∥ . The identity M = ( M ∗ ) ∗ \mathcal{M} = (\mathcal{M}_*)^* M = ( M ∗ ) ∗ therefore gives a unique T R ∈ M T_R \in\mathcal{M} T R ∈ M . Moreover, 0 ≤ τ t ( T ) ≤ I 0 \le\tau_t(T) \le I 0 ≤ τ t ( T ) ≤ I for every t t t . Evaluating (95) on positive normal functionals gives 0 ≤ ω ( T R ) ≤ ω ( I ) 0 \le\omega(T_R) \le\omega(I) 0 ≤ ω ( T R ) ≤ ω ( I ) , and hence 0 ≤ T R ≤ I 0 \le T_R \le I 0 ≤ T R ≤ I .
Using the covariance property of the action (90) yields
ϕ θ , ξ ( T R ) = 1 ∣ F R ∣ ∫ F R ϕ θ , t ∘ ξ ( T ) d t . (96) \phi_{\theta,\xi}(T_R) = \frac{1}{|F_R|} \int_{F_R} \phi_{\theta,t \circ\xi}(T)\,\mathrm{d}t. \tag*{(96)} ϕ θ , ξ ( T R ) = ∣ F R ∣ 1 ∫ F R ϕ θ , t ∘ ξ ( T ) d t . (96)
Consequently,
inf η ∈ Ξ ϕ θ , η ( T ) ≤ ϕ θ , ξ ( T R ) ≤ sup η ∈ Ξ ϕ θ , η ( T ) (97) \inf_{\eta\in\Xi} \phi_{\theta,\eta}(T) \le\phi_{\theta,\xi}(T_R) \le\sup_{\eta\in\Xi} \phi_{\theta,\eta}(T) \tag*{(97)} η ∈ Ξ inf ϕ θ , η ( T ) ≤ ϕ θ , ξ ( T R ) ≤ η ∈ Ξ sup ϕ θ , η ( T ) (97)
for every R R R , θ \theta θ , and ξ \xi ξ .
Approximate invariance. Fix s ∈ R q s \in\mathbb{R}^{q} s ∈ R q . Normality of τ s \tau_s τ s means that ω ∘ τ s ∈ M ∗ \omega\circ\tau_s \in\mathcal{M}_* ω ∘ τ s ∈ M ∗ whenever ω ∈ M ∗ \omega\in\mathcal{M}_* ω ∈ M ∗ . Thus (95) , and a change of variables imply
τ s ( T R ) = 1 ∣ F R ∣ ∫ F R τ s + t ( T ) d t = 1 ∣ F R ∣ ∫ F R + s τ t ( T ) d t . \tau_s(T_R) = \frac{1}{|F_R|} \int_{F_R} \tau_{s+t}(T)\,\mathrm{d}t = \frac{1}{|F_R|} \int_{F_R+s} \tau_t(T)\,\mathrm{d}t. τ s ( T R ) = ∣ F R ∣ 1 ∫ F R τ s + t ( T ) d t = ∣ F R ∣ 1 ∫ F R + s τ t ( T ) d t .
Subtracting (94) , the integrals over F R ∩ ( F R + s ) F_R \cap(F_R+s) F R ∩ ( F R + s ) cancel. Therefore
∣ ω ( τ s ( T R ) − T R ) ∣ ≤ ∥ ω ∥ ∥ T ∥ ∣ ( F R + s ) △ F R ∣ ∣ F R ∣ , ω ∈ M ∗ , \left|\omega\left(\tau_s(T_R)-T_R\right)\right| \le\|\omega\|\,\|T\|\,\frac{|(F_R+s)\mathbin{\triangle}F_R|}{|F_R|}, \qquad\omega\in\mathcal{M}_*, ∣ ω ( τ s ( T R ) − T R ) ∣ ≤ ∥ ω ∥ ∥ T ∥ ∣ F R ∣ ∣ ( F R + s ) △ F R ∣ , ω ∈ M ∗ ,
where △ \triangle △ denotes symmetric difference. Taking the supremum over ∥ ω ∥ ≤ 1 \|\omega\| \le1 ∥ ω ∥ ≤ 1 gives
∥ τ s ( T R ) − T R ∥ ≤ ∥ T ∥ ∣ ( F R + s ) △ F R ∣ ∣ F R ∣ . (98) \|\tau_s(T_R)-T_R\| \le\|T\|\frac{|(F_R+s)\mathbin{\triangle}F_R|}{|F_R|}. \tag*{(98)} ∥ τ s ( T R ) − T R ∥ ≤ ∥ T ∥ ∣ F R ∣ ∣ ( F R + s ) △ F R ∣ . (98)
For R > 1 2 max j ∣ s j ∣ R > \frac{1}{2}\max_j |s_j| R > 2 1 max j ∣ s j ∣ , the intersection of the two cubes has volume ∏ j = 1 q ( 2 R − ∣ s j ∣ ) \prod_{j=1}^{q}(2R-|s_j|) ∏ j = 1 q ( 2 R − ∣ s j ∣ ) , so
∣ ( F R + s ) △ F R ∣ ∣ F R ∣ = 2 { 1 − ∏ j = 1 q ( 1 − ∣ s j ∣ 2 R ) } ≤ 1 R ∑ j = 1 q ∣ s j ∣ → 0. (99) \frac{|(F_R+s)\mathbin{\triangle}F_R|}{|F_R|} = 2\left\{1-\prod_{j=1}^{q}\left(1-\frac{|s_j|}{2R}\right)\right\} \le\frac{1}{R}\sum_{j=1}^{q}|s_j| \to0. \tag*{(99)} ∣ F R ∣ ∣ ( F R + s ) △ F R ∣ = 2 { 1 − j = 1 ∏ q ( 1 − 2 R ∣ s j ∣ ) } ≤ R 1 j = 1 ∑ q ∣ s j ∣ → 0. (99)
This is the Følner property of the cubes. In particular, ∥ τ s ( T R ) − T R ∥ → 0 \|\tau_s(T_R)-T_R\| \to0 ∥ τ s ( T R ) − T R ∥ → 0 for every fixed s s s .
Compactness and an invariant limit. The effect interval [ 0 , I ] M [0,I]_{\mathcal{M}} [ 0 , I ] M is ultraweakly compact. Indeed, the closed unit ball is weak-star compact by the Banach–Alaoglu theorem, and [ 0 , I ] M [0,I]_{\mathcal{M}} [ 0 , I ] M is weak-star closed: its defining inequalities can be tested against positive normal functionals. Since M ∗ \mathcal{M}_* M ∗ is separable, this compact set is metrizable and therefore sequentially compact; see also [42 Chapter III] for the predual framework. Taking R = 1 , 2 , … R=1,2,\ldots R = 1 , 2 , … , choose a subsequence R j → ∞ R_j \to\infty R j → ∞ and an effect T ‾ ∈ M \overline{T} \in\mathcal{M} T ∈ M such that
T R j → T ‾ in σ ( M , M ∗ ) . (100) T_{R_j} \to\overline{T}\quad\text{in }\sigma(\mathcal{M},\mathcal{M}_*). \tag*{(100)} T R j → T in σ ( M , M ∗ ) . (100)
For fixed s ∈ R q s \in\mathbb{R}^{q} s ∈ R q and ω ∈ M ∗ \omega\in\mathcal{M}_* ω ∈ M ∗ , normality of τ s \tau_s τ s , (100) , and (98) give
ω ( τ s ( T ‾ ) − T ‾ ) = lim j → ∞ ω ( τ s ( T R j ) − T R j ) = 0. \omega\left(\tau_s(\overline{T})-\overline{T}\right)=\lim_{j\to\infty}\omega\left(\tau_s(T_{R_j})-T_{R_j}\right)=0. ω ( τ s ( T ) − T ) = j → ∞ lim ω ( τ s ( T R j ) − T R j ) = 0.
Because the predual separates points of M \mathcal{M} M , τ s ( T ‾ ) = T ‾ \tau_s(\overline{T})=\overline{T} τ s ( T ) = T . The element s s s was arbitrary, so T ‾ ∈ M G \overline{T} \in\mathcal{M}^{G} T ∈ M G .
Level, unbiasedness, and minimax values. Every ϕ θ , ξ \phi_{\theta,\xi} ϕ θ , ξ is normal. Passing to the limit in (97) proves (91) simultaneously for all θ ∈ Θ \theta\in\Theta θ ∈ Θ and ξ ∈ Ξ \xi\in\Xi ξ ∈ Ξ . If T ∈ T α T \in\mathcal{T}_{\alpha} T ∈ T α , its upper bound is at most α \alpha α on Θ 0 \Theta_0 Θ 0 , so T ‾ ∈ T α , G \overline{T} \in\mathcal{T}_{\alpha,G} T ∈ T α , G . If T ∈ T α , u T \in\mathcal{T}_{\alpha,u} T ∈ T α , u , its lower bound is at least α \alpha α on Θ 1 \Theta_1 Θ 1 , giving T ‾ ∈ T α , u , G \overline{T} \in\mathcal{T}_{\alpha,u,G} T ∈ T α , u , G as well. For every θ ∈ Θ 1 \theta\in\Theta_1 θ ∈ Θ 1 ,
sup ξ ∈ Ξ β T ‾ ( ϕ θ , ξ ) ≤ sup ξ ∈ Ξ β T ( ϕ θ , ξ ) . \sup_{\xi\in\Xi} \beta_{\overline{T}}(\phi_{\theta,\xi}) \le\sup_{\xi\in\Xi} \beta_T(\phi_{\theta,\xi}). ξ ∈ Ξ sup β T ( ϕ θ , ξ ) ≤ ξ ∈ Ξ sup β T ( ϕ θ , ξ ) .
Taking the infimum over either admissible class proves one direction of (92)–(93); the reverse direction follows because the invariant class is a subclass of the original class.
The proof requires neither completeness of the statistical family nor a Gaussian assumption. It selects an invariant cluster point for each test; it does not assert convergence of the full family of averages or construct a normal conditional expectation onto M G \mathcal{M}^G M G .
Corollary D.2 (Optimal invariant tests are nuisance-maximin). Assume the hypotheses of Theorem D.1 . Let C \mathcal{C} C be either T α \mathcal{T}_{\alpha} T α or T α , u \mathcal{T}_{\alpha,u} T α , u , and put C G = C ∩ M G \mathcal{C}_G = \mathcal{C} \cap\mathcal{M}^G C G = C ∩ M G . Suppose that T 0 ∈ C G T_0 \in\mathcal{C}_G T 0 ∈ C G is uniformly most powerful within C G \mathcal{C}_G C G , in the sense that
ϕ θ , ξ ( T 0 ) ≥ ϕ θ , ξ ( A ) , A ∈ C G , θ ∈ Θ 1 , ξ ∈ Ξ . \phi_{\theta,\xi}(T_0) \ge\phi_{\theta,\xi}(A), \qquad A \in\mathcal{C}_G,\quad\theta\in\Theta_1,\quad\xi\in\Xi. ϕ θ , ξ ( T 0 ) ≥ ϕ θ , ξ ( A ) , A ∈ C G , θ ∈ Θ 1 , ξ ∈ Ξ. Then
π T 0 ( θ ) = sup T ∈ C π T ( θ ) , θ ∈ Θ 1 . (101) \pi_{T_0}(\theta) = \sup_{T \in\mathcal{C}} \pi_T(\theta), \qquad\theta\in\Theta_1. \tag*{(101)} π T 0 ( θ ) = T ∈ C sup π T ( θ ) , θ ∈ Θ 1 . (101) Thus T 0 T_0 T 0 is uniformly maximin over the nuisance parameter in the same admissible class, including the unbiased class when it is used.
Proof. For T ∈ C T \in\mathcal{C} T ∈ C , let T ‾ ∈ C G \overline{T} \in\mathcal{C}_G T ∈ C G be supplied by Theorem D.1 . The assumed invariant optimality gives
π T 0 ( θ ) ≥ π T ‾ ( θ ) ≥ π T ( θ ) , θ ∈ Θ 1 . \pi_{T_0}(\theta) \ge\pi_{\overline{T}}(\theta) \ge\pi_T(\theta), \qquad\theta\in\Theta_1. π T 0 ( θ ) ≥ π T ( θ ) ≥ π T ( θ ) , θ ∈ Θ 1 .
Taking the supremum over T ∈ C T \in\mathcal{C} T ∈ C , and using T 0 ∈ C T_0 \in\mathcal{C} T 0 ∈ C , proves (101) .
Application to the Gaussian nuisance translations. In § 3 , take M = CCR ( S ) \mathcal{M} = \operatorname{CCR}(S) M = CCR ( S ) and the normal Gaussian states ϕ y , z ∼ N ( ( y , z ) , Σ ) \phi_{y,z} \sim\mathcal{N}((y,z), \Sigma) ϕ y , z ∼ N (( y , z ) , Σ ) , where Σ = V + i S ⪰ 0 \Sigma= V + iS \succeq0 Σ = V + i S ⪰ 0 , V ≻ 0 V \succ0 V ≻ 0 , y ∈ R y \in\mathbb{R} y ∈ R , and z ∈ R d − 1 z \in\mathbb{R}^{d-1} z ∈ R d − 1 . No nonsingularity assumption on S S S is required. If rank ( S ) = 2 k \operatorname{rank}(S) = 2k rank ( S ) = 2 k and r = d − 2 k r = d - 2k r = d − 2 k , the concrete representation described above gives
M ≃ L ∞ ( R r ) ⊗ ‾ B ( L 2 ( R k ) ) . \mathcal{M} \simeq L^{\infty}(\mathbb{R}^{r}) \mathbin{\overline{\otimes}} \mathcal{B}(L^{2}(\mathbb{R}^{k})). M ≃ L ∞ ( R r ) ⊗ B ( L 2 ( R k )) .
Here L 2 ( R 0 ) = C L^{2}(\mathbb{R}^{0}) = \mathbb{C} L 2 ( R 0 ) = C when k = 0 k = 0 k = 0 . This algebra acts on the separable Hilbert space H ~ = L 2 ( R r + k ) \widetilde{\mathcal{H}} = L^{2}(\mathbb{R}^{r+k}) H = L 2 ( R r + k ) and has separable predual: the predual is a quotient of T 1 ( H ~ ) \mathcal{T}_{1}(\widetilde{\mathcal{H}}) T 1 ( H ) under restriction of normal functionals, and the latter trace-class space is separable.
For t ∈ R d − 1 t \in\mathbb{R}^{d-1} t ∈ R d − 1 , set τ t : = α ( 0 , t ) \tau_t := \alpha_{(0,t)} τ t := α ( 0 , t ) , with α a \alpha_a α a as in (87) . Thus
τ t ( W S ( ξ ) ) = e i ξ T ( 0 , t ) W S ( ξ ) . \tau_t(W_S(\xi)) = e^{i\xi^{T}(0,t)}W_S(\xi). τ t ( W S ( ξ )) = e i ξ T ( 0 , t ) W S ( ξ ) .
These are normal automorphisms. Their implementations in the enlarged Schrödinger representation may be chosen as a strongly continuous family of Weyl displacement unitaries U t U_t U t . In particular, for a trace-class operator K K K and A ∈ M A \in\mathcal{M} A ∈ M ,
∣ Tr { K ( U t ∗ A U t − U s ∗ A U s ) } ∣ ≤ ∥ U t K U t ∗ − U s K U s ∗ ∥ 1 ∥ A ∥ → 0 ( t → s ) . \left|\operatorname{Tr}\{K(U_t^* A U_t - U_s^* A U_s)\}\right| \le\|U_t K U_t^* - U_s K U_s^*\|_1 \|A\| \to0 \qquad(t \to s). ∣ Tr { K ( U t ∗ A U t − U s ∗ A U s )} ∣ ≤ ∥ U t K U t ∗ − U s K U s ∗ ∥ 1 ∥ A ∥ → 0 ( t → s ) .
The trace-norm convergence follows first for finite-rank K K K from strong continuity and then for arbitrary K K K by trace-norm approximation. Restricting these normal functionals to M \mathcal{M} M verifies point-ultraweak continuity. The unitaries U t U_t U t may lie outside M \mathcal{M} M ; only the induced automorphisms of M \mathcal{M} M are used.
The Gaussian characteristic function gives
ϕ y , z ( τ t ( W S ( ξ ) ) ) = exp { i ξ T ( y , z + t ) − 1 2 ξ T V ξ } = ϕ y , z + t ( W S ( ξ ) ) . \phi_{y,z}(\tau_t(W_S(\xi)))=\exp\left\{i\xi^{T}(y,z+t)-\frac{1}{2}\xi^{T}V\xi\right\}=\phi_{y,z+t}(W_S(\xi)). ϕ y , z ( τ t ( W S ( ξ ))) = exp { i ξ T ( y , z + t ) − 2 1 ξ T V ξ } = ϕ y , z + t ( W S ( ξ )) .
Both sides define normal functionals of the observable. Since the linear span of the Weyl operators is ultraweakly dense in M \mathcal{M} M , this proves
ϕ y , z ( τ t ( A ) ) = ϕ y , z + t ( A ) , A ∈ M , (102) \phi_{y,z}(\tau_t(A))=\phi_{y,z+t}(A),\qquad A\in\mathcal{M}, \tag*{(102)} ϕ y , z ( τ t ( A )) = ϕ y , z + t ( A ) , A ∈ M , (102)
which is (90). Hence all hypotheses of Theorem D.1 are satisfied with q = d − 1 q=d-1 q = d − 1 and Ξ = R d − 1 \Xi=\mathbb{R}^{d-1} Ξ = R d − 1 . For d = 1 d=1 d = 1 , the nuisance group is trivial.
For one-sided testing use Θ 0 = ( − ∞ , 0 ] \Theta_0=(-\infty,0] Θ 0 = ( − ∞ , 0 ] and Θ 1 = ( 0 , ∞ ) \Theta_1=(0,\infty) Θ 1 = ( 0 , ∞ ) ; for two-sided testing use Θ 0 = { 0 } \Theta_0=\{0\} Θ 0 = { 0 } and Θ 1 = R ∖ { 0 } \Theta_1=\mathbb{R}\setminus\{0\} Θ 1 = R ∖ { 0 } , together with the unbiased class. The nuisance translations are transitive on Ξ \Xi Ξ , so invariant power is independent of z z z . For a general nonzero contrast c T h c^{T}h c T h , use the group ker ( c T ) \operatorname{ker}(c^{T}) ker ( c T ) , identified linearly with R d − 1 \mathbb{R}^{d-1} R d − 1 .
This establishes the symmetrization step required in the Gaussian testing proofs. The form of the invariant effects is supplied separately by Lemma C.1 ; no fixed-point identification is inferred from the averages themselves. The conclusion concerns minimum power over nuisance parameters, not pointwise domination of every competing test at each nuisance value.
Proofs of the main theorems
Proof of Theorem 2.2. We follow the proof of the asymptotic quantum representation theorem in [14 Theorem 2.4] . The hypothesis that L h ( n ) ⇒ L h \mathcal{L}^{(n)}_h\Rightarrow\mathcal{L}_h L h ( n ) ⇒ L h is not used in Steps 1 and 2 of that proof, nor in Step 3 before the final paragraph beginning “Fix h ∈ R d h\in\mathbb{R}^{d} h ∈ R d arbitrarily.” Consequently, the same argument up to that point gives the following objects:
a subsequence { n k } k ≥ 1 \{n_k\}_{k\geq1} { n k } k ≥ 1 , chosen independently of h h h ;
a coordinatewise nondecreasing family { M ‾ α : α ∈ Q s } ⊂ CCR ( Im Σ ) + \{\overline{M}_{\alpha}:\alpha\in\mathbb{Q}^{s}\}\subset\operatorname{CCR}(\operatorname{Im}\Sigma)_{+} { M α : α ∈ Q s } ⊂ CCR ( Im Σ ) + , with 0 ≤ M ‾ α ≤ I 0\leq\overline{M}_{\alpha}\leq I 0 ≤ M α ≤ I ;
a POVM M ‾ \overline{M} M on R ‾ s \overline{\mathbb{R}}^{s} R s , determined by
M ‾ t : = M ‾ ( ( − ∞ , t ] ) = inf α ∈ Q s α > t M ‾ α , t ∈ R s , (103) \overline{M}_t:=\overline{M}((-\infty,t])=\inf_{\substack{\alpha\in\mathbb{Q}^{s}\\ \alpha>t}}\overline{M}_{\alpha},\qquad t\in\mathbb{R}^{s}, \tag*{(103)} M t := M (( − ∞ , t ]) = α ∈ Q s α > t inf M α , t ∈ R s , (103)
where the infimum is taken in the weak operator topology; and
the POVM on R s \mathbb{R}^{s} R s defined by
M ( ∞ ) ( B ) : = M ‾ ( B ) + δ 0 ( B ) M ‾ ( R ‾ s ∖ R s ) , B ∈ B ( R s ) . (104) M^{(\infty)}(B):=\overline{M}(B)+\delta_{0}(B)\overline{M}(\overline{\mathbb{R}}^{s}\setminus\mathbb{R}^{s}),\qquad B\in\mathcal{B}(\mathbb{R}^{s}). \tag*{(104)} M ( ∞ ) ( B ) := M ( B ) + δ 0 ( B ) M ( R s ∖ R s ) , B ∈ B ( R s ) . (104)
Moreover, the sandwiched Le Cam third lemma, exactly as used in Step 3 of the cited proof, gives
lim k → ∞ L h ( n k ) ( ( − ∞ , α ] ) = ϱ h ( M ‾ α ) , h ∈ R d , α ∈ Q s . (105) \lim_{k\to\infty}\mathcal{L}^{(n_k)}_h((-\infty,\alpha])=\varrho_h(\overline{M}_{\alpha}),\qquad h\in\mathbb{R}^{d},\quad\alpha\in\mathbb{Q}^{s}. \tag*{(105)} k → ∞ lim L h ( n k ) (( − ∞ , α ]) = ϱ h ( M α ) , h ∈ R d , α ∈ Q s . (105)
Fix h ∈ R d h \in\mathbb{R}^{d} h ∈ R d , and define
μ h ( B ) : = ϱ h ( M ‾ ( ∞ ) ( B ) ) , B ∈ B ( R s ) . \mu_{h}(B) := \varrho_{h}\left(\overline{M}^{(\infty)}(B)\right), \qquad B \in\mathcal{B}(\mathbb{R}^{s}). μ h ( B ) := ϱ h ( M ( ∞ ) ( B ) ) , B ∈ B ( R s ) .
We prove that L h ( n k ) ⇒ μ h \mathcal{L}_{h}^{(n_{k})} \Rightarrow\mu_{h} L h ( n k ) ⇒ μ h . For α ∈ Q s \alpha\in\mathbb{Q}^{s} α ∈ Q s , set
a h ( α ) : = ϱ h ( M ~ α ) . a_{h}(\alpha) := \varrho_{h}\left(\widetilde{M}_{\alpha}\right). a h ( α ) := ϱ h ( M α ) .
We first show that M ‾ \overline{M} M places no mass at infinity under the state ϱ h \varrho_{h} ϱ h . Let ε > 0 \varepsilon> 0 ε > 0 . By tightness of { L h ( n ) } n ≥ 1 \{\mathcal{L}_{h}^{(n)}\}_{n \ge1} { L h ( n ) } n ≥ 1 , there exists R > 0 R > 0 R > 0 such that
L h ( n ) ( [ − R , R ] s ) ≥ 1 − ε for every n ≥ 1. \mathcal{L}_{h}^{(n)}\left([-R,R]^{s}\right) \ge1-\varepsilon\qquad\text{for every } n \ge1. L h ( n ) ( [ − R , R ] s ) ≥ 1 − ε for every n ≥ 1.
Choose r ∈ Q r \in\mathbb{Q} r ∈ Q with r > R r > R r > R , and let α r = ( r , … , r ) ∈ Q s \alpha_{r}=(r,\ldots,r) \in\mathbb{Q}^{s} α r = ( r , … , r ) ∈ Q s . Since [ − R , R ] s ⊂ ( − ∞ , α r ] [-R,R]^{s} \subset(-\infty,\alpha_{r}] [ − R , R ] s ⊂ ( − ∞ , α r ] , (105) gives
a h ( α r ) = lim k → ∞ L h ( n k ) ( ( − ∞ , α r ] ) ≥ 1 − ε . a_{h}(\alpha_{r})=\lim_{k\to\infty}\mathcal{L}_{h}^{(n_{k})}\left((-\infty,\alpha_{r}]\right)\ge1-\varepsilon. a h ( α r ) = k → ∞ lim L h ( n k ) ( ( − ∞ , α r ] ) ≥ 1 − ε .
Because α ↦ M ~ α \alpha\mapsto\widetilde{M}_{\alpha} α ↦ M α is coordinatewise nondecreasing, (103) implies
M ‾ α r ≥ M ~ α r , \overline{M}_{\alpha_{r}} \ge\widetilde{M}_{\alpha_{r}}, M α r ≥ M α r ,
and hence
ϱ h ( M ‾ α r ) ≥ 1 − ε . \varrho_{h}\left(\overline{M}_{\alpha_{r}}\right) \ge1-\varepsilon. ϱ h ( M α r ) ≥ 1 − ε .
Letting r → ∞ r \to\infty r → ∞ through the rationals and using normality of ϱ h \varrho_{h} ϱ h , we obtain
ϱ h ( M ‾ ( R s ) ) = 1. \varrho_{h}\left(\overline{M}(\mathbb{R}^{s})\right)=1. ϱ h ( M ( R s ) ) = 1.
Therefore,
ϱ h ( M ‾ ( R ‾ s ∖ R s ) ) = 0. (106) \varrho_{h}\left(\overline{M}\left(\overline{\mathbb{R}}^{s}\setminus\mathbb{R}^{s}\right)\right)=0. \tag*{(106)} ϱ h ( M ( R s ∖ R s ) ) = 0. (106)
It follows from (104) and (106) that
μ h ( B ) = ϱ h ( M ‾ ( B ) ) , B ∈ B ( R s ) . \mu_{h}(B)=\varrho_{h}\left(\overline{M}(B)\right), \qquad B\in\mathcal{B}(\mathbb{R}^{s}). μ h ( B ) = ϱ h ( M ( B ) ) , B ∈ B ( R s ) .
Consequently, for every t ∈ R s t\in\mathbb{R}^{s} t ∈ R s , normality of ϱ h \varrho_{h} ϱ h and (103) yield
μ h ( ( − ∞ , t ] ) = ϱ h ( M ‾ t ) = inf α ∈ Q s α > t ϱ h ( M ~ α ) = inf α ∈ Q s α > t a h ( α ) . (107) \mu_{h}((-\infty,t])=\varrho_{h}\left(\overline{M}_{t}\right)
=\inf_{\substack{\alpha\in\mathbb{Q}^{s}\\ \alpha>t}}\varrho_{h}\left(\widetilde{M}_{\alpha}\right)
=\inf_{\substack{\alpha\in\mathbb{Q}^{s}\\ \alpha>t}}a_{h}(\alpha). \tag*{(107)} μ h (( − ∞ , t ]) = ϱ h ( M t ) = α ∈ Q s α > t inf ϱ h ( M α ) = α ∈ Q s α > t inf a h ( α ) . (107)
Equations (105) and (107) verify the hypotheses of Lemma F.2 with P k = L h ( n k ) P_{k}=\mathcal{L}_{h}^{(n_{k})} P k = L h ( n k ) , P = μ h P=\mu_{h} P = μ h , a ( α ) = a h ( α ) a(\alpha)=a_{h}(\alpha) a ( α ) = a h ( α ) .
Hence
L h ( n k ) ⇒ μ h . \mathcal{L}_{h}^{(n_{k})}\Rightarrow\mu_{h}. L h ( n k ) ⇒ μ h .
Since h ∈ R d h\in\mathbb{R}^{d} h ∈ R d was arbitrary and the subsequence { n k } \{n_{k}\} { n k } was chosen independently of h h h , this proves the assertion for every fixed local alternative h h h .
Proof of Theorem 3.1. Write the mean parameter as h = ( y , z ) h=(y,z) h = ( y , z ) , where y ∈ R y\in\mathbb{R} y ∈ R is the parameter of interest and z ∈ R d − 1 z\in\mathbb{R}^{d-1} z ∈ R d − 1 is the nuisance parameter. Let
M = CCR ( S ) , ϕ y , z ∼ N ( ( y , z ) , Σ ) , Σ = V + − 1 S , \mathcal{M}=\operatorname{CCR}(S),\qquad\phi_{y,z}\sim N((y,z),\Sigma),\qquad\Sigma=V+\sqrt{-1}S, M = CCR ( S ) , ϕ y , z ∼ N (( y , z ) , Σ ) , Σ = V + − 1 S ,
where V ≻ 0 V\succ0 V ≻ 0 . Throughout, a test is an effect T ∈ M T\in\mathcal{M} T ∈ M .
Consider the nuisance-translation group G = ( R d − 1 , + ) G=(\mathbb{R}^{d-1},+) G = ( R d − 1 , + ) , acting on M \mathcal{M} M through
τ t : = α ( 0 , t ) , t ∈ R d − 1 , \tau_t:=\alpha_{(0,t)},\qquad t\in\mathbb{R}^{d-1}, τ t := α ( 0 , t ) , t ∈ R d − 1 ,
where α a \alpha_a α a denotes the normal automorphism constructed in Appendix § 12 . By the Gaussian covariance relation (88) ,
ϕ y , z ( τ t ( A ) ) = ϕ y , z + t ( A ) , A ∈ M . (108) \phi_{y,z}(\tau_t(A))=\phi_{y,z+t}(A),\qquad A\in\mathcal{M}. \tag*{(108)} ϕ y , z ( τ t ( A )) = ϕ y , z + t ( A ) , A ∈ M . (108)
The action is point-ultraweakly continuous, and hence satisfies the hypotheses of Theorem D.1 . Let T ∈ M T\in\mathcal{M} T ∈ M be any level-α \alpha α test. By Theorem D.1 , there exists an invariant level-α \alpha α test T ‾ ∈ M \overline{T}\in\mathcal{M} T ∈ M such that
inf z ∈ R d − 1 ϕ y , z ( T ) ≤ inf z ∈ R d − 1 ϕ y , z ( T ‾ ) , y > 0. (109) \inf_{z\in\mathbb{R}^{d-1}}\phi_{y,z}(T)\leq\inf_{z\in\mathbb{R}^{d-1}}\phi_{y,z}(\overline{T}),\qquad y>0. \tag*{(109)} z ∈ R d − 1 inf ϕ y , z ( T ) ≤ z ∈ R d − 1 inf ϕ y , z ( T ) , y > 0. (109)
Since the nuisance translations act transitively on R d − 1 \mathbb{R}^{d-1} R d − 1 , the power of T ‾ \overline{T} T is independent of z z z . By Lemma C.1 , the invariant algebra is
M G = A [ X 1 ] = { e − 1 t X 1 : t ∈ R } . \mathcal{M}^{G}=\mathcal{A}[X_1]=\left\{e^{\sqrt{-1}tX_1}:t\in\mathbb{R}\right\}. M G = A [ X 1 ] = { e − 1 t X 1 : t ∈ R } .
Thus every invariant test has the form T ‾ = f ( X 1 ) \overline{T}=f(X_1) T = f ( X 1 ) for some Borel function f : R → [ 0 , 1 ] f:\mathbb{R}\to[0,1] f : R → [ 0 , 1 ] . The Gaussian characteristic function gives
ϕ y , z ( e − 1 t X 1 ) = exp { − 1 t y − 1 2 t 2 V 11 } . \phi_{y,z}\left(e^{\sqrt{-1}tX_1}\right)=\exp\left\{\sqrt{-1}ty-\frac{1}{2}t^2V_{11}\right\}. ϕ y , z ( e − 1 t X 1 ) = exp { − 1 t y − 2 1 t 2 V 11 } .
Consequently, the restriction of ϕ y , z \phi_{y,z} ϕ y , z to A [ X 1 ] \mathcal{A}[X_1] A [ X 1 ] is the classical normal location model
X 1 ∼ N ( y , V 11 ) , X_1\sim N(y,V_{11}), X 1 ∼ N ( y , V 11 ) ,
which does not depend on the nuisance parameter z z z . In particular,
ϕ y , z ( T ‾ ) = ∫ R f ( x ) N ( y , V 11 ) ( d x ) . (110) \phi_{y,z}(\overline{T})=\int_{\mathbb{R}}f(x)N(y,V_{11})(dx). \tag*{(110)} ϕ y , z ( T ) = ∫ R f ( x ) N ( y , V 11 ) ( d x ) . (110)
Since T ‾ \overline{T} T has level α \alpha α , its restriction satisfies
∫ R f ( x ) N ( 0 , V 11 ) ( d x ) ≤ α . \int_{\mathbb{R}}f(x)N(0,V_{11})(dx)\leq\alpha. ∫ R f ( x ) N ( 0 , V 11 ) ( d x ) ≤ α .
The normal location family has a monotone likelihood ratio. Therefore, by the Neyman–Pearson lemma, for every y > 0 y>0 y > 0 ,
∫ R f ( x ) N ( y , V 11 ) ( d x ) ≤ 1 − Φ ( z α − y V 11 ) , \int_{\mathbb{R}}f(x)N(y,V_{11})(dx)\leq1-\Phi\left(z_\alpha-\frac{y}{\sqrt{V_{11}}}\right), ∫ R f ( x ) N ( y , V 11 ) ( d x ) ≤ 1 − Φ ( z α − V 11 y ) ,
where z α = Φ − 1 ( 1 − α ) z_\alpha=\Phi^{-1}(1-\alpha) z α = Φ − 1 ( 1 − α ) . Combining this inequality with (109) gives
inf z ∈ R d − 1 ϕ y , z ( T ) ≤ 1 − Φ ( z α − y V 11 ) , y > 0. \inf_{z\in\mathbb{R}^{d-1}}\phi_{y,z}(T) \le1-\Phi\left(z_\alpha-\frac{y}{\sqrt{V_{11}}}\right),\qquad y>0. z ∈ R d − 1 inf ϕ y , z ( T ) ≤ 1 − Φ ( z α − V 11 y ) , y > 0.
Finally, consider the test
T α = { X 1 > z α V 11 } . T_\alpha=\left\{X_1>z_\alpha\sqrt{V_{11}}\right\}. T α = { X 1 > z α V 11 } .
By the spectral theorem, T α ∈ A [ X 1 ] ⊆ M T_\alpha\in\mathcal{A}[X_1]\subseteq\mathcal{M} T α ∈ A [ X 1 ] ⊆ M . Its rejection probability is
ϕ y , z ( T α ) = 1 − Φ ( z α − y V 11 ) , \phi_{y,z}(T_\alpha)=1-\Phi\left(z_\alpha-\frac{y}{\sqrt{V_{11}}}\right), ϕ y , z ( T α ) = 1 − Φ ( z α − V 11 y ) ,
independently of z z z . This probability is increasing in y y y and equals α \alpha α at y = 0 y=0 y = 0 . Hence T α T_\alpha T α has level α \alpha α for the composite null y ≤ 0 y\le0 y ≤ 0 and attains the preceding bound for every y > 0 y>0 y > 0 .
Proof of Theorem 3.3. As in the proof of Theorem 3.2 , choose an invertible real matrix L L L whose first row is c T c^T c T . Introduce the transformed observables X ~ = L X \widetilde{X}=LX X = L X and mean parameter h ~ = L h = ( y , z ) \widetilde{h}=Lh=(y,z) h = L h = ( y , z ) , where
Y : = X ~ 1 = c T X , y = c T h , z ∈ R d − 1 . Y:=\widetilde{X}_1=c^T X,\qquad y=c^T h,\qquad z\in\mathbb{R}^{d-1}. Y := X 1 = c T X , y = c T h , z ∈ R d − 1 .
The transformed covariance is Σ ~ = L Σ L T \widetilde{\Sigma}=L\Sigma L^T Σ = L Σ L T , and
σ 2 : = c T V c = Σ ~ 11 > 0. \sigma^2:=c^T Vc=\widetilde{\Sigma}_{11}>0. σ 2 := c T V c = Σ 11 > 0.
Since L L L is invertible, the transformed observables generate the same algebra M = CCR ( S ) \mathcal{M}=\operatorname{CCR}(S) M = CCR ( S ) . The testing problem becomes
H 0 : y = 0 , H 1 : y ≠ 0. H_0:y=0,\qquad H_1:y\ne0. H 0 : y = 0 , H 1 : y = 0.
To distinguish the transformed parametrization from the original Gaussian shift family, write
ϕ ~ y , z : = ϕ L − 1 ( y , z ) , Σ , ( y , z ) ∈ R × R d − 1 . \widetilde{\phi}_{y,z}:=\phi_{L^{-1}(y,z),\Sigma},\qquad(y,z)\in\mathbb{R}\times\mathbb{R}^{d-1}. ϕ y , z := ϕ L − 1 ( y , z ) , Σ , ( y , z ) ∈ R × R d − 1 .
Equivalently, ϕ ~ y , z \widetilde{\phi}_{y,z} ϕ y , z is the Gaussian state with mean ( y , z ) (y,z) ( y , z ) and covariance Σ ~ \widetilde{\Sigma} Σ in the transformed coordinates. Let T ∈ M T\in\mathcal{M} T ∈ M be any level-α \alpha α unbiased test. The nuisance translations act through τ t = α L − 1 ( 0 , t ) \tau_t=\alpha L^{-1}(0,t) τ t = α L − 1 ( 0 , t ) , t ∈ R d − 1 t\in\mathbb{R}^{d-1} t ∈ R d − 1 , and satisfy
ϕ ~ y , z ( τ t ( A ) ) = ϕ ~ y , z + t ( A ) , A ∈ M . \widetilde{\phi}_{y,z}(\tau_t(A))=\widetilde{\phi}_{y,z+t}(A),\qquad A\in\mathcal{M}. ϕ y , z ( τ t ( A )) = ϕ y , z + t ( A ) , A ∈ M .
By Theorem D.1, there exists an invariant level-α \alpha α unbiased test T ‾ \overline{T} T such that
inf z ∈ R d − 1 ϕ ~ y , z ( T ) ≤ inf z ∈ R d − 1 ϕ ~ y , z ( T ‾ ) , y ≠ 0. (111) \inf_{z\in\mathbb{R}^{d-1}}\widetilde{\phi}_{y,z}(T)\le\inf_{z\in\mathbb{R}^{d-1}}\widetilde{\phi}_{y,z}(\overline{T}),\qquad y\ne0. \tag*{(111)} z ∈ R d − 1 inf ϕ y , z ( T ) ≤ z ∈ R d − 1 inf ϕ y , z ( T ) , y = 0. (111)
By Lemma C.1, applied in the transformed coordinates, every invariant test belongs to
M G = { e − 1 t Y : t ∈ R } ′ ′ . \mathcal{M}^{G}=\left\{e^{\sqrt{-1}tY}:t\in\mathbb{R}\right\}^{\prime\prime}. M G = { e − 1 t Y : t ∈ R } ′′ .
Thus T ‾ = f ( Y ) \overline{T}=f(Y) T = f ( Y ) for some Borel function f : R → [ 0 , 1 ] f:\mathbb{R}\to[0,1] f : R → [ 0 , 1 ] . Since
φ ~ y , z ( e − 1 t Y ) = exp ( − 1 t y − 1 2 t 2 σ 2 ) , \widetilde{\varphi}_{y,z}(e^{\sqrt{-1}tY})=\exp\left(\sqrt{-1}ty-\frac{1}{2}t^2\sigma^2\right), φ y , z ( e − 1 t Y ) = exp ( − 1 t y − 2 1 t 2 σ 2 ) ,
the invariant testing problem reduces to the classical normal location family Y ∼ N ( y , σ 2 ) Y\sim N(y,\sigma^2) Y ∼ N ( y , σ 2 ) , independently of the nuisance parameter z z z .
By the classical UMPU theorem for the normal location family, the level-α \alpha α unbiased test maximizing power against every y ≠ 0 y\ne0 y = 0 is
T α = { ∣ Y ∣ > z α / 2 σ } . T_\alpha=\{|Y|>z_{\alpha/2}\sigma\}. T α = { ∣ Y ∣ > z α /2 σ } .
where z α / 2 = Φ − 1 ( 1 − α / 2 ) z_{\alpha/2}=\Phi^{-1}(1-\alpha/2) z α /2 = Φ − 1 ( 1 − α /2 ) . Its power function is
β α ( y ) : = φ ~ y , z ( T α ) = 1 − Φ ( z α / 2 − y σ ) + Φ ( − z α / 2 − y σ ) . \begin{aligned}
\beta_\alpha(y):=\widetilde{\varphi}_{y,z}(T_\alpha) \\
&=1-\Phi\left(z_{\alpha/2}-\frac{y}{\sigma}\right)+\Phi\left(-z_{\alpha/2}-\frac{y}{\sigma}\right).
\end{aligned} β α ( y ) := φ y , z ( T α ) = 1 − Φ ( z α /2 − σ y ) + Φ ( − z α /2 − σ y ) .
Consequently, (111) implies
inf z ∈ R d − 1 φ ~ y , z ( T ) ≤ β α ( y ) , y ≠ 0. \inf_{z\in\mathbb{R}^{d-1}}\widetilde{\varphi}_{y,z}(T)\le\beta_\alpha(y),\qquad y\ne0. z ∈ R d − 1 inf φ y , z ( T ) ≤ β α ( y ) , y = 0.
The function β α \beta_\alpha β α is even, satisfies β α ( 0 ) = α \beta_\alpha(0)=\alpha β α ( 0 ) = α , and is nondecreasing in ∣ y ∣ |y| ∣ y ∣ . Hence, for every s > 0 s>0 s > 0 ,
inf ∣ y ∣ ≥ s z ∈ R d − 1 φ ~ y , z ( T ) ≤ inf z ∈ R d − 1 φ ~ s , z ( T ) ≤ β α ( s ) . \inf_{\substack{|y|\ge s\\ z\in\mathbb{R}^{d-1}}}\widetilde{\varphi}_{y,z}(T)
\le\inf_{z\in\mathbb{R}^{d-1}}\widetilde{\varphi}_{s,z}(T)
\le\beta_\alpha(s). ∣ y ∣ ≥ s z ∈ R d − 1 inf φ y , z ( T ) ≤ z ∈ R d − 1 inf φ s , z ( T ) ≤ β α ( s ) .
The proposed test T α T_\alpha T α is invariant, level-α \alpha α , and unbiased, and its minimum power over ∣ y ∣ ≥ s |y|\ge s ∣ y ∣ ≥ s equals β α ( s ) \beta_\alpha(s) β α ( s ) . Thus the bound is attained.
Returning to the original coordinates, where y = c T h y=c^T h y = c T h and σ 2 = c T V c \sigma^2=c^T Vc σ 2 = c T V c , proves the theorem. □
Proof of Theorem 4.1. Set
c : = ψ ˙ θ 0 T , σ 2 : = c T Σ c > 0 , b α ( t ) : = 1 − Φ ( z α − t σ ) , t ∈ R . c:=\dot{\psi}_{\theta_0}^T,\qquad\sigma^2:=c^T\Sigma c>0,\qquad b_\alpha(t):=1-\Phi\left(z_\alpha-\frac{t}{\sigma}\right),\qquad t\in\mathbb{R}. c := ψ ˙ θ 0 T , σ 2 := c T Σ c > 0 , b α ( t ) := 1 − Φ ( z α − σ t ) , t ∈ R .
For h ∈ R d h\in\mathbb{R}^d h ∈ R d , write β n ( h ) : = Tr ρ θ 0 + h / n ( n ) T ( n ) \beta_n(h):=\operatorname{Tr}\rho_{\theta_0+h/\sqrt{n}}^{(n)}T^{(n)} β n ( h ) := Tr ρ θ 0 + h / n ( n ) T ( n ) . We first prove the upper bound. Suppose, to the contrary, that it fails for some s > 0 s>0 s > 0 . Then there exists δ > 0 \delta>0 δ > 0 such that
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ( h ) > b α ( s ) + 2 δ . (112) \inf_{\substack{D\subset\mathbb{R}^d\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_n(h)>b_\alpha(s)+2\delta. \tag*{(112)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ( h ) > b α ( s ) + 2 δ . (112)
By continuity of b α b_\alpha b α , choose ϵ > 0 \epsilon>0 ϵ > 0 sufficiently small that b α ( s + ϵ ) < b α ( s ) + δ b_\alpha(s+\epsilon)<b_\alpha(s)+\delta b α ( s + ϵ ) < b α ( s ) + δ . Put
H 1 , s , ϵ : = { h ∈ R d : c T h ≥ s + ϵ } . H_{1,s,\epsilon}:=\{h\in\mathbb{R}^d:c^T h\ge s+\epsilon\}. H 1 , s , ϵ := { h ∈ R d : c T h ≥ s + ϵ } .
Since H 1 , s , ϵ H_{1,s,\epsilon} H 1 , s , ϵ is closed and separable, choose a countable dense subset { h 1 , h 2 , … } ⊂ H 1 , s , ϵ \{h_1,h_2,\ldots\}\subset H_{1,s,\epsilon} { h 1 , h 2 , … } ⊂ H 1 , s , ϵ and define D m : = { h 1 , … , h m } D_m:=\{h_1,\ldots,h_m\} D m := { h 1 , … , h m } . For each fixed m m m , differentiability of ψ \psi ψ at θ 0 \theta_0 θ 0 gives
n ψ ( θ 0 + h n ) = c T h + o ( 1 ) \sqrt{n}\,\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)=c^{T}h+o(1) n ψ ( θ 0 + n h ) = c T h + o ( 1 )
for every h ∈ D m h\in D_m h ∈ D m . Since D m D_m D m is finite and c T h ≥ s + ϵ c^{T}h\ge s+\epsilon c T h ≥ s + ϵ on D m D_m D m , we have D m ⊂ H 1 , s ( n ) D_m\subset H_{1,s}^{(n)} D m ⊂ H 1 , s ( n ) for all sufficiently large n n n . Consequently, (112) implies
lim sup n → ∞ min 1 ≤ j ≤ m β n ( h j ) > b α ( s ) + 2 δ ( m ≥ 1 ) . \limsup_{n\to\infty}\min_{1\le j\le m}\beta_n(h_j)>b_\alpha(s)+2\delta\qquad(m\ge1). n → ∞ lim sup 1 ≤ j ≤ m min β n ( h j ) > b α ( s ) + 2 δ ( m ≥ 1 ) .
We may therefore choose a strictly increasing sequence { n m } m ≥ 1 \{n_m\}_{m\ge1} { n m } m ≥ 1 such that
β n m ( h j ) ≥ b α ( s ) + δ , j = 1 , … , m . (113) \beta_{n_m}(h_j)\ge b_\alpha(s)+\delta,\qquad j=1,\ldots,m. \tag*{(113)} β n m ( h j ) ≥ b α ( s ) + δ , j = 1 , … , m . (113)
For each m m m , regard T ( n m ) T^{(n_m)} T ( n m ) as the binary POVM { T ( n m ) , I − T ( n m ) } \{T^{(n_m)},I-T^{(n_m)}\} { T ( n m ) , I − T ( n m ) } . Its outcome distributions are supported on { 0 , 1 } \{0,1\} { 0 , 1 } and hence are tight under every fixed local alternative. By Theorem 2.2, there exist a further subsequence { n m k } k ≥ 1 \{n_{m_k}\}_{k\ge1} { n m k } k ≥ 1 and a limiting Gaussian test 0 ≤ T ( ∞ ) ≤ I 0\le T^{(\infty)}\le I 0 ≤ T ( ∞ ) ≤ I such that
β n m k ( h ) → ϱ h ( T ( ∞ ) ) for every fixed h ∈ R d . (114) \beta_{n_{m_k}}(h)\to\varrho_h(T^{(\infty)})\qquad\text{for every fixed }h\in\mathbb{R}^{d}. \tag*{(114)} β n m k ( h ) → ϱ h ( T ( ∞ ) ) for every fixed h ∈ R d . (114)
Fix j ≥ 1 j\ge1 j ≥ 1 . Since m k ≥ j m_k\ge j m k ≥ j eventually, (112) and (113) give
ϱ h j ( T ( ∞ ) ) ≥ b α ( s ) + δ , j ≥ 1. (115) \varrho_{h_j}(T^{(\infty)})\ge b_\alpha(s)+\delta,\qquad j\ge1. \tag*{(115)} ϱ h j ( T ( ∞ ) ) ≥ b α ( s ) + δ , j ≥ 1. (115)
We next verify that T ( ∞ ) T^{(\infty)} T ( ∞ ) is a level-α \alpha α test for the Gaussian hypotheses H 0 : c T h ≤ 0 H_0:c^{T}h\le0 H 0 : c T h ≤ 0 versus H 1 : c T h > 0 H_1:c^{T}h>0 H 1 : c T h > 0 .
Let h ∈ R d h\in\mathbb{R}^{d} h ∈ R d satisfy c T h < 0 c^{T}h<0 c T h < 0 . Differentiability of ψ \psi ψ and ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 give
n ψ ( θ 0 + h n ) = c T h + o ( 1 ) < 0 \sqrt{n}\,\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)=c^{T}h+o(1)<0 n ψ ( θ 0 + n h ) = c T h + o ( 1 ) < 0
for all sufficiently large n n n . Since T ( n ) T^{(n)} T ( n ) has level α \alpha α under the original null, β n ( h ) ≤ α \beta_n(h)\le\alpha β n ( h ) ≤ α eventually. Hence (113) implies
ϱ h ( T ( ∞ ) ) ≤ α , c T h < 0. (116) \varrho_h(T^{(\infty)})\le\alpha,\qquad c^{T}h<0. \tag*{(116)} ϱ h ( T ( ∞ ) ) ≤ α , c T h < 0. (116)
By Lemma F.1, the map h ↦ ϱ h ( T ( ∞ ) ) h\mapsto\varrho_h(T^{(\infty)}) h ↦ ϱ h ( T ( ∞ ) ) is continuous. It follows that
ϱ h ( T ( ∞ ) ) ≤ α , c T h ≤ 0 , (117) \varrho_h(T^{(\infty)})\le\alpha,\qquad c^{T}h\le0, \tag*{(117)} ϱ h ( T ( ∞ ) ) ≤ α , c T h ≤ 0 , (117)
so T ( ∞ ) T^{(\infty)} T ( ∞ ) is level α \alpha α in the limiting Gaussian experiment.
By the same continuity argument, (114) extends from the dense set { h j } j ≥ 1 \{h_j\}_{j\ge1} { h j } j ≥ 1 to all of H 1 , s , ϵ H_{1,s,\epsilon} H 1 , s , ϵ :
ϱ h ( T ( ∞ ) ) ≥ b α ( s ) + δ , h ∈ H 1 , s , ϵ . (118) \varrho_h(T^{(\infty)})\ge b_\alpha(s)+\delta,\qquad h\in H_{1,s,\epsilon}. \tag*{(118)} ϱ h ( T ( ∞ ) ) ≥ b α ( s ) + δ , h ∈ H 1 , s , ϵ . (118)
In particular,
inf c T h ≥ s + ϵ ϱ h ( T ( ∞ ) ) ≥ b α ( s ) + δ . \inf_{c^{T}h\ge s+\epsilon}\varrho_h(T^{(\infty)})\ge b_\alpha(s)+\delta. c T h ≥ s + ϵ inf ϱ h ( T ( ∞ ) ) ≥ b α ( s ) + δ .
On the other hand, since T ( ∞ ) T^{(\infty)} T ( ∞ ) is level α \alpha α , Theorem 3.2 gives
inf c T h ≥ s + ϵ ϱ h ( T ( ∞ ) ) ≤ b α ( s + ϵ ) < b α ( s ) + δ , \inf_{c^{T}h\ge s+\epsilon}\varrho_h(T^{(\infty)})\le b_\alpha(s+\epsilon)<b_\alpha(s)+\delta, c T h ≥ s + ϵ inf ϱ h ( T ( ∞ ) ) ≤ b α ( s + ϵ ) < b α ( s ) + δ ,
contradicting the preceding inequality. This proves the upper bound.
We finally verify attainment of the Gaussian local power envelope. Consider
T α ( n ) = { ψ ˙ θ 0 X ( n ) > z α σ } . T_{\alpha}^{(n)}=\left\{\dot{\psi}_{\theta_0}X^{(n)}>z_{\alpha}\sigma\right\}. T α ( n ) = { ψ ˙ θ 0 X ( n ) > z α σ } .
Applying the sandwiched Le Cam third lemma under D \mathcal{D} D -extendibility [14 Corollary 3.7] , we obtain, for every fixed h ∈ R d h\in\mathbb{R}^{d} h ∈ R d ,
( ψ ˙ θ 0 X ( n ) , ρ θ 0 + h / n ( n ) ) ⇒ N ( c T h , σ 2 ) . \left(\dot{\psi}_{\theta_0}X^{(n)},\rho_{\theta_0+h/\sqrt{n}}^{(n)}\right)\Rightarrow N(c^{T}h,\sigma^{2}). ( ψ ˙ θ 0 X ( n ) , ρ θ 0 + h / n ( n ) ) ⇒ N ( c T h , σ 2 ) .
Since the limiting normal distribution assigns zero mass to z α σ z_{\alpha}\sigma z α σ ,
β n ∗ ( h ) : = Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) → 1 − Φ ( z α − c T h σ ) = b α ( c T h ) . (119) \begin{aligned}
\beta_{n}^{\ast}(h)&:=\operatorname{Tr}\left(\rho_{\theta_0+h/\sqrt{n}}^{(n)}T_{\alpha}^{(n)}\right)\\
&\to1-\Phi\left(z_{\alpha}-\frac{c^{T}h}{\sigma}\right)=b_{\alpha}(c^{T}h). \tag*{(119)}
\end{aligned} β n ∗ ( h ) := Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) → 1 − Φ ( z α − σ c T h ) = b α ( c T h ) . (119)
In particular, setting h = 0 h=0 h = 0 gives lim n → ∞ β n ∗ ( 0 ) = α \lim_{n\to\infty}\beta_{n}^{\ast}(0)=\alpha lim n → ∞ β n ∗ ( 0 ) = α .
We now show that these tests attain the stated power bound for the original n n n -dependent alternative sets.
Fix s > 0 s>0 s > 0 , a nonempty finite set D ⊂ R d D\subset\mathbb{R}^{d} D ⊂ R d , and η ∈ ( 0 , s ) \eta\in(0,s) η ∈ ( 0 , s ) . Define
D − η : = { h ∈ D : c T h ≥ s − η } . D_{-\eta}:=\{h\in D:c^{T}h\ge s-\eta\}. D − η := { h ∈ D : c T h ≥ s − η } .
For every h ∈ D ∖ D − η h\in D\setminus D_{-\eta} h ∈ D ∖ D − η , differentiability of ψ \psi ψ gives
n ψ ( θ 0 + h n ) = c T h + o ( 1 ) < s \sqrt{n}\,\psi\left(\theta_{0}+\frac{h}{\sqrt{n}}\right)=c^{T}h+o(1)<s n ψ ( θ 0 + n h ) = c T h + o ( 1 ) < s
for all sufficiently large n n n . Since D D D is finite, it follows that D ∩ H 1 , s ( n ) ⊆ D − η D\cap H_{1,s}^{(n)}\subseteq D_{-\eta} D ∩ H 1 , s ( n ) ⊆ D − η eventually. Consequently,
inf h ∈ D ∩ H 1 , s ( n ) β n ∗ ( h ) ≥ inf h ∈ D − η β n ∗ ( h ) \inf_{h\in D\cap H_{1,s}^{(n)}}\beta_{n}^{\ast}(h)\ge\inf_{h\in D_{-\eta}}\beta_{n}^{\ast}(h) h ∈ D ∩ H 1 , s ( n ) inf β n ∗ ( h ) ≥ h ∈ D − η inf β n ∗ ( h )
for all sufficiently large $n.
If D − η ≠ ∅ D_{-\eta}\ne\varnothing D − η = ∅ , its finiteness and (119) imply
lim inf n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ∗ ( h ) ≥ min h ∈ D − η b α ( c T h ) ≥ b α ( s − η ) . \liminf_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_{n}^{\ast}(h)\ge\min_{h\in D_{-\eta}}b_{\alpha}(c^{T}h)\ge b_{\alpha}(s-\eta). n → ∞ lim inf h ∈ D ∩ H 1 , s ( n ) inf β n ∗ ( h ) ≥ h ∈ D − η min b α ( c T h ) ≥ b α ( s − η ) .
If D − η = ∅ D_{-\eta}=\varnothing D − η = ∅ , then D ∩ H 1 , s ( n ) D\cap H_{1,s}^{(n)} D ∩ H 1 , s ( n ) is eventually empty and the same lower bound holds by the convention inf ∅ = + ∞ \inf\varnothing=+\infty inf ∅ = + ∞ .
Since η > 0 \eta>0 η > 0 is arbitrary and b α b_{\alpha} b α is continuous, letting η ↓ 0 \eta\downarrow0 η ↓ 0 yields
lim inf n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ∗ ( h ) ≥ b α ( s ) . \liminf_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_{n}^{\ast}(h)\ge b_{\alpha}(s). n → ∞ lim inf h ∈ D ∩ H 1 , s ( n ) inf β n ∗ ( h ) ≥ b α ( s ) .
Since D D D was arbitrary,
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ∗ ( h ) ≥ b α ( s ) . (120) \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_{n}^{\ast}(h)\ge b_{\alpha}(s). \tag*{(120)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ∗ ( h ) ≥ b α ( s ) . (120)
Conversely, fix η > 0 \eta> 0 η > 0 and put h η : = ( s + η ) c / ∥ c ∥ 2 h_\eta:= (s+\eta)c/\lVert c\rVert^2 h η := ( s + η ) c / ∥ c ∥ 2 . Then c T h η = s + η c^{T}h_\eta=s+\eta c T h η = s + η , and differentiability gives
n ψ ( θ 0 + h η n ) = s + η + o ( 1 ) > s \sqrt{n}\,\psi\left(\theta_0+\frac{h_\eta}{\sqrt{n}}\right)=s+\eta+o(1)>s n ψ ( θ 0 + n h η ) = s + η + o ( 1 ) > s
for all sufficiently large n n n . Thus h η ∈ H 1 , s ( n ) h_\eta\in H^{(n)}_{1,s} h η ∈ H 1 , s ( n ) eventually. Taking D = { h η } D=\{h_\eta\} D = { h η } and using (119) , we get
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ∗ ( h ) ≤ lim n → ∞ β n ∗ ( h η ) = b α ( s + η ) . \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H^{(n)}_{1,s}}\beta_n^*(h)\leq\lim_{n\to\infty}\beta_n^*(h_\eta)=b_\alpha(s+\eta). D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ∗ ( h ) ≤ n → ∞ lim β n ∗ ( h η ) = b α ( s + η ) .
Letting η ↓ 0 \eta\downarrow0 η ↓ 0 and using continuity of b α b_\alpha b α gives
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ∗ ( h ) ≤ b α ( s ) . (121) \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H^{(n)}_{1,s}}\beta_n^*(h)\leq b_\alpha(s). \tag*{(121)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ∗ ( h ) ≤ b α ( s ) . (121)
Combining (120) and (121) , we obtain
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) = b α ( s ) . \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H^{(n)}_{1,s}}\operatorname{Tr}\left(\rho^{(n)}_{\theta_0+h/\sqrt{n}}T_\alpha^{(n)}\right)=b_\alpha(s). D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) = b α ( s ) .
Thus the proposed scalar tests attain the Gaussian local power envelope in the stated finite-submodel criterion for every s > 0 s>0 s > 0 . □
Proof of Theorem 4.2. Set
c : = ψ ˙ θ 0 T , σ 2 : = c T Σ c > 0 , c:=\dot{\psi}_{\theta_0}^{T},\qquad\sigma^2:=c^{T}\Sigma c>0, c := ψ ˙ θ 0 T , σ 2 := c T Σ c > 0 ,
and define
b α ( t ) : = 1 − Φ ( − z α / 2 − t σ , z α / 2 − t σ ) , t ∈ R . b_\alpha(t):=1-\Phi\left(-z_{\alpha/2}-\frac{t}{\sigma},z_{\alpha/2}-\frac{t}{\sigma}\right),\qquad t\in\mathbb{R}. b α ( t ) := 1 − Φ ( − z α /2 − σ t , z α /2 − σ t ) , t ∈ R .
For h ∈ R d h\in\mathbb{R}^{d} h ∈ R d , write β n ( h ) : = Tr ρ θ 0 + h / n ( n ) T ( n ) \beta_n(h):=\operatorname{Tr}\rho^{(n)}_{\theta_0+h/\sqrt{n}}T^{(n)} β n ( h ) := Tr ρ θ 0 + h / n ( n ) T ( n ) . We first prove the upper bound. Suppose, to the contrary, that it fails for some s > 0 s>0 s > 0 . Then there exists δ > 0 \delta>0 δ > 0 such that
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ( h ) > b α ( s ) + 3 δ . (122) \inf_{\substack{D\subset\mathbb{R}^{d}\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H^{(n)}_{1,s}}\beta_n(h)>b_\alpha(s)+3\delta. \tag*{(122)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ( h ) > b α ( s ) + 3 δ . (122)
By continuity of b α b_\alpha b α , choose η > 0 \eta>0 η > 0 sufficiently small that
b α ( s + η ) < b α ( s ) + δ . (123) b_\alpha(s+\eta)<b_\alpha(s)+\delta. \tag*{(123)} b α ( s + η ) < b α ( s ) + δ . (123)
Define
A s + η : = { h ∈ R d : ∣ c T h ∣ ≥ s + η } . A_{s+\eta}:=\{h\in\mathbb{R}^{d}:|c^{T}h|\geq s+\eta\}. A s + η := { h ∈ R d : ∣ c T h ∣ ≥ s + η } .
Since A s + η A_{s+\eta} A s + η is closed and separable, choose a countable dense subset { h 1 , h 2 , … } ⊆ A s + η \{h_1,h_2,\ldots\}\subseteq A_{s+\eta} { h 1 , h 2 , … } ⊆ A s + η and put D m : = { h 1 , … , h m } D_m:=\{h_1,\ldots,h_m\} D m := { h 1 , … , h m } . For each fixed h j ∈ A s + η h_j\in A_{s+\eta} h j ∈ A s + η , differentiability of ψ \psi ψ gives
n ψ ( θ 0 + h j n ) = c T h j + o ( 1 ) . \sqrt{n}\,\psi\left(\theta_0+\frac{h_j}{\sqrt{n}}\right)=c^{T}h_j+o(1). n ψ ( θ 0 + n h j ) = c T h j + o ( 1 ) .
Since ∣ c T h j ∣ ≥ s + η |c^{T}h_j| \ge s+\eta ∣ c T h j ∣ ≥ s + η , we have h j ∈ H 1 , s ( n ) h_j \in H_{1,s}^{(n)} h j ∈ H 1 , s ( n ) for all sufficiently large n n n . Because D m D_m D m is finite, D m ⊆ H 1 , s ( n ) D_m \subseteq H_{1,s}^{(n)} D m ⊆ H 1 , s ( n ) eventually for each fixed m m m . Thus (122) implies
lim sup n → ∞ min 1 ≤ j ≤ m β n ( h j ) > b α ( s ) + 3 δ , m ≥ 1. \limsup_{n\to\infty}\min_{1\le j\le m}\beta_n(h_j)>b_\alpha(s)+3\delta,\qquad m\ge1. n → ∞ lim sup 1 ≤ j ≤ m min β n ( h j ) > b α ( s ) + 3 δ , m ≥ 1.
Consequently, we may choose a strictly increasing sequence { n m } m ≥ 1 \{n_m\}_{m\ge1} { n m } m ≥ 1 such that
β n m ( h j ) ≥ b α ( s ) + 2 δ , j = 1 , … , m . (124) \beta_{n_m}(h_j)\ge b_\alpha(s)+2\delta,\qquad j=1,\ldots,m. \tag*{(124)} β n m ( h j ) ≥ b α ( s ) + 2 δ , j = 1 , … , m . (124)
For each m m m , regard T ( n m ) T^{(n_m)} T ( n m ) as the binary POVM { T ( n m ) , I − T ( n m ) } \{T^{(n_m)},I-T^{(n_m)}\} { T ( n m ) , I − T ( n m ) } . Its outcome distributions are supported on { 0 , 1 } \{0,1\} { 0 , 1 } and hence are tight under every fixed local alternative. By Theorem 2.2 , there exist a further subsequence { n m k } k ≥ 1 \{n_{m_k}\}_{k\ge1} { n m k } k ≥ 1 and a limiting Gaussian test 0 ≤ T ( ∞ ) ≤ I 0\le T^{(\infty)}\le I 0 ≤ T ( ∞ ) ≤ I such that
β n m k ( h ) → ϱ h ( T ( ∞ ) ) for every fixed h ∈ R d . (125) \beta_{n_{m_k}}(h)\to\varrho_h(T^{(\infty)})
\qquad\text{for every fixed }h\in\mathbb{R}^d. \tag*{(125)} β n m k ( h ) → ϱ h ( T ( ∞ ) ) for every fixed h ∈ R d . (125)
Fix j ≥ 1 j\ge1 j ≥ 1 . Since m k ≥ j m_k\ge j m k ≥ j eventually, the preceding two displays give
ϱ h j ( T ( ∞ ) ) ≥ b α ( s ) + 2 δ , j ≥ 1. (126) \varrho_{h_j}(T^{(\infty)})\ge b_\alpha(s)+2\delta,\qquad j\ge1. \tag*{(126)} ϱ h j ( T ( ∞ ) ) ≥ b α ( s ) + 2 δ , j ≥ 1. (126)
By Lemma F.1 , the map h ↦ ϱ h ( T ( ∞ ) ) h\mapsto\varrho_h(T^{(\infty)}) h ↦ ϱ h ( T ( ∞ ) ) is continuous. Hence this inequality extends to all of A s + η A_{s+\eta} A s + η :
ϱ h ( T ( ∞ ) ) ≥ b α ( s ) + 2 δ , ∣ c T h ∣ ≥ s + η . (127) \varrho_h(T^{(\infty)})\ge b_\alpha(s)+2\delta,\qquad|c^{T}h|\ge s+\eta. \tag*{(127)} ϱ h ( T ( ∞ ) ) ≥ b α ( s ) + 2 δ , ∣ c T h ∣ ≥ s + η . (127)
We next verify that T ( ∞ ) T^{(\infty)} T ( ∞ ) is a level-α \alpha α unbiased test for the Gaussian hypotheses H 0 : c T h = 0 H_0:c^{T}h=0 H 0 : c T h = 0 versus H 1 : c T h ≠ 0 H_1:c^{T}h\ne0 H 1 : c T h = 0 . Let h ∈ R d h\in\mathbb{R}^d h ∈ R d satisfy c T h ≠ 0 c^{T}h\ne0 c T h = 0 . Differentiability of ψ \psi ψ at θ 0 \theta_0 θ 0 , together with ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 , gives
n ψ ( θ 0 + h n ) = c T h + o ( 1 ) . \sqrt{n}\,\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)=c^{T}h+o(1). n ψ ( θ 0 + n h ) = c T h + o ( 1 ) .
Thus ψ ( θ 0 + h / n ) ≠ 0 \psi(\theta_0+h/\sqrt{n})\ne0 ψ ( θ 0 + h / n ) = 0 for all sufficiently large n n n . Since T ( n ) T^{(n)} T ( n ) is unbiased for the original two-sided testing problem, β n ( h ) ≥ α \beta_n(h)\ge\alpha β n ( h ) ≥ α eventually. By (125) ,
ϱ h ( T ( ∞ ) ) ≥ α , c T h ≠ 0. (128) \varrho_h(T^{(\infty)})\ge\alpha,\qquad c^{T}h\ne0. \tag*{(128)} ϱ h ( T ( ∞ ) ) ≥ α , c T h = 0. (128)
It remains to verify the level condition on the limiting null hyperplane. Fix h ∈ R d h\in\mathbb{R}^d h ∈ R d such that c T h = 0 c^{T}h=0 c T h = 0 . By Lemma F.3 , there exists h n → h h_n\to h h n → h satisfying
ψ ( θ 0 + h n n ) = 0 \psi\left(\theta_0+\frac{h_n}{\sqrt{n}}\right)=0 ψ ( θ 0 + n h n ) = 0
for all sufficiently large n n n . The finite-sample level condition therefore gives
Tr ( ρ θ 0 + h n / n ( n ) T ( n ) ) ≤ α \operatorname{Tr}\left(\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}T^{(n)}\right)\le\alpha Tr ( ρ θ 0 + h n / n ( n ) T ( n ) ) ≤ α
eventually. By assumption (25) ,
∥ ρ θ 0 + h n / n ( n ) − ρ θ 0 + h / n ( n ) ∥ 1 → 0. \left\|\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}-\rho_{\theta_0+h/\sqrt{n}}^{(n)}\right\|_1\to0. ρ θ 0 + h n / n ( n ) − ρ θ 0 + h / n ( n ) 1 → 0.
Since 0 ≤ T ( n ) ≤ I 0 \le T^{(n)} \le I 0 ≤ T ( n ) ≤ I ,
β n ( h ) ≤ Tr ( ρ θ 0 + h n / n ( n ) T ( n ) ) + ∥ ρ θ 0 + h n / n ( n ) − ρ θ 0 + h / n ( n ) ∥ 1 ≤ α + o ( 1 ) . \beta_n(h) \le\operatorname{Tr}\left(\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}T^{(n)}\right)+\left\|\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}-\rho_{\theta_0+h/\sqrt{n}}^{(n)}\right\|_1
\le\alpha+o(1). β n ( h ) ≤ Tr ( ρ θ 0 + h n / n ( n ) T ( n ) ) + ρ θ 0 + h n / n ( n ) − ρ θ 0 + h / n ( n ) 1 ≤ α + o ( 1 ) .
Consequently, lim sup n → ∞ β n ( h ) ≤ α \limsup_{n\to\infty}\beta_n(h)\le\alpha lim sup n → ∞ β n ( h ) ≤ α . Passing to the extracted subsequence, we obtain
ϱ h ( T ( ∞ ) ) ≤ α , c T h = 0. (129) \varrho_h(T^{(\infty)})\le\alpha,\qquad c^T h=0. \tag*{(129)} ϱ h ( T ( ∞ ) ) ≤ α , c T h = 0. (129)
Combining (128) and (129) , we conclude that T ( ∞ ) T^{(\infty)} T ( ∞ ) is level-α \alpha α unbiased in the limiting Gaussian experiment. Therefore, Theorem 3.3 gives
inf ∣ c T h ∣ ≥ s + η ϱ h ( T ( ∞ ) ) ≤ 1 − Φ ( − z α / 2 − s + η σ , z α / 2 − s + η σ ) = b α ( s + η ) < b α ( s ) + δ , \inf_{\lvert c^T h\rvert\ge s+\eta}\varrho_h(T^{(\infty)})\le1-\Phi\left(-z_{\alpha/2}-\frac{s+\eta}{\sigma},z_{\alpha/2}-\frac{s+\eta}{\sigma}\right)
=b_\alpha(s+\eta)<b_\alpha(s)+\delta, ∣ c T h ∣ ≥ s + η inf ϱ h ( T ( ∞ ) ) ≤ 1 − Φ ( − z α /2 − σ s + η , z α /2 − σ s + η ) = b α ( s + η ) < b α ( s ) + δ ,
where the last inequality follows from (123) . This contradicts (127) and proves the asserted upper bound.
We finally verify attainment of the Gaussian local power envelope. Consider
T α ( n ) = { ∣ ψ ˙ θ 0 X ( n ) ∣ > z α / 2 σ } . T_\alpha^{(n)}=\left\{\left|\dot{\psi}_{\theta_0}X^{(n)}\right|>z_{\alpha/2}\sigma\right\}. T α ( n ) = { ψ ˙ θ 0 X ( n ) > z α /2 σ } .
The sandwiched Le Cam third lemma under D \mathcal{D} D -extendibility [14 Corollary 3.7] gives, for every fixed h ∈ R d h\in\mathbb{R}^d h ∈ R d ,
( ψ ˙ θ 0 X ( n ) , ρ θ 0 + h / n ( n ) ) ⇒ N ( c T h , σ 2 ) . \left(\dot{\psi}_{\theta_0}X^{(n)},\rho_{\theta_0+h/\sqrt{n}}^{(n)}\right)\Rightarrow N(c^T h,\sigma^2). ( ψ ˙ θ 0 X ( n ) , ρ θ 0 + h / n ( n ) ) ⇒ N ( c T h , σ 2 ) .
Since the limiting normal law assigns zero mass to the rejection thresholds, we obtain
β n ∗ ( h ) : = Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) → b α ( c T h ) . (130) \beta_n^*(h):=\operatorname{Tr}\left(\rho_{\theta_0+h/\sqrt{n}}^{(n)}T_\alpha^{(n)}\right)\to b_\alpha(c^T h). \tag*{(130)} β n ∗ ( h ) := Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) → b α ( c T h ) . (130)
The function b α b_\alpha b α is even, nondecreasing in ∣ t ∣ \lvert t\rvert ∣ t ∣ , and satisfies b α ( 0 ) = α b_\alpha(0)=\alpha b α ( 0 ) = α . In particular, lim n → ∞ β n ∗ ( 0 ) = α \lim_{n\to\infty}\beta_n^*(0)=\alpha lim n → ∞ β n ∗ ( 0 ) = α .
We now show attainment for the original n n n -dependent alternative sets
H 1 , s ( n ) = { h ∈ R d : ∣ n ψ ( θ 0 + h n ) ∣ ≥ s } . H_{1,s}^{(n)}=\left\{h\in\mathbb{R}^d:\left|\sqrt{n}\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)\right|\ge s\right\}. H 1 , s ( n ) = { h ∈ R d : n ψ ( θ 0 + n h ) ≥ s } .
Fix s > 0 s>0 s > 0 , a nonempty finite set D ⊂ R d D\subset\mathbb{R}^d D ⊂ R d , and ϵ ∈ ( 0 , s ) \epsilon\in(0,s) ϵ ∈ ( 0 , s ) . Put
D − ϵ : = { h ∈ D : ∣ c T h ∣ ≥ s − ϵ } . D_{-\epsilon}:=\{h\in D:\lvert c^T h\rvert\ge s-\epsilon\}. D − ϵ := { h ∈ D : ∣ c T h ∣ ≥ s − ϵ } .
For every h ∈ D ∖ D − ϵ h\in D\setminus D_{-\epsilon} h ∈ D ∖ D − ϵ , differentiability gives
n ψ ( θ 0 + h n ) = c T h + o ( 1 ) , \sqrt{n}\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)=c^T h+o(1), n ψ ( θ 0 + n h ) = c T h + o ( 1 ) ,
so ∣ n ψ ( θ 0 + h / n ) ∣ < s \left|\sqrt{n}\psi(\theta_0+h/\sqrt{n})\right|<s ∣ n ψ ( θ 0 + h / n ) ∣ < s for all sufficiently large n n n . Since D D D is finite,
D ∩ H 1 , s ( n ) ⊆ D − ϵ . D\cap H_{1,s}^{(n)}\subseteq D_{-\epsilon}. D ∩ H 1 , s ( n ) ⊆ D − ϵ .
eventually. Hence
inf h ∈ D ∩ H 1 , s ( n ) β n ⋆ ( h ) ≥ inf h ∈ D − ϵ β n ⋆ ( h ) \inf_{h\in D\cap H_{1,s}^{(n)}} \beta_n^\star(h) \ge\inf_{h\in D_{-\epsilon}} \beta_n^\star(h) h ∈ D ∩ H 1 , s ( n ) inf β n ⋆ ( h ) ≥ h ∈ D − ϵ inf β n ⋆ ( h )
for all sufficiently large n n n .
If D − ϵ ≠ ∅ D_{-\epsilon}\ne\varnothing D − ϵ = ∅ , its finiteness and (130) give
lim inf n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ⋆ ( h ) ≥ min h ∈ D − ϵ b α ( c T h ) ≥ b α ( s − ϵ ) . \liminf_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_n^\star(h)\ge\min_{h\in D_{-\epsilon}}b_\alpha(c^T h)\ge b_\alpha(s-\epsilon). n → ∞ lim inf h ∈ D ∩ H 1 , s ( n ) inf β n ⋆ ( h ) ≥ h ∈ D − ϵ min b α ( c T h ) ≥ b α ( s − ϵ ) .
If D − ϵ = ∅ D_{-\epsilon}=\varnothing D − ϵ = ∅ , then D ∩ H 1 , s ( n ) D\cap H_{1,s}^{(n)} D ∩ H 1 , s ( n ) is eventually empty and the same inequality holds under the convention inf ∅ = + ∞ \inf\varnothing=+\infty inf ∅ = + ∞ . Letting ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 , we obtain
lim inf n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ⋆ ( h ) ≥ b α ( s ) . \liminf_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_n^\star(h)\ge b_\alpha(s). n → ∞ lim inf h ∈ D ∩ H 1 , s ( n ) inf β n ⋆ ( h ) ≥ b α ( s ) .
Since D D D was arbitrary,
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ⋆ ( h ) ≥ b α ( s ) . (131) \inf_{\substack{D\subset\mathbb{R}^d\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_n^\star(h)\ge b_\alpha(s). \tag*{(131)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ⋆ ( h ) ≥ b α ( s ) . (131)
Conversely, fix ϵ > 0 \epsilon>0 ϵ > 0 and put h ϵ : = ( s + ϵ ) c / ∥ c ∥ 2 h_\epsilon:=(s+\epsilon)c/\lVert c\rVert^2 h ϵ := ( s + ϵ ) c / ∥ c ∥ 2 . Then c T h ϵ = s + ϵ c^T h_\epsilon=s+\epsilon c T h ϵ = s + ϵ , and differentiability yields
n ψ ( θ 0 + h ϵ n ) = s + ϵ + o ( 1 ) . \sqrt{n}\,\psi\left(\theta_0+\frac{h_\epsilon}{\sqrt{n}}\right)=s+\epsilon+o(1). n ψ ( θ 0 + n h ϵ ) = s + ϵ + o ( 1 ) .
Thus h ϵ ∈ H 1 , s ( n ) h_\epsilon\in H_{1,s}^{(n)} h ϵ ∈ H 1 , s ( n ) for all sufficiently large n n n . Taking D = { h ϵ } D=\{h_\epsilon\} D = { h ϵ } and using (130) , we obtain
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ⋆ ( h ) ≤ lim n → ∞ β n ⋆ ( h ϵ ) = b α ( s + ϵ ) . \inf_{\substack{D\subset\mathbb{R}^d\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_n^\star(h)\le\lim_{n\to\infty}\beta_n^\star(h_\epsilon)=b_\alpha(s+\epsilon). D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ⋆ ( h ) ≤ n → ∞ lim β n ⋆ ( h ϵ ) = b α ( s + ϵ ) .
Letting ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 and using continuity of b α b_\alpha b α gives
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) β n ⋆ ( h ) ≤ b α ( s ) . (132) \inf_{\substack{D\subset\mathbb{R}^d\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\beta_n^\star(h)\le b_\alpha(s). \tag*{(132)} D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf β n ⋆ ( h ) ≤ b α ( s ) . (132)
Combining (131) and (132) , we conclude that
inf D ⊂ R d 0 < ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D ∩ H 1 , s ( n ) Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) = b α ( s ) . \inf_{\substack{D\subset\mathbb{R}^d\\0<|D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D\cap H_{1,s}^{(n)}}\operatorname{Tr}\left(\rho_{\theta_0+h/\sqrt{n}}^{(n)}T_\alpha^{(n)}\right)=b_\alpha(s). D ⊂ R d 0 < ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D ∩ H 1 , s ( n ) inf Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) = b α ( s ) .
Thus the proposed scalar tests attain the Gaussian local power envelope for the original n n n -dependent alternative sets H 1 , s ( n ) H_{1,s}^{(n)} H 1 , s ( n ) .
Proof of Theorem 4.3 . The proof follows the asymptotic minimax argument of [14 ] . Set
c : = ψ ˙ θ 0 T , σ 2 : = c T Σ c > 0 , c:=\dot{\psi}_{\theta_0}^{T},\qquad\sigma^2:=c^T\Sigma c>0, c := ψ ˙ θ 0 T , σ 2 := c T Σ c > 0 ,
and define
b α ( t ) : = 1 − Φ ( − z α / 2 − t σ , z α / 2 − t σ ) , t ∈ R . b_\alpha(t):=1-\Phi\left(-z_{\alpha/2}-\frac{t}{\sigma},z_{\alpha/2}-\frac{t}{\sigma}\right),\qquad t\in\mathbb{R}. b α ( t ) := 1 − Φ ( − z α /2 − σ t , z α /2 − σ t ) , t ∈ R .
For h ∈ R d h \in\mathbb{R}^{d} h ∈ R d , write β n ( h ) : = Tr ρ θ 0 + h / n ( n ) T ( n ) \beta_{n}(h) := \operatorname{Tr}\rho_{\theta_{0}+h/\sqrt{n}}^{(n)}T^{(n)} β n ( h ) := Tr ρ θ 0 + h / n ( n ) T ( n ) .
We first prove the upper bound. Fix s > 0 s>0 s > 0 and choose a countable dense subset { h 1 , h 2 , … } \{h_{1},h_{2},\ldots\} { h 1 , h 2 , … } of H 1 , s H_{1,s} H 1 , s . Let D k : = { h 1 , … , h k } D_{k}:=\{h_{1},\ldots,h_{k}\} D k := { h 1 , … , h k } and put
γ k ( n ) : = min h ∈ D k β n ( h ) , γ k : = lim sup n → ∞ γ k ( n ) , γ ∞ : = lim k → ∞ γ k . \gamma_{k}^{(n)}:=\min_{h\in D_{k}}\beta_{n}(h),\qquad\gamma_{k}:=\limsup_{n\to\infty}\gamma_{k}^{(n)},\qquad\gamma_{\infty}:=\lim_{k\to\infty}\gamma_{k}. γ k ( n ) := h ∈ D k min β n ( h ) , γ k := n → ∞ lim sup γ k ( n ) , γ ∞ := k → ∞ lim γ k .
The last limit exists because γ k \gamma_{k} γ k is nonincreasing and belongs to [ 0 , 1 ] [0,1] [ 0 , 1 ] . Since each D k D_{k} D k is an admissible finite subset of H 1 , s H_{1,s} H 1 , s , the left-hand side of (28) is at most γ ∞ \gamma_{\infty} γ ∞ . It therefore suffices to establish γ ∞ ≤ b α ( s ) \gamma_{\infty}\le b_{\alpha}(s) γ ∞ ≤ b α ( s ) .
Choose a strictly increasing sequence ( n k ) (n_{k}) ( n k ) such that γ k ( n k ) − γ k → 0 \gamma_{k}^{(n_{k})}-\gamma_{k}\to0 γ k ( n k ) − γ k → 0 . Then γ k ( n k ) → γ ∞ \gamma_{k}^{(n_{k})}\to\gamma_{\infty} γ k ( n k ) → γ ∞ . The binary POVMs { T ( n k ) , I − T ( n k ) } \{T^{(n_{k})},I-T^{(n_{k})}\} { T ( n k ) , I − T ( n k ) } are locally tight. By Theorem 2.2 , there exist a further subsequence ( n k t ) (n_{k_{t}}) ( n k t ) and a Gaussian test 0 ≤ T ( ∞ ) ≤ I 0\le T^{(\infty)}\le I 0 ≤ T ( ∞ ) ≤ I such that
β n k t ( h ) → ϱ h ( T ( ∞ ) ) for every fixed h ∈ R d , (133) \beta_{n_{k_{t}}}(h)\to\varrho_{h}(T^{(\infty)})
\quad\text{for every fixed }h\in\mathbb{R}^{d}, \tag*{(133)} β n k t ( h ) → ϱ h ( T ( ∞ ) ) for every fixed h ∈ R d , (133)
where ϱ h = ϕ h , Σ \varrho_{h}=\phi_{h,\Sigma} ϱ h = ϕ h , Σ .
By the finite-sample level and unbiasedness conditions, for every fixed h h h , β n ( h ) ≤ α \beta_{n}(h)\le\alpha β n ( h ) ≤ α eventually when c T h = 0 c^{T}h=0 c T h = 0 , and β n ( h ) ≥ α \beta_{n}(h)\ge\alpha β n ( h ) ≥ α eventually when c T h ≠ 0 c^{T}h\ne0 c T h = 0 . Passing to the limit in (133) gives
ϱ h ( T ( ∞ ) ) ≤ α ( c T h = 0 ) , ϱ h ( T ( ∞ ) ) ≥ α ( c T h ≠ 0 ) . \varrho_{h}(T^{(\infty)})\le\alpha\quad(c^{T}h=0),\qquad\varrho_{h}(T^{(\infty)})\ge\alpha\quad(c^{T}h\ne0). ϱ h ( T ( ∞ ) ) ≤ α ( c T h = 0 ) , ϱ h ( T ( ∞ ) ) ≥ α ( c T h = 0 ) .
Thus T ( ∞ ) T^{(\infty)} T ( ∞ ) is a level-α \alpha α unbiased test in the limiting Gaussian experiment. By Lemma F.1 , its power function h ↦ ϱ h ( T ( ∞ ) ) h\mapsto\varrho_{h}(T^{(\infty)}) h ↦ ϱ h ( T ( ∞ ) ) is continuous.
For every fixed h ∈ ⋃ k D k h\in\bigcup_{k}D_{k} h ∈ ⋃ k D k , we have h ∈ D k t h\in D_{k_{t}} h ∈ D k t eventually. Consequently,
γ ∞ = lim t → ∞ γ k t ( n k t ) = lim t → ∞ min u ∈ D k t β n k t ( u ) ≤ lim t → ∞ β n k t ( h ) = ϱ h ( T ( ∞ ) ) . \begin{aligned}
\gamma_{\infty}
&=\lim_{t\to\infty}\gamma_{k_{t}}^{(n_{k_{t}})}\\
&=\lim_{t\to\infty}\min_{u\in D_{k_{t}}}\beta_{n_{k_{t}}}(u)\\
&\le\lim_{t\to\infty}\beta_{n_{k_{t}}}(h)=\varrho_{h}(T^{(\infty)}).
\end{aligned} γ ∞ = t → ∞ lim γ k t ( n k t ) = t → ∞ lim u ∈ D k t min β n k t ( u ) ≤ t → ∞ lim β n k t ( h ) = ϱ h ( T ( ∞ ) ) .
Density and continuity extend this inequality to every h ∈ H 1 , s h\in H_{1,s} h ∈ H 1 , s . Therefore, by Theorem 3.3 ,
γ ∞ ≤ inf h ∈ H 1 , s ϱ h ( T ( ∞ ) ) ≤ b α ( s ) . \gamma_{\infty}\le\inf_{h\in H_{1,s}}\varrho_{h}(T^{(\infty)})\le b_{\alpha}(s). γ ∞ ≤ h ∈ H 1 , s inf ϱ h ( T ( ∞ ) ) ≤ b α ( s ) .
This proves the upper bound in (28) .
We next verify attainment of the Gaussian local power envelope. Consider
T α ( n ) = { ∣ ψ θ 0 X ( n ) ∣ > z α / 2 σ } . T_{\alpha}^{(n)}=\left\{\left|\psi_{\theta_{0}}X^{(n)}\right|>z_{\alpha/2}\sigma\right\}. T α ( n ) = { ψ θ 0 X ( n ) > z α /2 σ } .
By the sandwiched Le Cam third lemma under D \mathcal{D} D -extendibility [14 Corollary 3.7] , for every fixed h ∈ R d h\in\mathbb{R}^{d} h ∈ R d ,
( ψ θ 0 X ( n ) , ρ θ 0 + h / n ( n ) ) ⇒ N ( c T h , σ 2 ) . \left(\psi_{\theta_{0}}X^{(n)},\rho_{\theta_{0}+h/\sqrt{n}}^{(n)}\right)\Rightarrow N(c^{T}h,\sigma^{2}). ( ψ θ 0 X ( n ) , ρ θ 0 + h / n ( n ) ) ⇒ N ( c T h , σ 2 ) .
Since the limiting normal distribution assigns zero mass to the rejection thresholds, we obtain
β n ∗ ( h ) : = Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) → b α ( c T h ) . (134) \beta_{n}^{*}(h):=\operatorname{Tr}\left(\rho_{\theta_{0}+h/\sqrt{n}}^{(n)}T_{\alpha}^{(n)}\right)
\to b_{\alpha}(c^{T}h). \tag*{(134)} β n ∗ ( h ) := Tr ( ρ θ 0 + h / n ( n ) T α ( n ) ) → b α ( c T h ) . (134)
The function b α b_\alpha b α is even, nondecreasing in ∣ t ∣ |t| ∣ t ∣ , and satisfies b α ( 0 ) = α b_\alpha(0)=\alpha b α ( 0 ) = α . Consequently, the proposed tests have limiting rejection probability α \alpha α for every fixed h h h with c T h = 0 c^{T}h=0 c T h = 0 , and at least α \alpha α for every fixed h h h with c T h ≠ 0 c^{T}h\ne0 c T h = 0 .
For every nonempty finite D ⊂ H 1 , s D\subset H_{1,s} D ⊂ H 1 , s , finiteness and (134) give
lim n → ∞ min h ∈ D β n ⋆ ( h ) = min h ∈ D b α ( c T h ) ≥ b α ( s ) . \lim_{n\to\infty}\min_{h\in D}\beta_n^\star(h)=\min_{h\in D}b_\alpha(c^{T}h)
\ge b_\alpha(s). n → ∞ lim h ∈ D min β n ⋆ ( h ) = h ∈ D min b α ( c T h ) ≥ b α ( s ) .
Conversely, the vector h s : = s c / ∥ c ∥ 2 h_s:=sc/\|c\|^2 h s := sc /∥ c ∥ 2 belongs to H 1 , s H_{1,s} H 1 , s and satisfies c T h s = s c^{T}h_s=s c T h s = s . Taking D = { h s } D=\{h_s\} D = { h s } gives
lim n → ∞ β n ⋆ ( h s ) = b α ( s ) . \lim_{n\to\infty}\beta_n^\star(h_s)=b_\alpha(s). n → ∞ lim β n ⋆ ( h s ) = b α ( s ) .
Hence
inf ∅ ≠ D ⊂ H 1 , s ∣ D ∣ < ∞ lim sup n → ∞ inf h ∈ D β n ⋆ ( h ) = b α ( s ) , \inf_{\substack{\varnothing\ne D\subset H_{1,s}\\ |D|<\infty}}\limsup_{n\to\infty}\inf_{h\in D}\beta_n^\star(h)=b_\alpha(s), ∅ = D ⊂ H 1 , s ∣ D ∣ < ∞ inf n → ∞ lim sup h ∈ D inf β n ⋆ ( h ) = b α ( s ) ,
which proves attainment of the Gaussian local power envelope for every s > 0 s>0 s > 0 .
Proofs of the technical lemmas
Lemma F.1 (Continuity of Gaussian-shift power functions). Let S ∈ R m × m S\in\mathbb{R}^{m\times m} S ∈ R m × m be real skew-symmetric and Σ = V + i S ⪰ 0 \Sigma=V+iS\succeq0 Σ = V + i S ⪰ 0 , with V ≻ 0 V\succ0 V ≻ 0 . Let M = CCR ( S ) \mathcal{M}=\operatorname{CCR}(S) M = CCR ( S ) be the von Neumann algebra constructed in Appendix § 12 , and let B ∈ R m × d B\in\mathbb{R}^{m\times d} B ∈ R m × d . For h ∈ R d h\in\mathbb{R}^{d} h ∈ R d , let ϕ B h , Σ \phi_{Bh,\Sigma} ϕ B h , Σ denote the normal Gaussian state with mean B h Bh B h and covariance Σ \Sigma Σ . Then
lim g → h sup A ∈ M ∥ A ∥ ≤ 1 ∣ ϕ B g , Σ ( A ) − ϕ B h , Σ ( A ) ∣ = 0. (135) \lim_{g\to h}\sup_{\substack{A\in\mathcal{M}\\ \|A\|\le1}}\left|\phi_{Bg,\Sigma}(A)-\phi_{Bh,\Sigma}(A)\right|=0. \tag*{(135)} g → h lim A ∈ M ∥ A ∥ ≤ 1 sup ∣ ϕ B g , Σ ( A ) − ϕ B h , Σ ( A ) ∣ = 0. (135) In particular , h ↦ ϕ B h , Σ ( T ) h\mapsto\phi_{Bh,\Sigma}(T) h ↦ ϕ B h , Σ ( T ) is continuous for every bounded T ∈ M T\in\mathcal{M} T ∈ M , including every test 0 ≤ T ≤ I 0\le T\le I 0 ≤ T ≤ I .
Proof. Recall from Appendix § 12 that M \mathcal{M} M acts on a separable Hilbert space H ~ \widetilde{\mathcal{H}} H and that translations of the Gaussian mean are implemented by strongly continuous displacement unitaries U ( a ) U(a) U ( a ) , with
α a ( A ) = U ( a ) ∗ A U ( a ) , ϕ B h , Σ = ϕ 0 , Σ ∘ α B h . \alpha_a(A)=U(a)^*AU(a),\qquad\phi_{Bh,\Sigma}=\phi_{0,\Sigma}\circ\alpha_{Bh}. α a ( A ) = U ( a ) ∗ A U ( a ) , ϕ B h , Σ = ϕ 0 , Σ ∘ α B h . Since ϕ 0 , Σ \phi_{0,\Sigma} ϕ 0 , Σ is a normal state on the concrete von Neumann algebra M \mathcal{M} M , the normal-state representation theorem [26 Theorem 7.1.12] provides a positive trace-class operator K 0 K_0 K 0 on H ~ \widetilde{\mathcal{H}} H such that
ϕ 0 , Σ ( A ) = Tr ( K 0 A ) , A ∈ M . \phi_{0,\Sigma}(A)=\operatorname{Tr}(K_0A),\qquad A\in\mathcal{M}. ϕ 0 , Σ ( A ) = Tr ( K 0 A ) , A ∈ M . Consequently,
ϕ B h , Σ ( A ) = Tr ( U ( B h ) K 0 U ( B h ) ∗ A ) . \phi_{Bh,\Sigma}(A)=\operatorname{Tr}\left(U(Bh)K_0U(Bh)^*A\right). ϕ B h , Σ ( A ) = Tr ( U ( B h ) K 0 U ( B h ) ∗ A ) . For every A ∈ M A\in\mathcal{M} A ∈ M with ∥ A ∥ ≤ 1 \|A\|\le1 ∥ A ∥ ≤ 1 , the trace inequality ∣ Tr ( K A ) ∣ ≤ ∥ K ∥ 1 ∥ A ∥ |\operatorname{Tr}(KA)|\le\|K\|_1\|A\| ∣ Tr ( K A ) ∣ ≤ ∥ K ∥ 1 ∥ A ∥ therefore gives
∣ ϕ B g , Σ ( A ) − ϕ B h , Σ ( A ) ∣ ≤ ∥ U ( B g ) K 0 U ( B g ) ∗ − U ( B h ) K 0 U ( B h ) ∗ ∥ 1 . (136) \left|\phi_{Bg,\Sigma}(A)-\phi_{Bh,\Sigma}(A)\right|
\le\left\|U(Bg)K_0U(Bg)^*-U(Bh)K_0U(Bh)^*\right\|_1. \tag*{(136)} ∣ ϕ B g , Σ ( A ) − ϕ B h , Σ ( A ) ∣ ≤ ∥ U ( B g ) K 0 U ( B g ) ∗ − U ( B h ) K 0 U ( B h ) ∗ ∥ 1 . (136)
We show that the right-hand side tends to zero. For a rank-one operator K = ∣ u ⟩ ⟨ v ∣ K = |u\rangle\langle v| K = ∣ u ⟩ ⟨ v ∣ , the identity ∥ ∣ u ⟩ ⟨ v ∣ ∥ 1 = ∥ u ∥ ∥ v ∥ \lVert|u\rangle\langle v| \rVert_1 = \lVert u\rVert\lVert v\rVert ∥ ∣ u ⟩ ⟨ v ∣ ∥ 1 = ∥ u ∥ ∥ v ∥ and the triangle inequality yield
∥ U ( B g ) K U ( B g ) ∗ − U ( B h ) K U ( B h ) ∗ ∥ 1 ≤ ∥ ( U ( B g ) − U ( B h ) ) u ∥ ∥ v ∥ + ∥ u ∥ ∥ ( U ( B g ) − U ( B h ) ) v ∥ . \begin{aligned}
\lVert U(Bg)KU(Bg)^* - U(Bh)KU(Bh)^* \rVert_1
&\le\lVert(U(Bg)-U(Bh))u\rVert\lVert v\rVert\\
&\quad+ \lVert u\rVert\lVert(U(Bg)-U(Bh))v\rVert.
\end{aligned} ∥ U ( B g ) K U ( B g ) ∗ − U ( B h ) K U ( B h ) ∗ ∥ 1 ≤ ∥( U ( B g ) − U ( B h )) u ∥ ∥ v ∥ + ∥ u ∥ ∥( U ( B g ) − U ( B h )) v ∥ .
This tends to zero as g → h g\to h g → h by strong continuity of U U U . By linearity, the same holds for every finite-rank operator.
Finally, let K K K be finite-rank and approximate K 0 K_0 K 0 in trace norm. Unitary invariance gives
∥ U ( B g ) K 0 U ( B g ) ∗ − U ( B h ) K 0 U ( B h ) ∗ ∥ 1 ≤ 2 ∥ K 0 − K ∥ 1 + ∥ U ( B g ) K U ( B g ) ∗ − U ( B h ) K U ( B h ) ∗ ∥ 1 . \begin{aligned}
\lVert U(Bg)K_0U(Bg)^* - U(Bh)K_0U(Bh)^* \rVert_1
&\le2\lVert K_0-K\rVert_1 \\
&\quad+ \lVert U(Bg)KU(Bg)^* - U(Bh)KU(Bh)^* \rVert_1.
\end{aligned} ∥ U ( B g ) K 0 U ( B g ) ∗ − U ( B h ) K 0 U ( B h ) ∗ ∥ 1 ≤ 2 ∥ K 0 − K ∥ 1 + ∥ U ( B g ) K U ( B g ) ∗ − U ( B h ) K U ( B h ) ∗ ∥ 1 .
Since finite-rank operators are dense in the trace class, the right-hand side tends to zero as g → h g\to h g → h . Taking the supremum over ∥ A ∥ ≤ 1 \lVert A\rVert\le1 ∥ A ∥ ≤ 1 in (136) proves (135) . The asserted continuity of ϕ B h , Σ ( T ) \phi_{Bh,\Sigma}(T) ϕ B h , Σ ( T ) follows immediately.
Lemma F.2. Let P n P_n P n , n ≥ 1 n\ge1 n ≥ 1 , and P P P be probability measures on R m \mathbb{R}^m R m . For t = ( t 1 , … , t m ) ∈ R m t=(t_1,\ldots,t_m)\in\mathbb{R}^m t = ( t 1 , … , t m ) ∈ R m , write
( − ∞ , t ] : = ∏ j = 1 m ( − ∞ , t j ] . (-\infty,t] := \prod_{j=1}^{m}(-\infty,t_j]. ( − ∞ , t ] := j = 1 ∏ m ( − ∞ , t j ] . Suppose that, for every q ∈ Q m q\in\mathbb{Q}^m q ∈ Q m , the limit
a ( q ) : = lim n → ∞ P n ( ( − ∞ , q ] ) a(q):=\lim_{n\to\infty}P_n((-\infty,q]) a ( q ) := n → ∞ lim P n (( − ∞ , q ]) exists. Assume moreover that
P ( ( − ∞ , t ] ) = inf q ∈ Q m q > t a ( q ) , t ∈ R m , (137) P((-\infty,t])=\inf_{\substack{q\in\mathbb{Q}^m\\q>t}}a(q), \qquad t\in\mathbb{R}^m, \tag*{(137)} P (( − ∞ , t ]) = q ∈ Q m q > t inf a ( q ) , t ∈ R m , (137) where q > t q>t q > t means that q j > t j q_j>t_j q j > t j for every j = 1 , … , m j=1,\ldots,m j = 1 , … , m . Then
P n ⇒ P . P_n\Rightarrow P. P n ⇒ P .
Proof of Lemma F.2. Set
F n ( t ) : = P n ( ( − ∞ , t ] ) and F ( t ) : = P ( ( − ∞ , t ] ) . F_n(t):=P_n((-\infty,t]) \qquad\text{and}\qquad F(t):=P((-\infty,t]). F n ( t ) := P n (( − ∞ , t ]) and F ( t ) := P (( − ∞ , t ]) .
It suffices to prove that F n ( t ) → F ( t ) F_n(t)\to F(t) F n ( t ) → F ( t ) at every continuity point t t t of F F F . Fix such a point t t t . For every q ∈ Q m q\in\mathbb{Q}^m q ∈ Q m satisfying q > t q>t q > t , monotonicity gives F n ( t ) ≤ F n ( q ) F_n(t)\le F_n(q) F n ( t ) ≤ F n ( q ) . Therefore,
lim sup n → ∞ F n ( t ) ≤ a ( q ) . \limsup_{n\to\infty}F_n(t)\le a(q). n → ∞ lim sup F n ( t ) ≤ a ( q ) .
Taking the infimum over all rational q > t q>t q > t and using (137) , we obtain
lim sup n → ∞ F n ( t ) ≤ F ( t ) . (138) \limsup_{n\to\infty}F_n(t)\le F(t). \tag*{(138)} n → ∞ lim sup F n ( t ) ≤ F ( t ) . (138)
For ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 , let
y ℓ : = t − 1 ℓ ( 1 , … , 1 ) , y_\ell:= t-\frac{1}{\ell}(1,\ldots,1), y ℓ := t − ℓ 1 ( 1 , … , 1 ) ,
and choose p ℓ ∈ Q m p_\ell\in\mathbb{Q}^m p ℓ ∈ Q m such that
y ℓ < p ℓ < t y_\ell< p_\ell< t y ℓ < p ℓ < t
coordinatewise. Since p ℓ < t p_\ell< t p ℓ < t ,
lim inf n → ∞ F n ( t ) ≥ lim n → ∞ F n ( p ℓ ) = a ( p ℓ ) . \liminf_{n\to\infty} F_n(t) \ge\lim_{n\to\infty} F_n(p_\ell)=a(p_\ell). n → ∞ lim inf F n ( t ) ≥ n → ∞ lim F n ( p ℓ ) = a ( p ℓ ) .
Since p ℓ > y ℓ p_\ell> y_\ell p ℓ > y ℓ , (137) also gives
F ( y ℓ ) = inf q ∈ Q m q > y ℓ a ( q ) ≤ a ( p ℓ ) . F(y_\ell)=\inf_{\substack{q\in\mathbb{Q}^m\\q>y_\ell}}a(q)\le a(p_\ell). F ( y ℓ ) = q ∈ Q m q > y ℓ inf a ( q ) ≤ a ( p ℓ ) .
Consequently,
lim inf n → ∞ F n ( t ) ≥ F ( y ℓ ) ( ℓ ≥ 1 ) . \liminf_{n\to\infty}F_n(t)\ge F(y_\ell)\qquad(\ell\ge1). n → ∞ lim inf F n ( t ) ≥ F ( y ℓ ) ( ℓ ≥ 1 ) .
Because y ℓ → t y_\ell\to t y ℓ → t and t t t is a continuity point of F F F , letting ℓ → ∞ \ell\to\infty ℓ → ∞ yields
lim inf n → ∞ F n ( t ) ≥ F ( t ) . (139) \liminf_{n\to\infty}F_n(t)\ge F(t). \tag*{(139)} n → ∞ lim inf F n ( t ) ≥ F ( t ) . (139)
Combining (138) and (139) proves that F n ( t ) → F ( t ) F_n(t)\to F(t) F n ( t ) → F ( t ) at every continuity point of F F F , and hence P n ⇒ P P_n\Rightarrow P P n ⇒ P . □
Lemma F.3 (Exact-null approximation). Let ψ : Θ → R \psi:\Theta\to\mathbb{R} ψ : Θ → R be continuous in a neighborhood of θ 0 \theta_0 θ 0 and differentiable at θ 0 \theta_0 θ 0 , with ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 and c : = ψ ˙ θ 0 ≠ 0 c:=\dot{\psi}_{\theta_0}\ne0 c := ψ ˙ θ 0 = 0 . For every h ∈ R d h\in\mathbb{R}^d h ∈ R d satisfying c T h = 0 c^T h=0 c T h = 0 , there exists a sequence h n → h h_n\to h h n → h such that
ψ ( θ 0 + h n n ) = 0 \psi\left(\theta_0+\frac{h_n}{\sqrt{n}}\right)=0 ψ ( θ 0 + n h n ) = 0 for all sufficiently large $n.
Proof of Lemma F.3. Fix h ∈ R d h\in\mathbb{R}^d h ∈ R d such that c T h = 0 c^T h=0 c T h = 0 , and choose v ∈ R d v\in\mathbb{R}^d v ∈ R d such that c T v = 1 c^T v=1 c T v = 1 . For ∣ t ∣ ≤ 1 |t|\le1 ∣ t ∣ ≤ 1 , define
g n ( t ) : = n ψ ( θ 0 + h + t v n ) . g_n(t):=\sqrt{n}\,\psi\left(\theta_0+\frac{h+tv}{\sqrt{n}}\right). g n ( t ) := n ψ ( θ 0 + n h + t v ) .
By differentiability of ψ \psi ψ at θ 0 \theta_0 θ 0 ,
g n ( t ) = t + r n ( t ) , g_n(t)=t+r_n(t), g n ( t ) = t + r n ( t ) ,
where
sup ∣ t ∣ ≤ 1 ∣ r n ( t ) ∣ ⟶ 0. (140) \sup_{|t|\le1}|r_n(t)|\longrightarrow0. \tag*{(140)} ∣ t ∣ ≤ 1 sup ∣ r n ( t ) ∣ ⟶ 0. (140)
Indeed, writing
ψ ( θ 0 + u ) = c T u + R ( u ) , ∣ R ( u ) ∣ ∥ u ∥ ⟶ 0 ( u → 0 ) , \psi(\theta_0+u)=c^T u+R(u),\qquad\frac{|R(u)|}{\lVert u\rVert}\longrightarrow0\quad(u\to0), ψ ( θ 0 + u ) = c T u + R ( u ) , ∥ u ∥ ∣ R ( u ) ∣ ⟶ 0 ( u → 0 ) ,
we have, for ∣ t ∣ ≤ 1 |t|\le1 ∣ t ∣ ≤ 1 ,
r n ( t ) = n R ( h + t v n ) . r_n(t)=\sqrt{n}\,R\left(\frac{h+tv}{\sqrt{n}}\right). r n ( t ) = n R ( n h + t v ) .
Since ∥ h + t v ∥ ≤ C \lVert h+tv\rVert\le C ∥ h + t v ∥ ≤ C for ∣ t ∣ ≤ 1 |t|\le1 ∣ t ∣ ≤ 1 , it follows that
sup ∣ t ∣ ≤ 1 ∣ r n ( t ) ∣ ≤ C sup ∥ u ∥ ≤ C / n ∣ R ( u ) ∣ ∥ u ∥ → 0. \sup_{|t|\le1}|r_n(t)|\le C\sup_{\lVert u\rVert\le C/\sqrt{n}}\frac{|R(u)|}{\lVert u\rVert}\to0. ∣ t ∣ ≤ 1 sup ∣ r n ( t ) ∣ ≤ C ∥ u ∥ ≤ C / n sup ∥ u ∥ ∣ R ( u ) ∣ → 0.
Let ε n : = sup ∣ t ∣ ≤ 1 ∣ r n ( t ) ∣ \varepsilon_n:=\sup_{|t|\le1}|r_n(t)| ε n := sup ∣ t ∣ ≤ 1 ∣ r n ( t ) ∣ . For all sufficiently large n n n , 2 ε n ≤ 1 2\varepsilon_n\le1 2 ε n ≤ 1 . Moreover,
g n ( − 2 ε n ) ≤ − ε n ≤ 0 , g n ( 2 ε n ) ≥ ε n ≥ 0. g_n(-2\varepsilon_n)\le-\varepsilon_n\le0,\qquad g_n(2\varepsilon_n)\ge\varepsilon_n\ge0. g n ( − 2 ε n ) ≤ − ε n ≤ 0 , g n ( 2 ε n ) ≥ ε n ≥ 0.
By continuity, there exists t n ∈ [ − 2 ε n , 2 ε n ] t_n\in[-2\varepsilon_n,2\varepsilon_n] t n ∈ [ − 2 ε n , 2 ε n ] such that g n ( t n ) = 0 g_n(t_n)=0 g n ( t n ) = 0 . Set h n : = h + t n v h_n:=h+t_nv h n := h + t n v . Then h n → h h_n\to h h n → h , and
ψ ( θ 0 + h n n ) = 0. (141) \psi\left(\theta_0+\frac{h_n}{\sqrt{n}}\right)=0. \tag*{(141)} ψ ( θ 0 + n h n ) = 0. (141)
Trace norm bound under local perturbations
Lemma G.1 (Trace-norm bound for local product states). Let Θ ⊂ R d \Theta\subset\mathbb{R}^d Θ ⊂ R d be open and let { ρ ϑ : ϑ ∈ Θ } \{\rho_\vartheta:\vartheta\in\Theta\} { ρ ϑ : ϑ ∈ Θ } be a quantum statistical model on a finite-dimensional Hilbert space H \mathcal{H} H . Assume that ϑ ↦ ρ ϑ \vartheta\mapsto\rho_\vartheta ϑ ↦ ρ ϑ is continuously differentiable in trace norm. Let L j ( ϑ ) : = L ϑ , j L_j(\vartheta):=L_{\vartheta,j} L j ( ϑ ) := L ϑ , j denote the SLDs defined in (12) , and let
G ϑ = ( G ϑ , i j ) i , j = 1 d G_\vartheta=(G_{\vartheta,ij})_{i,j=1}^{d} G ϑ = ( G ϑ , ij ) i , j = 1 d be the SLD Fisher information matrix defined in (14) . Assume that ϑ ↦ G ϑ \vartheta\mapsto G_\vartheta ϑ ↦ G ϑ is continuous. For h ∈ R d h\in\mathbb{R}^d h ∈ R d , write
L h ( ϑ ) : = ∑ j = 1 d h j L j ( ϑ ) , ρ ϑ ( n ) : = ρ ϑ ⊗ n . L_h(\vartheta):=\sum_{j=1}^{d}h_jL_j(\vartheta),\qquad\rho_\vartheta^{(n)}:=\rho_\vartheta^{\otimes n}. L h ( ϑ ) := j = 1 ∑ d h j L j ( ϑ ) , ρ ϑ ( n ) := ρ ϑ ⊗ n . For every θ ∈ Θ \theta\in\Theta θ ∈ Θ , h ∈ R d h\in\mathbb{R}^d h ∈ R d , and integer n ≥ 1 n\ge1 n ≥ 1 such that θ + t h / n ∈ Θ \theta+th/\sqrt{n}\in\Theta θ + t h / n ∈ Θ for all t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] ,
∥ ρ θ + h / n ( n ) − ρ θ ( n ) ∥ 1 ≤ ∫ 0 1 h T G θ + t h / n h d t ≤ sup 0 ≤ t ≤ 1 h T G θ + t h / n h . (142) \left\|\rho_{\theta+h/\sqrt{n}}^{(n)}-\rho_\theta^{(n)}\right\|_1
\le\int_0^1\sqrt{h^{\mathsf T}G_{\theta+th/\sqrt{n}}h}\,\mathrm{d}t
\le\sup_{0\le t\le1}\sqrt{h^{\mathsf T}G_{\theta+th/\sqrt{n}}h}. \tag*{(142)} ρ θ + h / n ( n ) − ρ θ ( n ) 1 ≤ ∫ 0 1 h T G θ + t h / n h d t ≤ 0 ≤ t ≤ 1 sup h T G θ + t h / n h . (142) In particular, for fixed θ \theta θ and h h h ,
lim sup n → ∞ ∥ ρ θ + h / n ( n ) − ρ θ ( n ) ∥ 1 ≤ h T G θ h . (143) \limsup_{n\to\infty}\left\|\rho_{\theta+h/\sqrt{n}}^{(n)}-\rho_\theta^{(n)}\right\|_1\le\sqrt{h^{\mathsf T}G_\theta h}. \tag*{(143)} n → ∞ lim sup ρ θ + h / n ( n ) − ρ θ ( n ) 1 ≤ h T G θ h . (143) Since h h h is real, h T J ϑ h = h T G ϑ h h^{\mathsf T}J_\vartheta h=h^{\mathsf T}G_\vartheta h h T J ϑ h = h T G ϑ h , so G ϑ G_\vartheta G ϑ may equivalently be replaced by J ϑ J_\vartheta J ϑ inside these quadratic forms.
Proof. Taking the trace in the SLD equation and differentiating Tr ( ρ ϑ ) = 1 \operatorname{Tr}(\rho_\vartheta)=1 Tr ( ρ ϑ ) = 1 gives
Tr ( ρ ϑ L j ( ϑ ) ) = Tr ( ∂ j ρ ϑ ) = 0. \operatorname{Tr}(\rho_\vartheta L_j(\vartheta))=\operatorname{Tr}(\partial_j\rho_\vartheta)=0. Tr ( ρ ϑ L j ( ϑ )) = Tr ( ∂ j ρ ϑ ) = 0.
Consequently,
Tr ( ρ ϑ L h ( ϑ ) ) = 0 , Tr ( ρ ϑ L h ( ϑ ) 2 ) = h T G ϑ h = h T J ϑ h . (144) \operatorname{Tr}(\rho_\vartheta L_h(\vartheta))=0,\qquad\operatorname{Tr}(\rho_\vartheta L_h(\vartheta)^2)=h^{\mathsf T}G_\vartheta h=h^{\mathsf T}J_\vartheta h. \tag*{(144)} Tr ( ρ ϑ L h ( ϑ )) = 0 , Tr ( ρ ϑ L h ( ϑ ) 2 ) = h T G ϑ h = h T J ϑ h . (144)
The second identity follows by expanding L h ( ϑ ) 2 L_h(\vartheta)^2 L h ( ϑ ) 2 . In particular, G ϑ G_\vartheta G ϑ is real symmetric and positive semidefinite.
We first give a trace-norm estimate. For any density matrix σ \sigma σ and self-adjoint matrix B B B , the triangle inequality, invariance of the trace norm under taking adjoints, and ∥ C D ∥ 1 ≤ ∥ C ∥ 2 ∥ D ∥ 2 \lVert CD\rVert_1 \le\lVert C\rVert_2\lVert D\rVert_2 ∥ C D ∥ 1 ≤ ∥ C ∥ 2 ∥ D ∥ 2 give
1 2 ∥ σ B + B σ ∥ 1 ≤ ∥ σ B ∥ 1 = ∥ σ 1 / 2 ( σ 1 / 2 B ) ∥ 1 ≤ ∥ σ 1 / 2 ∥ 2 ∥ σ 1 / 2 B ∥ 2 = Tr ( σ B 2 ) , (145) \begin{aligned}
\frac{1}{2}\lVert\sigma B+B\sigma\rVert_1 \le\lVert\sigma B\rVert_1 \\
&= \lVert\sigma^{1/2}(\sigma^{1/2}B)\rVert_1 \\
&\le\lVert\sigma^{1/2}\rVert_2\lVert\sigma^{1/2}B\rVert_2 = \sqrt{\operatorname{Tr}(\sigma B^2)}, \tag*{(145)}
\end{aligned} 2 1 ∥ σ B + B σ ∥ 1 ≤ ∥ σ B ∥ 1 = ∥ σ 1/2 ( σ 1/2 B ) ∥ 1 ≤ ∥ σ 1/2 ∥ 2 ∥ σ 1/2 B ∥ 2 = Tr ( σ B 2 ) , (145)
where ∥ C ∥ 2 : = { Tr ( C ∗ C ) } 1 / 2 \lVert C\rVert_2 := \{\operatorname{Tr}(C^*C)\}^{1/2} ∥ C ∥ 2 := { Tr ( C ∗ C ) } 1/2 is the Hilbert--Schmidt norm.
Fix θ \theta θ , h h h , n n n as in the statement, and put θ t = θ + t h / n \theta_t=\theta+th/\sqrt{n} θ t = θ + t h / n and σ t = ρ θ t ⊗ n \sigma_t=\rho_{\theta_t}^{\otimes n} σ t = ρ θ t ⊗ n . Continuous differentiability of the model implies that t ↦ σ t t\mapsto\sigma_t t ↦ σ t is continuously differentiable in trace norm. The fundamental theorem of calculus therefore gives
ρ θ + h / n ( n ) − ρ θ ( n ) = ∫ 0 1 σ ˙ t d t . (146) \rho_{\theta+h/\sqrt{n}}^{(n)}-\rho_\theta^{(n)}=\int_0^1\dot{\sigma}_t\,dt. \tag*{(146)} ρ θ + h / n ( n ) − ρ θ ( n ) = ∫ 0 1 σ ˙ t d t . (146)
Define the collective directional SLD by
L h ( n ) ( ϑ ) : = ∑ a = 1 n L h [ a ] ( ϑ ) , L h [ a ] ( ϑ ) : = I ⊗ ( a − 1 ) ⊗ L h ( ϑ ) ⊗ I ⊗ ( n − a ) . L_h^{(n)}(\vartheta):=\sum_{a=1}^{n}L_h^{[a]}(\vartheta),\qquad L_h^{[a]}(\vartheta):=I^{\otimes(a-1)}\otimes L_h(\vartheta)\otimes I^{\otimes(n-a)}. L h ( n ) ( ϑ ) := a = 1 ∑ n L h [ a ] ( ϑ ) , L h [ a ] ( ϑ ) := I ⊗ ( a − 1 ) ⊗ L h ( ϑ ) ⊗ I ⊗ ( n − a ) .
The chain rule gives
d d t ρ θ t = 1 2 n ( ρ θ t L h ( θ t ) + L h ( θ t ) ρ θ t ) . \frac{d}{dt}\rho_{\theta_t}=\frac{1}{2\sqrt{n}}\left(\rho_{\theta_t}L_h(\theta_t)+L_h(\theta_t)\rho_{\theta_t}\right). d t d ρ θ t = 2 n 1 ( ρ θ t L h ( θ t ) + L h ( θ t ) ρ θ t ) .
Applying the product rule to the n n n tensor factors consequently yields
σ ˙ t = 1 2 n ( σ t L h ( n ) ( θ t ) + L h ( n ) ( θ t ) σ t ) . (147) \dot{\sigma}_t=\frac{1}{2\sqrt{n}}\left(\sigma_tL_h^{(n)}(\theta_t)+L_h^{(n)}(\theta_t)\sigma_t\right). \tag*{(147)} σ ˙ t = 2 n 1 ( σ t L h ( n ) ( θ t ) + L h ( n ) ( θ t ) σ t ) . (147)
Combining (146) , (147) , and (145) , we obtain
∥ ρ θ + h / n ( n ) − ρ θ ( n ) ∥ 1 ≤ 1 n ∫ 0 1 Tr [ σ t ( L h ( n ) ( θ t ) ) 2 ] d t . (148) \left\lVert\rho_{\theta+h/\sqrt{n}}^{(n)}-\rho_\theta^{(n)}\right\rVert_1\le\frac{1}{\sqrt{n}}\int_0^1\sqrt{\operatorname{Tr}\left[\sigma_t\left(L_h^{(n)}(\theta_t)\right)^2\right]}\,dt. \tag*{(148)} ρ θ + h / n ( n ) − ρ θ ( n ) 1 ≤ n 1 ∫ 0 1 Tr [ σ t ( L h ( n ) ( θ t ) ) 2 ] d t . (148)
To evaluate the integrand, expand the square of the collective SLD. Each diagonal term has expectation Tr ( ρ θ t L h ( θ t ) 2 ) \operatorname{Tr}(\rho_{\theta_t}L_h(\theta_t)^2) Tr ( ρ θ t L h ( θ t ) 2 ) . For a ≠ b a\ne b a = b , the tensor-product structure and (144) give
Tr [ σ t L h [ a ] ( θ t ) L h [ b ] ( θ t ) ] = [ Tr ( ρ θ t L h ( θ t ) ) ] 2 = 0. \operatorname{Tr}\left[\sigma_tL_h^{[a]}(\theta_t)L_h^{[b]}(\theta_t)\right]=\left[\operatorname{Tr}(\rho_{\theta_t}L_h(\theta_t))\right]^2=0. Tr [ σ t L h [ a ] ( θ t ) L h [ b ] ( θ t ) ] = [ Tr ( ρ θ t L h ( θ t )) ] 2 = 0.
Hence
Tr [ σ t ( L h ( n ) ( θ t ) ) 2 ] = n Tr ( ρ θ t L h ( θ t ) 2 ) = n h ⊤ G θ t h . \operatorname{Tr}\left[\sigma_t\left(L_h^{(n)}(\theta_t)\right)^2\right]=n\operatorname{Tr}\left(\rho_{\theta_t}L_h(\theta_t)^2\right)=nh^\top G_{\theta_t}h. Tr [ σ t ( L h ( n ) ( θ t ) ) 2 ] = n Tr ( ρ θ t L h ( θ t ) 2 ) = n h ⊤ G θ t h .
Substituting into (148) proves the integral estimate in (142) . The supremum estimate follows because the interval of integration has length one.
Finally, for fixed θ \theta θ , h h h , the path lies in Θ \Theta Θ for all sufficiently large n n n , and sup 0 ≤ t ≤ 1 ∥ θ t − θ ∥ ≤ ∥ h ∥ / n → 0 \sup_{0\le t\le1}\lVert\theta_t-\theta\rVert\le\lVert h\rVert/\sqrt{n}\to0 sup 0 ≤ t ≤ 1 ∥ θ t − θ ∥ ≤ ∥ h ∥ / n → 0 . Continuity of G ϑ G_\vartheta G ϑ at θ \theta θ therefore implies
sup 0 ≤ t ≤ 1 ∥ G θ t − G θ ∥ o p → 0 , \sup_{0\le t\le1}\lVert G_{\theta_t}-G_\theta\rVert_{\mathrm{op}}\to0, 0 ≤ t ≤ 1 sup ∥ G θ t − G θ ∥ op → 0 ,
where ∥ ⋅ ∥ o p \lVert\cdot\rVert_{\mathrm{op}} ∥ ⋅ ∥ op is the matrix operator norm. It follows that sup 0 ≤ t ≤ 1 h ⊤ G θ t h → h ⊤ G θ h \sup_{0\le t\le1}\sqrt{h^\top G_{\theta_t}h}\to\sqrt{h^\top G_\theta h} sup 0 ≤ t ≤ 1 h ⊤ G θ t h → h ⊤ G θ h , proving (143) .
Corollary G.2 (Comparison of moving local parameters). Under the assumptions of Lemma G.1, fix θ 0 ∈ Θ \theta_0 \in\Theta θ 0 ∈ Θ and let ( h n ) (h_n) ( h n ) and ( k n ) (k_n) ( k n ) be bounded sequences in R d \mathbb{R}^d R d . Put
δ n = h n − k n , ϑ n , t = θ 0 + k n + t δ n n , 0 ≤ t ≤ 1. \delta_n = h_n-k_n,\qquad\vartheta_{n,t}=\theta_0+\frac{k_n+t\delta_n}{\sqrt{n}},\qquad0\leq t\leq1. δ n = h n − k n , ϑ n , t = θ 0 + n k n + t δ n , 0 ≤ t ≤ 1. For all sufficiently large n n n ,
∥ ρ θ 0 + h n / n ( n ) − ρ θ 0 + k n / n ( n ) ∥ 1 ≤ ∫ 0 1 δ n T G ϑ n , t δ n d t ≤ ∥ δ n ∥ 0 sup 0 ≤ t ≤ 1 ∥ G ϑ n , t ∥ o p 1 / 2 . (149) \left\|\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}-\rho_{\theta_0+k_n/\sqrt{n}}^{(n)}\right\|_1
\leq\int_0^1\sqrt{\delta_n^{\mathsf{T}}G_{\vartheta_{n,t}}\delta_n}\,\mathrm{d}t
\leq\|\delta_n\|_0\sup_{0\leq t\leq1}\|G_{\vartheta_{n,t}}\|_{\mathrm{op}}^{1/2}. \tag*{(149)} ρ θ 0 + h n / n ( n ) − ρ θ 0 + k n / n ( n ) 1 ≤ ∫ 0 1 δ n T G ϑ n , t δ n d t ≤ ∥ δ n ∥ 0 0 ≤ t ≤ 1 sup ∥ G ϑ n , t ∥ op 1/2 . (149) Consequently, if h n − k n → 0 h_n-k_n\to0 h n − k n → 0 , then
∥ ρ θ 0 + h n / n ( n ) − ρ θ 0 + k n / n ( n ) ∥ 1 → 0. (150) \left\|\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}-\rho_{\theta_0+k_n/\sqrt{n}}^{(n)}\right\|_1\to0. \tag*{(150)} ρ θ 0 + h n / n ( n ) − ρ θ 0 + k n / n ( n ) 1 → 0. (150) The convergence of rejection probabilities is uniform over all tests:
sup 0 ≤ T ≤ I H ⊗ n ∣ Tr ( ρ θ 0 + h n / n ( n ) T ) − Tr ( ρ θ 0 + k n / n ( n ) T ) ∣ → 0. \sup_{0\leq T\leq I_{\mathcal{H}\otimes n}}\left|\operatorname{Tr}\left(\rho_{\theta_0+h_n/\sqrt{n}}^{(n)}T\right)-\operatorname{Tr}\left(\rho_{\theta_0+k_n/\sqrt{n}}^{(n)}T\right)\right|\to0. 0 ≤ T ≤ I H ⊗ n sup Tr ( ρ θ 0 + h n / n ( n ) T ) − Tr ( ρ θ 0 + k n / n ( n ) T ) → 0. Proof. Boundedness of h n h_n h n and k n k_n k n gives
sup 0 ≤ t ≤ 1 ∥ ϑ n , t − θ 0 ∥ ≤ max { ∥ h n ∥ , ∥ k n ∥ } n → 0. \sup_{0\leq t\leq1}\|\vartheta_{n,t}-\theta_0\|\leq\frac{\max\{\|h_n\|,\|k_n\|\}}{\sqrt{n}}\to0. 0 ≤ t ≤ 1 sup ∥ ϑ n , t − θ 0 ∥ ≤ n max { ∥ h n ∥ , ∥ k n ∥ } → 0. Since Θ \Theta Θ is open, the entire segment { ϑ n , t : 0 ≤ t ≤ 1 } \{\vartheta_{n,t}:0\leq t\leq1\} { ϑ n , t : 0 ≤ t ≤ 1 } lies in Θ \Theta Θ for all sufficiently large n n n . Apply Lemma G.1 with basepoint θ = θ 0 + k n / n \theta=\theta_0+k_n/\sqrt{n} θ = θ 0 + k n / n and local displacement h = δ n h=\delta_n h = δ n . This gives the first inequality in (148); the second follows from
δ n T G ϑ n , t δ n ≤ ∥ G ϑ n , t ∥ o p ∥ δ n ∥ 2 . \delta_n^{\mathsf{T}}G_{\vartheta_{n,t}}\delta_n\leq\|G_{\vartheta_{n,t}}\|_{\mathrm{op}}\|\delta_n\|^2. δ n T G ϑ n , t δ n ≤ ∥ G ϑ n , t ∥ op ∥ δ n ∥ 2 . Continuity of G G G at θ 0 \theta_0 θ 0 implies
sup 0 ≤ t ≤ 1 ∥ G ϑ n , t − G θ 0 ∥ o p → 0. \sup_{0\leq t\leq1}\|G_{\vartheta_{n,t}}-G_{\theta_0}\|_{\mathrm{op}}\to0. 0 ≤ t ≤ 1 sup ∥ G ϑ n , t − G θ 0 ∥ op → 0. The supremum on the right-hand side of (148) is therefore bounded. If δ n → 0 \delta_n\to0 δ n → 0 , this proves (149). Finally, for any effect 0 ≤ T ≤ I 0\leq T\leq I 0 ≤ T ≤ I ,
∣ Tr ( ( ρ − σ ) T ) ∣ ≤ ∥ ρ − σ ∥ 1 ∥ T ∥ o p ≤ ∥ ρ − σ ∥ 1 . \left|\operatorname{Tr}((\rho-\sigma)T)\right|\leq\|\rho-\sigma\|_1\|T\|_{\mathrm{op}}\leq\|\rho-\sigma\|_1. ∣ Tr (( ρ − σ ) T ) ∣ ≤ ∥ ρ − σ ∥ 1 ∥ T ∥ op ≤ ∥ ρ − σ ∥ 1 . Applying this inequality to the two product states proves the last assertion.
Proofs for the adaptive scalar Wald tests
Regularity and the one-step identities
Let Q η Q_\eta Q η be the orthogonal projection onto ker ρ η \ker\rho_\eta ker ρ η . Constant rank and the C 2 C^2 C 2 assumption imply that Q η Q_\eta Q η is C 2 C^2 C 2 . Differentiating ρ η Q η = 0 \rho_\eta Q_\eta=0 ρ η Q η = 0 gives Q η ( ∂ j ρ η ) Q η = 0 Q_\eta(\partial_j\rho_\eta)Q_\eta=0 Q η ( ∂ j ρ η ) Q η = 0 . The C 2 C^2 C 2 family of linear maps on Hermitian matrices
T η ( B ) : = ρ η B + B ρ η 2 + Q η B Q η T_\eta(B):=\frac{\rho_\eta B+B\rho_\eta}{2}+Q_\eta BQ_\eta T η ( B ) := 2 ρ η B + B ρ η + Q η B Q η
is invertible: in an eigenbasis of ρ η \rho_\eta ρ η , it multiplies each matrix entry by ( λ a + λ b ) / 2 (\lambda_a+\lambda_b)/2 ( λ a + λ b ) /2 if at least one eigenvalue is positive, and by 1 1 1 otherwise. Thus L η , j ( 0 ) : = T η − 1 ( ∂ j ρ η ) L_{\eta,j}^{(0)}:=\mathcal{T}_{\eta}^{-1}(\partial_j\rho_\eta) L η , j ( 0 ) := T η − 1 ( ∂ j ρ η ) is a C 1 C^1 C 1 choice of SLDs. Any other SLD differs from this one only on the kernel block, which does not affect J η J_\eta J η or G η G_\eta G η . For any continuous SLD choice as in Subsection 5.1, G η G_\eta G η and Y η Y_\eta Y η are therefore continuous, and C Y : = sup η ∈ K ∥ Y η ∥ o p < ∞ C_Y:=\sup_{\eta\in K}\|Y_\eta\|_{\mathrm{op}}<\infty C Y := sup η ∈ K ∥ Y η ∥ op < ∞ .
All derivatives and bounds used below are therefore ordinary finite-dimensional ones. Taking traces in (12) and differentiating Tr ρ η = 1 \operatorname{Tr}\rho_\eta=1 Tr ρ η = 1 gives Tr ( ρ η L η , j ) = 0 \operatorname{Tr}(\rho_\eta L_{\eta,j})=0 Tr ( ρ η L η , j ) = 0 and hence μ ( η , η ) = 0 \mu(\eta,\eta)=0 μ ( η , η ) = 0 . Holding η \eta η fixed and differentiating only the true-state argument, cyclicity of the trace yields
∂ ϑ j μ ( ϑ , η ) ∣ ϑ = η = Tr { ( ∂ j ρ η ) Y η } = 1 2 Tr { ρ η ( L η , j Y η + Y η L η , j ) } = ( G η v η ) j = ( c η ) j . \left.\partial_{\vartheta_j}\mu(\vartheta,\eta)\right|_{\vartheta=\eta}
=\operatorname{Tr}\{(\partial_j\rho_\eta)Y_\eta\}
=\frac{1}{2}\operatorname{Tr}\{\rho_\eta(L_{\eta,j}Y_\eta+Y_\eta L_{\eta,j})\}
=(G_\eta v_\eta)_j=(c_\eta)_j. ∂ ϑ j μ ( ϑ , η ) ϑ = η = Tr {( ∂ j ρ η ) Y η } = 2 1 Tr { ρ η ( L η , j Y η + Y η L η , j )} = ( G η v η ) j = ( c η ) j .
Likewise,
Tr ( ρ η Y η 2 ) = v η T G η v η = c η T G η − 1 c η . \operatorname{Tr}(\rho_\eta Y_\eta^2)=v_\eta^{T}G_\eta v_\eta=c_\eta^{T}G_\eta^{-1}c_\eta. Tr ( ρ η Y η 2 ) = v η T G η v η = c η T G η − 1 c η .
This proves (32) .
To establish (33) , set B ( ϑ , η ) = ψ ( η ) + μ ( ϑ , η ) − ψ ( ϑ ) B(\vartheta,\eta)=\psi(\eta)+\mu(\vartheta,\eta)-\psi(\vartheta) B ( ϑ , η ) = ψ ( η ) + μ ( ϑ , η ) − ψ ( ϑ ) . The identities just proved give B ( η , η ) = 0 B(\eta,\eta)=0 B ( η , η ) = 0 and ∇ ϑ B ( η , η ) = 0 \nabla_\vartheta B(\eta,\eta)=0 ∇ ϑ B ( η , η ) = 0 . On a closed ball contained in Θ \Theta Θ , both second derivatives of ρ ϑ \rho_\vartheta ρ ϑ and of ψ ( ϑ ) \psi(\vartheta) ψ ( ϑ ) are bounded. Moreover,
D ϑ 2 μ ( ϑ , η ) [ u , u ] = Tr { D 2 ρ ϑ [ u , u ] Y η } , ∣ D ϑ 2 μ ( ϑ , η ) [ u , u ] ∣ ≤ ∥ D 2 ρ ϑ [ u , u ] ∥ 1 ∥ Y η ∥ o p . D_\vartheta^2\mu(\vartheta,\eta)[u,u]=\operatorname{Tr}\{D^2\rho_\vartheta[u,u]Y_\eta\},\qquad
\left|D_\vartheta^2\mu(\vartheta,\eta)[u,u]\right|\leq\|D^2\rho_\vartheta[u,u]\|_1\|Y_\eta\|_{\mathrm{op}}. D ϑ 2 μ ( ϑ , η ) [ u , u ] = Tr { D 2 ρ ϑ [ u , u ] Y η } , D ϑ 2 μ ( ϑ , η ) [ u , u ] ≤ ∥ D 2 ρ ϑ [ u , u ] ∥ 1 ∥ Y η ∥ op .
Hence ∣ D ϑ 2 B ( ϑ , η ) [ u , u ] ∣ ≤ C 1 ∥ u ∥ 2 \left|D_\vartheta^2B(\vartheta,\eta)[u,u]\right|\leq C_1\|u\|^2 D ϑ 2 B ( ϑ , η ) [ u , u ] ≤ C 1 ∥ u ∥ 2 uniformly for both parameters in a sufficiently small closed ball. Writing δ = ϑ − η \delta=\vartheta-\eta δ = ϑ − η , Taylor’s formula along the segment gives
B ( ϑ , η ) = ∫ 0 1 ( 1 − t ) D ϑ 2 B ( η + t δ , η ) [ δ , δ ] d t , ∣ B ( ϑ , η ) ∣ ≤ C 1 2 ∥ δ ∥ 2 . B(\vartheta,\eta)=\int_0^1(1-t)D_\vartheta^2B(\eta+t\delta,\eta)[\delta,\delta]\,dt,\qquad
|B(\vartheta,\eta)|\leq\frac{C_1}{2}\|\delta\|^2. B ( ϑ , η ) = ∫ 0 1 ( 1 − t ) D ϑ 2 B ( η + t δ , η ) [ δ , δ ] d t , ∣ B ( ϑ , η ) ∣ ≤ 2 C 1 ∥ δ ∥ 2 .
No derivative of Y η Y_\eta Y η or of its spectral projections is taken in this expansion.
Proof of Proposition 5.1
Proof of Proposition 5.1. We verify that the POVM in (36) induces a regular identifiable finite-dimensional classical model, after which the result follows from standard maximum-likelihood theory.
First, since ∥ H j ∥ o p ≤ 1 \|H_j\|_{\mathrm{op}}\leq1 ∥ H j ∥ op ≤ 1 and ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) ,
I ± ε H j ⪰ ( 1 − ε ) I . I\pm\varepsilon H_j\succeq(1-\varepsilon)I. I ± ε H j ⪰ ( 1 − ε ) I .
Hence every M j , ± p i l M_{j,\pm}^{\mathrm{pil}} M j , ± pil is strictly positive. Moreover,
M j , + p i l + M j , − p i l = I s , M_{j,+}^{\mathrm{pil}}+M_{j,-}^{\mathrm{pil}}=\frac{I}{s}, M j , + pil + M j , − pil = s I ,
and therefore
∑ j = 1 s ( M j , + p i l + M j , − p i l ) = I . \sum_{j=1}^{s}\left(M_{j,+}^{\mathrm{pil}}+M_{j,-}^{\mathrm{pil}}\right)=I. j = 1 ∑ s ( M j , + pil + M j , − pil ) = I .
Thus { M j , ± p i l } \{M_{j,\pm}^{\mathrm{pil}}\} { M j , ± pil } is a POVM. Furthermore,
M j , + p i l − M j , − p i l = ε s H j . M_{j,+}^{\mathrm{pil}}-M_{j,-}^{\mathrm{pil}}=\frac{\varepsilon}{s}H_j. M j , + pil − M j , − pil = s ε H j .
Since H 1 , … , H s H_1,\ldots,H_s H 1 , … , H s form a real basis of the traceless Hermitian matrices, the real span of the POVM effects contains I I I and every H j H_j H j , and therefore equals the space of all Hermitian matrices. Hence the POVM is informationally complete.
For x ∈ X x \in\mathcal{X} x ∈ X , write
p η ( x ) = Tr ( ρ η M x p i l ) . p_\eta(x) = \operatorname{Tr}\left(\rho_\eta M_x^{\mathrm{pil}}\right). p η ( x ) = Tr ( ρ η M x pil ) .
Strict positivity of the effects gives the uniform lower bound
p η ( x ) ≥ 1 − ε 2 s , η ∈ K , x ∈ X . p_\eta(x) \ge\frac{1-\varepsilon}{2s}, \qquad\eta\in K,\quad x \in\mathcal{X}. p η ( x ) ≥ 2 s 1 − ε , η ∈ K , x ∈ X .
Since η ↦ ρ η \eta\mapsto\rho_\eta η ↦ ρ η is C 2 C^2 C 2 in trace norm, each η ↦ p η ( x ) \eta\mapsto p_\eta(x) η ↦ p η ( x ) is C 2 C^2 C 2 . In particular, log p η ( x ) \log p_\eta(x) log p η ( x ) and its first two derivatives are continuous and bounded on compact subsets of int K \operatorname{int} K int K .
The induced classical model is identifiable on K K K . Indeed, suppose that p η ( x ) = p ϑ ( x ) p_\eta(x) = p_\vartheta(x) p η ( x ) = p ϑ ( x ) for every x ∈ X x \in\mathcal{X} x ∈ X . Then
Tr [ ( ρ η − ρ ϑ ) M x p i l ] = 0 for every x . \operatorname{Tr}\left[(\rho_\eta-\rho_\vartheta)M_x^{\mathrm{pil}}\right] = 0 \qquad\text{for every }x. Tr [ ( ρ η − ρ ϑ ) M x pil ] = 0 for every x .
Since the POVM effects span the Hermitian matrices, this implies ρ η = ρ ϑ \rho_\eta= \rho_\vartheta ρ η = ρ ϑ . By the assumed injectivity of the parametrization on K K K , η = ϑ \eta= \vartheta η = ϑ .
We next verify nonsingularity of the classical Fisher information. For a ∈ R d a \in\mathbb{R}^d a ∈ R d ,
a T I θ p i l a = ∑ x ∈ X { a T ∇ p θ ( x ) } 2 p θ ( x ) . a^{T}I_\theta^{\mathrm{pil}}a = \sum_{x\in\mathcal{X}} \frac{\left\{a^{T}\nabla p_\theta(x)\right\}^{2}}{p_\theta(x)}. a T I θ pil a = x ∈ X ∑ p θ ( x ) { a T ∇ p θ ( x ) } 2 .
If this quantity is zero, strict positivity of all p θ ( x ) p_\theta(x) p θ ( x ) implies
a T ∇ p θ ( x ) = Tr [ ( ∂ a ρ θ ) M x p i l ] = 0 for every x , a^{T}\nabla p_\theta(x) = \operatorname{Tr}\left[(\partial_a\rho_\theta)M_x^{\mathrm{pil}}\right] = 0 \qquad\text{for every }x, a T ∇ p θ ( x ) = Tr [ ( ∂ a ρ θ ) M x pil ] = 0 for every x ,
where ∂ a ρ θ : = ∑ j = 1 d a j ∂ j ρ θ \partial_a\rho_\theta:= \sum_{j=1}^{d} a_j\partial_j\rho_\theta ∂ a ρ θ := ∑ j = 1 d a j ∂ j ρ θ . Informational completeness then gives ∂ a ρ θ = 0 \partial_a\rho_\theta= 0 ∂ a ρ θ = 0 . If
L θ , a : = ∑ j = 1 d a j L θ , j , L_{\theta,a} := \sum_{j=1}^{d} a_jL_{\theta,j}, L θ , a := j = 1 ∑ d a j L θ , j ,
the SLD equations imply
∂ a ρ θ = 1 2 ( ρ θ L θ , a + L θ , a ρ θ ) . \partial_a\rho_\theta= \frac{1}{2}\left(\rho_\theta L_{\theta,a} + L_{\theta,a}\rho_\theta\right). ∂ a ρ θ = 2 1 ( ρ θ L θ , a + L θ , a ρ θ ) .
Multiplying by L θ , a L_{\theta,a} L θ , a and taking the trace gives
a T G θ a = Tr ( ρ θ L θ , a 2 ) = Tr { ( ∂ a ρ θ ) L θ , a } = 0. a^{T}G_\theta a = \operatorname{Tr}\left(\rho_\theta L_{\theta,a}^{2}\right) = \operatorname{Tr}\left\{(\partial_a\rho_\theta)L_{\theta,a}\right\} = 0. a T G θ a = Tr ( ρ θ L θ , a 2 ) = Tr { ( ∂ a ρ θ ) L θ , a } = 0.
Because G θ ≻ 0 G_\theta\succ0 G θ ≻ 0 , we conclude that a = 0 a=0 a = 0 . Thus I θ p i l ≻ 0 I_\theta^{\mathrm{pil}} \succ0 I θ pil ≻ 0 . It remains only to invoke standard likelihood theory. For completeness, let
M m ( η ) : = 1 m ∑ a = 1 m log p η ( W a ) , M ( η ) : = ∑ x ∈ X p θ ( x ) log p η ( x ) . M_m(\eta) := \frac{1}{m}\sum_{a=1}^{m}\log p_\eta(W_a), \qquad M(\eta) := \sum_{x\in\mathcal{X}}p_\theta(x)\log p_\eta(x). M m ( η ) := m 1 a = 1 ∑ m log p η ( W a ) , M ( η ) := x ∈ X ∑ p θ ( x ) log p η ( x ) .
Because X \mathcal{X} X is finite and inf η ∈ K , x p η ( x ) > 0 \inf_{\eta\in K,x}p_\eta(x)>0 inf η ∈ K , x p η ( x ) > 0 , the strong law for the empirical frequencies gives
sup η ∈ K ∣ M m ( η ) − M ( η ) ∣ ⟶ 0 almost surely . \sup_{\eta\in K}\left|M_m(\eta)-M(\eta)\right| \longrightarrow0 \qquad\text{almost surely}. η ∈ K sup ∣ M m ( η ) − M ( η ) ∣ ⟶ 0 almost surely .
Moreover,
M ( θ ) − M ( η ) = ∑ x ∈ X p θ ( x ) log p θ ( x ) p η ( x ) ≥ 0 , M(\theta)-M(\eta)=\sum_{x\in\mathcal{X}}p_{\theta}(x)\log\frac{p_{\theta}(x)}{p_{\eta}(x)}\geq0, M ( θ ) − M ( η ) = x ∈ X ∑ p θ ( x ) log p η ( x ) p θ ( x ) ≥ 0 ,
with equality only when p η = p θ p_{\eta}=p_{\theta} p η = p θ , and hence, by identifiability, only when η = θ \eta=\theta η = θ . Thus the population criterion has the unique maximizer θ \theta θ . Since K K K is compact, the consistency theorem for M M M -estimators [44 Theorem 5.7] yields
θ ~ n → θ in probability . \widetilde{\theta}_{n}\to\theta\quad\text{in probability}. θ n → θ in probability .
Finally, choose a closed ball U ⊂ int K U\subset\operatorname{int}K U ⊂ int K with θ \theta θ in its interior. Consistency implies that θ ~ n ∈ U \widetilde{\theta}_{n}\in U θ n ∈ U with probability tending to one. On U U U , the induced finite-outcome model
P p i l = { p η : η ∈ U } , p η ( x ) = Tr ( ρ η M x p i l ) , \mathcal{P}^{\mathrm{pil}}=\{p_{\eta}:\eta\in U\},\qquad p_{\eta}(x)=\operatorname{Tr}(\rho_{\eta}M_{x}^{\mathrm{pil}}), P pil = { p η : η ∈ U } , p η ( x ) = Tr ( ρ η M x pil ) ,
is a regular dominated parametric model with common support X \mathcal{X} X . Write s θ ( x ) : = ∇ θ log p θ ( x ) s_{\theta}(x):=\nabla_{\theta}\log p_{\theta}(x) s θ ( x ) := ∇ θ log p θ ( x ) . We first verify differentiability in quadratic mean at θ \theta θ . By the preceding positivity bound, there exists δ > 0 \delta>0 δ > 0 such that
inf η ∈ U inf x ∈ X p η ( x ) ≥ δ . \inf_{\eta\in U}\inf_{x\in\mathcal{X}}p_{\eta}(x)\geq\delta. η ∈ U inf x ∈ X inf p η ( x ) ≥ δ .
Moreover, since η ↦ ρ η \eta\mapsto\rho_{\eta} η ↦ ρ η is C 2 C^{2} C 2 in trace norm, each map η ↦ p η ( x ) \eta\mapsto p_{\eta}(x) η ↦ p η ( x ) is C 2 C^{2} C 2 on U U U , and consequently, η ↦ p η ( x ) \eta\mapsto\sqrt{p_{\eta}(x)} η ↦ p η ( x ) is also C 2 C^{2} C 2 , and Taylor expansion at θ \theta θ gives, for h → 0 h\to0 h → 0 ,
p θ + h ( x ) = p θ ( x ) + 1 2 h T s θ ( x ) p θ ( x ) + R x ( h ) , \sqrt{p_{\theta+h}(x)}=\sqrt{p_{\theta}(x)}+\frac{1}{2}h^{T}s_{\theta}(x)\sqrt{p_{\theta}(x)}+R_{x}(h), p θ + h ( x ) = p θ ( x ) + 2 1 h T s θ ( x ) p θ ( x ) + R x ( h ) ,
where, uniformly over x ∈ X x\in\mathcal{X} x ∈ X ,
∣ R x ( h ) ∣ ≤ C ∥ h ∥ 2 |R_{x}(h)|\leq C\lVert h\rVert^{2} ∣ R x ( h ) ∣ ≤ C ∥ h ∥ 2
for all sufficiently small h h h . Since X \mathcal{X} X is finite,
∑ x ∈ X R x ( h ) 2 = O ( ∥ h ∥ 4 ) = o ( ∥ h ∥ 2 ) . \sum_{x\in\mathcal{X}}R_{x}(h)^{2}=O(\lVert h\rVert^{4})=o(\lVert h\rVert^{2}). x ∈ X ∑ R x ( h ) 2 = O (∥ h ∥ 4 ) = o (∥ h ∥ 2 ) .
Hence
∑ x ∈ X [ p θ + h ( x ) − p θ ( x ) − 1 2 h T s θ ( x ) p θ ( x ) ] 2 = o ( ∥ h ∥ 2 ) , \sum_{x\in\mathcal{X}}\left[\sqrt{p_{\theta+h}(x)}-\sqrt{p_{\theta}(x)}-\frac{1}{2}h^{T}s_{\theta}(x)\sqrt{p_{\theta}(x)}\right]^{2}=o(\lVert h\rVert^{2}), x ∈ X ∑ [ p θ + h ( x ) − p θ ( x ) − 2 1 h T s θ ( x ) p θ ( x ) ] 2 = o (∥ h ∥ 2 ) ,
which is precisely differentiability in quadratic mean at θ \theta θ , with score s θ s_{\theta} s θ .
The score has mean zero, and its covariance matrix is the classical Fisher information I θ p i l ≻ 0 I_{\theta}^{\mathrm{pil}}\succ0 I θ pil ≻ 0 .
It remains to verify the local Lipschitz condition on the log likelihood. Since p η ( x ) p_{\eta}(x) p η ( x ) is C 2 C^{2} C 2 and uniformly bounded away from zero on the compact set U U U ,
( η , x ) ⟼ ∇ η log p η ( x ) = ∇ η p η ( x ) p η ( x ) (\eta,x)\longmapsto\nabla_{\eta}\log p_{\eta}(x)=\frac{\nabla_{\eta}p_{\eta}(x)}{p_{\eta}(x)} ( η , x ) ⟼ ∇ η log p η ( x ) = p η ( x ) ∇ η p η ( x )
is continuous and uniformly bounded. Thus there is a finite constant C U C_{U} C U such that
sup η ∈ U sup x ∈ X ∥ ∇ η log p η ( x ) ∥ ≤ C U . \sup_{\eta\in U}\sup_{x\in\mathcal{X}}\lVert\nabla_{\eta}\log p_{\eta}(x)\rVert\leq C_{U}. η ∈ U sup x ∈ X sup ∥ ∇ η log p η ( x )∥ ≤ C U .
By the mean-value theorem, for every η , ϑ ∈ U \eta,\vartheta\in U η , ϑ ∈ U and x ∈ X x \in\mathcal{X} x ∈ X ,
∣ log p η ( x ) − log p ϑ ( x ) ∣ ≤ C U ∥ η − ϑ ∥ . \lvert\log p_{\eta}(x)-\log p_{\vartheta}(x)\rvert\le C_U\lVert\eta-\vartheta\rVert. ∣ log p η ( x ) − log p ϑ ( x )∣ ≤ C U ∥ η − ϑ ∥ .
Equivalently, the required Lipschitz envelope may be taken to be the constant function F ( x ) ≡ C U F(x)\equiv C_U F ( x ) ≡ C U and the finiteness of the outcome space gives
E θ [ F ( X ) 2 ] = C U 2 < ∞ . \mathbb{E}_{\theta}[F(X)^2]=C_U^2<\infty. E θ [ F ( X ) 2 ] = C U 2 < ∞.
Thus the local Lipschitz and square-integrable-envelope conditions required in [44 Theorem 5.39] are satisfied. Hence the same theorem gives
m n ( θ ~ n − θ ) ⇒ N d ( 0 , ( I θ p i l ) − 1 ) . \sqrt{m_n}(\widetilde{\theta}_n-\theta)\Rightarrow N_d\left(0,(I_{\theta}^{\mathrm{pil}})^{-1}\right). m n ( θ n − θ ) ⇒ N d ( 0 , ( I θ pil ) − 1 ) .
This proves (38) . In particular,
m n ( θ ~ n − θ ) = O P θ ( 1 ) , \sqrt{m_n}(\widetilde{\theta}_n-\theta)=O_{\mathbb{P}_{\theta}}(1), m n ( θ n − θ ) = O P θ ( 1 ) ,
which is (35) .
Well-definedness of the adaptive measurement
Since η ↦ Y η \eta\mapsto Y_{\eta} η ↦ Y η is continuous and bounded on K K K , for every bounded continuous real function f f f the map η ↦ f ( Y η ) \eta\mapsto f(Y_{\eta}) η ↦ f ( Y η ) is continuous. This follows by uniform polynomial approximation on an interval containing all the spectra. Taking bounded monotone limits and using finite-dimensional spectral calculus then shows that η ↦ M η ( B ) \eta\mapsto M_{\eta}(B) η ↦ M η ( B ) is Borel measurable for every Borel set B B B . This justifies choosing the measurement at the random pilot estimate, even when eigenvalues of Y η Y_{\eta} Y η coincide.
More explicitly, let P n p i l ( d w ) P_n^{\mathrm{pil}}(dw) P n pil ( d w ) be the pilot POVM, including any classical processing, and write θ ~ n = τ n ( w ) \widetilde{\theta}_n=\tau_n(w) θ n = τ n ( w ) for its measurable estimate map. The joint law of the pilot and second-stage outcomes is
Tr { ρ ϑ ⊗ m n P n p i l ( d w ) } ∏ a = 1 N n Tr { ρ ϑ M τ n ( w ) ( d z a ) } . (151) \operatorname{Tr}\{\rho_{\vartheta}^{\otimes m_n}P_n^{\mathrm{pil}}(dw)\}\prod_{a=1}^{N_n}\operatorname{Tr}\{\rho_{\vartheta}M_{\tau_n(w)}(dz_a)\}. \tag*{(151)} Tr { ρ ϑ ⊗ m n P n pil ( d w )} a = 1 ∏ N n Tr { ρ ϑ M τ n ( w ) ( d z a )} . (151)
This is a parameter-independent measurement rule on the product experiment: only the outcome law depends on ϑ \vartheta ϑ . For a measurable rejection region D D D in the recorded outcomes, its quantum effect is obtained by integrating P n p i l ( d w ) ⊗ M τ n ( w ) ⊗ N n ( D w ) P_n^{\mathrm{pil}}(dw)\otimes M_{\tau_n(w)}^{\otimes N_n}(D_w) P n pil ( d w ) ⊗ M τ n ( w ) ⊗ N n ( D w ) , where D w D_w D w is the section of D D D at w w w . Positivity, normalization, and countable additivity follow from those of the conditional POVMs; in finite dimension the integrals may be evaluated entrywise. This establishes 0 ≤ T n , α ≤ I 0\le T_{n,\alpha}\le I 0 ≤ T n , α ≤ I . Formula (151) also proves conditional independence of the second-stage outcomes.
Proof of Theorem 5.2
Proof. Fix h n → h h_n\to h h n → h and put θ n = θ 0 + h n / n \theta_n=\theta_0+h_n/\sqrt{n} θ n = θ 0 + h n / n . All limits in probability in this proof are under P θ n ( n ) \mathbb{P}_{\theta_n}^{(n)} P θ n ( n ) unless another law is specified. By Lemma G.1, we have
∥ ρ θ 0 + h / m n ⊗ m n − ρ θ ⊗ m n ∥ 1 ≤ ∫ 0 1 b T G θ + t b / m n b d t (152) \left\lVert\rho_{\theta_0+h/\sqrt{m_n}}^{\otimes m_n}-\rho_{\theta}^{\otimes m_n}\right\rVert_1\le\int_0^1\sqrt{b^{\mathsf{T}}G_{\theta+tb/\sqrt{m_n}}b}\,dt \tag*{(152)} ρ θ 0 + h / m n ⊗ m n − ρ θ ⊗ m n 1 ≤ ∫ 0 1 b T G θ + t b / m n b d t (152)
valid whenever the segment is contained in Θ \Theta Θ . Apply this bound with m = m n m=m_n m = m n and b = b n : = m n / n h n b=b_n:=\sqrt{m_n/n}\,h_n b = b n := m n / n h n . The path is θ 0 + t h n / n \theta_0+th_n/\sqrt{n} θ 0 + t h n / n , so boundedness of G G G near θ 0 \theta_0 θ 0 gives
d n : = ∥ P θ n ⊗ m n − P θ 0 ⊗ m n ∥ 1 ≤ C m n / n ∥ h n ∥ → 0. (153) d_n:=\left\|\mathbb{P}_{\theta_n}^{\otimes m_n}-\mathbb{P}_{\theta_0}^{\otimes m_n}\right\|_1\le C\sqrt{m_n/n}\,\lVert h_n\rVert\to0. \tag*{(153)} d n := P θ n ⊗ m n − P θ 0 ⊗ m n 1 ≤ C m n / n ∥ h n ∥ → 0. (153)
Every event determined by the pilot has an associated effect on its m n m_n m n copies. Therefore, for any such event B n B_n B n ,
∣ P θ n ⊗ m n ( B n ) − P θ 0 ⊗ m n ( B n ) ∣ ≤ d n . \left|\mathbb{P}_{\theta_n}^{\otimes m_n}(B_n)-\mathbb{P}_{\theta_0}^{\otimes m_n}(B_n)\right|\le d_n. P θ n ⊗ m n ( B n ) − P θ 0 ⊗ m n ( B n ) ≤ d n .
Applying the event bound to the tail events of m n ∥ θ ~ n − θ 0 ∥ \sqrt{m_n}\lVert\widetilde{\theta}_n-\theta_0\rVert m n ∥ θ n − θ 0 ∥ transfers (35) from θ 0 \theta_0 θ 0 to θ n \theta_n θ n . Since m n ( θ n − θ 0 ) = b n → 0 \sqrt{m_n}(\theta_n-\theta_0)=b_n\to0 m n ( θ n − θ 0 ) = b n → 0 , we obtain
m n ( θ ~ n − θ n ) = O P θ n ⊗ m n ( 1 ) , θ ~ n → θ 0 . (154) \sqrt{m_n}(\widetilde{\theta}_n-\theta_n)=O_{\mathbb{P}_{\theta_n}^{\otimes m_n}}(1),\qquad\widetilde{\theta}_n\to\theta_0. \tag*{(154)} m n ( θ n − θ n ) = O P θ n ⊗ m n ( 1 ) , θ n → θ 0 . (154)
Set μ n = μ ( θ n , θ ~ n ) \mu_n=\mu(\theta_n,\widetilde{\theta}_n) μ n = μ ( θ n , θ n ) and B n = ψ ( θ ~ n ) + μ n − ψ ( θ n ) B_n=\psi(\widetilde{\theta}_n)+\mu_n-\psi(\theta_n) B n = ψ ( θ n ) + μ n − ψ ( θ n ) . With probability tending to one, both parameters lie in the neighborhood where (33) holds. Consequently,
n ∣ B n ∣ ≤ C n ∥ θ ~ n − θ n ∥ 2 = O P θ n ⊗ m n ( n / m n ) = o P θ n ⊗ m n ( 1 ) . (155) \sqrt{n}|B_n|\le C\sqrt{n}\lVert\widetilde{\theta}_n-\theta_n\rVert^2=O_{\mathbb{P}_{\theta_n}^{\otimes m_n}}(\sqrt{n}/m_n)=o_{\mathbb{P}_{\theta_n}^{\otimes m_n}}(1). \tag*{(155)} n ∣ B n ∣ ≤ C n ∥ θ n − θ n ∥ 2 = O P θ n ⊗ m n ( n / m n ) = o P θ n ⊗ m n ( 1 ) . (155)
The displayed bound is used on that event; its complement has probability tending to zero, which suffices for the probability conclusion. No moment bound on the pilot’s exceptional event is needed.
Let σ n 2 = ν ( θ n , θ ~ n ) \sigma_n^2=\nu(\theta_n,\widetilde{\theta}_n) σ n 2 = ν ( θ n , θ n ) . Continuity of (31) and (154) imply
μ n → 0 , σ n 2 → ν ( θ 0 , θ 0 ) = V 0 . (156) \mu_n\to0,\qquad\sigma_n^2\to\nu(\theta_0,\theta_0)=V_0. \tag*{(156)} μ n → 0 , σ n 2 → ν ( θ 0 , θ 0 ) = V 0 . (156)
Conditional on F n p i l \mathcal{F}_n^{\mathrm{pil}} F n pil , the centered variables ξ n , a = Z n , a − μ n \xi_{n,a}=Z_{n,a}-\mu_n ξ n , a = Z n , a − μ n are independent and identically distributed, with mean zero, variance σ n 2 \sigma_n^2 σ n 2 , and absolute value at most 2 C Y 2C_Y 2 C Y . For fixed t ∈ R t\in\mathbb{R} t ∈ R , Taylor’s formula for the scalar exponential gives
E θ n ( n ) [ e i t ξ n , 1 / N n ∣ F n p i l ] = 1 − t 2 σ n 2 2 N n + R n , t , ∣ R n , t ∣ ≤ ∣ t ∣ 3 ( 2 C Y ) 3 6 N n 3 / 2 . \mathbb{E}_{\theta_n}^{(n)}\left[e^{it\xi_{n,1}/\sqrt{N_n}}\mid\mathcal{F}_n^{\mathrm{pil}}\right]
=1-\frac{t^2\sigma_n^2}{2N_n}+R_{n,t},\qquad
|R_{n,t}|\le\frac{|t|^3(2C_Y)^3}{6N_n^{3/2}}. E θ n ( n ) [ e i t ξ n , 1 / N n ∣ F n pil ] = 1 − 2 N n t 2 σ n 2 + R n , t , ∣ R n , t ∣ ≤ 6 N n 3/2 ∣ t ∣ 3 ( 2 C Y ) 3 .
Thus conditional independence and the expansion log ( 1 + z ) = z + O ( ∣ z ∣ 2 ) \log(1+z)=z+O(|z|^2) log ( 1 + z ) = z + O ( ∣ z ∣ 2 ) near zero imply
E θ n ( n ) [ e i t N n ( Z ‾ n − μ n ) ∣ F n p i l ] = ( 1 − t 2 σ n 2 2 N n + R n , t ) N n → e − t 2 V 0 / 2 in probability . (157) \begin{aligned}
\mathbb{E}_{\theta_n}^{(n)}\left[e^{it\sqrt{N_n}(\overline{Z}_n-\mu_n)}\mid\mathcal{F}_n^{\mathrm{pil}}\right]
&=\left(1-\frac{t^2\sigma_n^2}{2N_n}+R_{n,t}\right)^{N_n}\\
&\to e^{-t^2V_0/2}\quad\text{in probability}. \tag*{(157)}
\end{aligned} E θ n ( n ) [ e i t N n ( Z n − μ n ) ∣ F n pil ] = ( 1 − 2 N n t 2 σ n 2 + R n , t ) N n → e − t 2 V 0 /2 in probability . (157)
The remainder in the logarithmic expansion is uniform over pilot outcomes, since σ n 2 ≤ 4 C Y 2 \sigma_n^2\le4C_Y^2 σ n 2 ≤ 4 C Y 2 . Conditional characteristic functions have modulus at most one; convergence in probability to the constant in (157) therefore also holds in mean absolute value. Taking expectations and applying the characteristic-function continuity theorem yields
N n ( Z ‾ n − μ n ) ⇒ N ( 0 , V 0 ) . (158) \sqrt{N_n}(\overline{Z}_n-\mu_n)\Rightarrow N(0,V_0). \tag*{(158)} N n ( Z n − μ n ) ⇒ N ( 0 , V 0 ) . (158)
Since N n / n → 1 N_n/n\to1 N n / n → 1 , the same limit holds with n \sqrt{n} n .
Finally the exact identity
ψ ^ n − ψ ( θ n ) = B n + ( Z ‾ n − μ n ) \widehat{\psi}_n-\psi(\theta_n)=B_n+(\overline{Z}_n-\mu_n) ψ n − ψ ( θ n ) = B n + ( Z n − μ n )
and (154)–(157) prove the centered normal limit in (41) . Continuity of V e f f V_{\mathrm{eff}} V eff and (154) imply V ^ n → V 0 > 0 \widehat{V}_n\to V_0>0 V n → V 0 > 0 ; the lower bound n − 1 n^{-1} n − 1 is irrelevant with probability tending to one. Finally, if ψ ( θ 0 ) = 0 \psi(\theta_0)=0 ψ ( θ 0 ) = 0 , differentiability gives
n ψ ( θ n ) = c 0 T h n + o ( 1 ) → c 0 T h . \sqrt{n}\psi(\theta_n)=c_0^{\mathsf{T}}h_n+o(1)\to c_0^{\mathsf{T}}h. n ψ ( θ n ) = c 0 T h n + o ( 1 ) → c 0 T h .
Adding this shift and applying Slutsky’s theorem proves (42) .
Proof of Corollary 5.3
Proof. For every h n → h h_n\to h h n → h , Theorem 5.2 gives the normal limit of W n W_n W n under θ 0 + h n / n \theta_0+h_n/\sqrt{n} θ 0 + h n / n . Therefore the one- and two-sided rejection probabilities converge to β α + ( c 0 T h ; V 0 ) \beta_\alpha^+(c_0^{\mathsf{T}}h;V_0) β α + ( c 0 T h ; V 0 ) and β α ± ( c 0 T h ; V 0 ) \beta_\alpha^\pm(c_0^{\mathsf{T}}h;V_0) β α ± ( c 0 T h ; V 0 ) , respectively.
To prove compact uniformity, suppose that the supremum in (44) does not converge to zero for a compact H H H . Then along a subsequence there are h n ∈ H h_n\in H h n ∈ H for which the error is bounded below by a fixed positive constant. Passing to a further subsequence gives h n → h ∈ H h_n\to h\in H h n → h ∈ H . The sequential local limit just proved and continuity of the functions in (43) force that error to zero, a contradiction.
Differentiability at θ 0 \theta_0 θ 0 also gives, uniformly for h ∈ H h\in H h ∈ H ,
n ψ ( θ 0 + h / n ) = c 0 T h + o ( 1 ) . \sqrt{n}\psi(\theta_0+h/\sqrt{n})=c_0^{\mathsf{T}}h+o(1). n ψ ( θ 0 + h / n ) = c 0 T h + o ( 1 ) .
At an actual one-sided local null, this implies c 0 T h ≤ o ( 1 ) c_0^{\mathsf{T}}h\le o(1) c 0 T h ≤ o ( 1 ) ; since β α + ( t ; V 0 ) \beta_\alpha^+(t;V_0) β α + ( t ; V 0 ) is increasing in t t t and equals α \alpha α at zero, (44) gives the first level bound. At an actual equality null, c 0 T h = o ( 1 ) c_0^{\mathsf{T}}h=o(1) c 0 T h = o ( 1 ) , and continuity of β α + \beta_\alpha^+ β α + at zero gives the second. At h = 0 h=0 h = 0 , both limiting powers equal α \alpha α directly. Finally, β α ± ( t ; V 0 ) \beta_\alpha^\pm(t;V_0) β α ± ( t ; V 0 ) is even and nondecreasing in ∣ t ∣ |t| ∣ t ∣ , with value α \alpha α at t = 0 t=0 t = 0 , which also gives the stated local asymptotic unbiasedness for the tangent hypotheses.
Proofs for weak Schur sampling
Proof of Lemma 6.1. We first establish the discrete Schur–multinomial approximation used in the separated-spectrum LAN construction of [27 Section 7.5] . Let K K K be a compact neighborhood of α 0 \alpha_0 α 0 in the ordered simplex on which every eigenvalue and adjacent gap is at least δ > 0 \delta>0 δ > 0 . Write P n , α = S W n ( α ) P_{n,\alpha}=\mathrm{SW}^{n}(\alpha) P n , α = SW n ( α ) , extended by zero to unordered count vectors, and Q n , α = Mult ( n , α ) Q_{n,\alpha}=\operatorname{Mult}(n,\alpha) Q n , α = Mult ( n , α ) on
K n : = { k ∈ Z ≥ 0 r : ∑ i k i = n } . \mathcal{K}_n:=\left\{k\in\mathbb{Z}_{\ge0}^{r}:\sum_i k_i=n\right\}. K n := { k ∈ Z ≥ 0 r : i ∑ k i = n } .
We claim that
sup α ∈ K ∥ P n , α − Q n , α ∥ T V → 0. (159) \sup_{\alpha\in K}\lVert P_{n,\alpha}-Q_{n,\alpha}\rVert_{\mathrm{TV}}\to0. \tag*{(159)} α ∈ K sup ∥ P n , α − Q n , α ∥ TV → 0. (159)
For a partition λ \lambda λ , the dimension and Schur polynomial identities give
P n , α ( λ ) = n ! ∏ i < j ( λ i − λ j + j − i ) ∏ i ( λ i + r − i ) ! det [ α i λ j + r − j ] i , j = 1 r ∏ i < j ( α i − α j ) . P_{n,\alpha}(\lambda)=
\frac{n!\prod_{i<j}(\lambda_i-\lambda_j+j-i)}
{\prod_i(\lambda_i+r-i)!}
\frac{\det[\alpha_i^{\lambda_j+r-j}]_{i,j=1}^{r}}
{\prod_{i<j}(\alpha_i-\alpha_j)}. P n , α ( λ ) = ∏ i ( λ i + r − i )! n ! ∏ i < j ( λ i − λ j + j − i ) ∏ i < j ( α i − α j ) det [ α i λ j + r − j ] i , j = 1 r .
Consider the typical set
A n , α : = { k ∈ K n : max i ∣ k i − n α i ∣ ≤ n 3 / 4 } . \mathcal{A}_{n,\alpha}:=
\left\{k\in\mathcal{K}_n:\max_i|k_i-n\alpha_i|\le n^{3/4}\right\}. A n , α := { k ∈ K n : i max ∣ k i − n α i ∣ ≤ n 3/4 } .
Uniformly over α ∈ K \alpha\in K α ∈ K , its elements are positive, strictly decreasing partitions for all sufficiently large n n n . Moreover,
Q n , α ( A n , α c ) ≤ 2 r exp ( − 2 n ) . Q_{n,\alpha}(\mathcal{A}_{n,\alpha}^{c}) \le2r\exp(-2\sqrt{n}). Q n , α ( A n , α c ) ≤ 2 r exp ( − 2 n ) .
On this set, the identity-permutation term in the determinant is ∏ i α i λ i + r − i \prod_i \alpha_i^{\lambda_i+r-i} ∏ i α i λ i + r − i . Every other permutation term is smaller in absolute value by a factor at most e − c n e^{-cn} e − c n , uniformly in α ∈ K \alpha\in K α ∈ K : the adjacent gaps of λ i + r − i \lambda_i+r-i λ i + r − i are at least n δ / 2 n\delta/2 n δ /2 , those of log α i \log\alpha_i log α i have a fixed positive lower bound, and sorting any nonidentity permutation increases the corresponding exponent by at least a constant times n n n . Consequently,
P n , α ( λ ) Q n , α ( λ ) = ∏ i < j ( λ i − λ j + j − i ) ∏ i < j ( α i − α j ) ∏ i λ i ! α i r − i ( λ i + r − i ) ! { 1 + O ( e − c n ) } = 1 + O ( n − 1 / 4 ) , λ ∈ A n , α , \begin{aligned}
\frac{P_{n,\alpha}(\lambda)}{Q_{n,\alpha}(\lambda)}
=
\frac{\prod_{i<j}(\lambda_i-\lambda_j+j-i)}{\prod_{i<j}(\alpha_i-\alpha_j)}
\prod_i \frac{\lambda_i!\alpha_i^{r-i}}{(\lambda_i+r-i)!}\{1+O(e^{-cn})\} \\
&=1+O(n^{-1/4}), \qquad\lambda\in\mathcal{A}_{n,\alpha},
\end{aligned} Q n , α ( λ ) P n , α ( λ ) = ∏ i < j ( α i − α j ) ∏ i < j ( λ i − λ j + j − i ) i ∏ ( λ i + r − i )! λ i ! α i r − i { 1 + O ( e − c n )} = 1 + O ( n − 1/4 ) , λ ∈ A n , α ,
uniformly over K K K . Indeed, each factor in the first product is n { 1 + O ( n − 1 / 4 ) } n\{1+O(n^{-1/4})\} n { 1 + O ( n − 1/4 )} , and the i i i th factorial factor is n − ( r − i ) { 1 + O ( n − 1 / 4 ) } n^{-(r-i)}\{1+O(n^{-1/4})\} n − ( r − i ) { 1 + O ( n − 1/4 )} ; the powers of n n n cancel because ∑ i ( r − i ) = r ( r − 1 ) / 2 \sum_i(r-i)=r(r-1)/2 ∑ i ( r − i ) = r ( r − 1 ) /2 .
If the uniform error in this likelihood ratio is η n → 0 \eta_n \to0 η n → 0 and q n = Q n , α ( A n , α c ) q_n=Q_{n,\alpha}(\mathcal{A}_{n,\alpha}^{c}) q n = Q n , α ( A n , α c ) , normalization gives
∑ A n , α ∣ P n , α − Q n , α ∣ ≤ η n , P n , α ( A n , α c ) ≤ q n + η n . \sum_{\mathcal{A}_{n,\alpha}}|P_{n,\alpha}-Q_{n,\alpha}|\le\eta_n,\qquad P_{n,\alpha}(\mathcal{A}_{n,\alpha}^{c})\le q_n+\eta_n. A n , α ∑ ∣ P n , α − Q n , α ∣ ≤ η n , P n , α ( A n , α c ) ≤ q n + η n .
Thus ∥ P n , α − Q n , α ∥ T V ≤ η n + q n \lVert P_{n,\alpha}-Q_{n,\alpha}\rVert_{\mathrm{TV}}\le\eta_n+q_n ∥ P n , α − Q n , α ∥ TV ≤ η n + q n , proving (159) .
Eventually α n ∈ K \boldsymbol{\alpha}_n\in K α n ∈ K . Represent a multinomial vector as N n = ∑ k = 1 n e I n , k N_n=\sum_{k=1}^{n}e_{I_{n,k}} N n = ∑ k = 1 n e I n , k , where the I n , k I_{n,k} I n , k are independent within each row and P ( I n , k = i ) = α n , i \mathbb{P}(I_{n,k}=i)=\alpha_{n,i} P ( I n , k = i ) = α n , i . The centered summands are uniformly bounded and their covariance matrices converge to C 0 C_0 C 0 . The triangular-array central limit theorem and Cramér–Wold therefore imply
( N n − n α n ) / n ⇒ N r ( 0 , C 0 ) . (N_n-n\boldsymbol{\alpha}_n)/\sqrt{n}\Rightarrow N_r(0,C_0). ( N n − n α n ) / n ⇒ N r ( 0 , C 0 ) .
For scalar directions with zero limiting variance, convergence to zero follows from Chebyshev’s inequality. Total variation contracts under measurable maps, so (159) transfers this limit to WSS and proves (52) . The deterministic centering shift gives (53) .
In particular, α ^ n → α 0 \widehat{\boldsymbol{\alpha}}_n\to\boldsymbol{\alpha}_0 α n → α 0 in probability and n ( α ^ n − α n ) \sqrt{n}(\widehat{\boldsymbol{\alpha}}_n-\boldsymbol{\alpha}_n) n ( α n − α n ) is tight. On a convex neighborhood on which φ \varphi φ is C 1 C^1 C 1 ,
φ ( α ^ n ) − φ ( α n ) = b 0 T ( α ^ n − α n ) + R n , \varphi(\widehat{\boldsymbol{\alpha}}_n)-\varphi(\boldsymbol{\alpha}_n)
=b_0^T(\widehat{\boldsymbol{\alpha}}_n-\boldsymbol{\alpha}_n)+R_n, φ ( α n ) − φ ( α n ) = b 0 T ( α n − α n ) + R n ,
where
∣ R n ∣ ≤ ∥ α ^ n − α n ∥ sup 0 ≤ t ≤ 1 ∥ ∇ φ ( α n + t ( α ^ n − α n ) ) − b 0 ∥ . |R_n|\le\lVert\widehat{\boldsymbol{\alpha}}_n-\boldsymbol{\alpha}_n\rVert
\sup_{0\le t\le1}\left\lVert\nabla\varphi(\boldsymbol{\alpha}_n+t(\widehat{\boldsymbol{\alpha}}_n-\boldsymbol{\alpha}_n))-b_0\right\rVert. ∣ R n ∣ ≤ ∥ α n − α n ∥ 0 ≤ t ≤ 1 sup ∥ ∇ φ ( α n + t ( α n − α n )) − b 0 ∥ .
The supremum is o P ( 1 ) o_P(1) o P ( 1 ) , and the event on which this expansion applies has probability tending to one. Hence n R n = o P ( 1 ) \sqrt{n}R_n=o_P(1) n R n = o P ( 1 ) , and (54) follows from (52) and Slutsky’s theorem. □ \square □
Proof of Theorem 6.2. ⟨ 1 ⟩ \langle1\rangle ⟨ 1 ⟩ By differentiability of ψ \psi ψ at θ 0 \theta_0 θ 0 ,
n [ ψ ( θ 0 + h n ) − ψ ( θ 0 ) ] = c 0 T h + o ( 1 ) . (160) \sqrt{n}\left[\psi\left(\theta_0+\frac{h}{\sqrt{n}}\right)-\psi(\theta_0)\right]=c_0^T h+o(1). \tag*{(160)} n [ ψ ( θ 0 + n h ) − ψ ( θ 0 ) ] = c 0 T h + o ( 1 ) . (160)
⟨ 2 ⟩ \langle2\rangle ⟨ 2 ⟩ By the q-LAN convergence under local alternatives,
( X ( n ) , ρ θ 0 + h / n ⊗ n ) ⇒ N ( h , Σ θ 0 ) . \left(X^{(n)},\rho_{\theta_0+h/\sqrt{n}}^{\otimes n}\right)\Rightarrow N(h,\Sigma_{\theta_0}). ( X ( n ) , ρ θ 0 + h / n ⊗ n ) ⇒ N ( h , Σ θ 0 ) .
Therefore,
c 0 T X ( n ) ⇒ N ( c 0 T h , V e f f ) . c_0^T X^{(n)} \Rightarrow N(c_0^T h,V_{\mathrm{eff}}). c 0 T X ( n ) ⇒ N ( c 0 T h , V eff ) .
Combining this convergence with (160) proves (55) .
⟨3⟩ By Lemma 6.1 , applied with α n = α θ 0 + h / n \boldsymbol{\alpha}_n=\boldsymbol{\alpha}_{\theta_0+h/\sqrt{n}} α n = α θ 0 + h / n ,
n [ φ ( λ ( n ) n ) − φ ( α θ 0 + h / n ) ] ⇒ N ( 0 , V W S S ) . \sqrt{n}\left[\varphi\left(\frac{\lambda^{(n)}}{n}\right)-\varphi\left(\boldsymbol{\alpha}_{\theta_0+h/\sqrt{n}}\right)\right]\Rightarrow N(0,V_{\mathrm{WSS}}). n [ φ ( n λ ( n ) ) − φ ( α θ 0 + h / n ) ] ⇒ N ( 0 , V WSS ) .
This is exactly (56) .
⟨4⟩ The rejection region in (49) , together with the convergence obtained in Step ⟨2⟩, gives
lim n → ∞ Tr [ ρ θ 0 + h / n ⊗ n T n , 1 ] = 1 − Φ ( z γ − κ ( h ) V e f f ) . \lim_{n\to\infty}\operatorname{Tr}\left[\rho_{\theta_0+h/\sqrt{n}}^{\otimes n}T_{n,1}\right]=1-\Phi\left(z_\gamma-\frac{\kappa(h)}{\sqrt{V_{\mathrm{eff}}}}\right). n → ∞ lim Tr [ ρ θ 0 + h / n ⊗ n T n , 1 ] = 1 − Φ ( z γ − V eff κ ( h ) ) .
The boundary of the limiting rejection region has zero Gaussian probability. The same argument applied to (56) proves (58) .
⟨5⟩ Fix a direction h h h with κ ( h ) > 0 \kappa(h)>0 κ ( h ) > 0 , a significance level γ ∈ ( 0 , 1 ) \gamma\in(0,1) γ ∈ ( 0 , 1 ) , and a desired power
π ∈ ( γ , 1 ) . \pi\in(\gamma,1). π ∈ ( γ , 1 ) .
Write
z π : = Φ − 1 ( 1 − π ) . z_\pi:=\Phi^{-1}(1-\pi). z π := Φ − 1 ( 1 − π ) .
At the alternative θ 0 + t h \theta_0+th θ 0 + t h , a test with asymptotic variance V j V_j V j , j ∈ { 1 , 2 } j\in\{1,2\} j ∈ { 1 , 2 } , has asymptotic power
1 − Φ ( z γ − n t κ ( h ) V j ) . 1-\Phi\left(z_\gamma-\frac{\sqrt{n}t\kappa(h)}{\sqrt{V_j}}\right). 1 − Φ ( z γ − V j n t κ ( h ) ) .
Equating this expression to π \pi π gives
n t κ ( h ) V j = z γ − z π . \frac{\sqrt{n}t\kappa(h)}{\sqrt{V_j}}=z_\gamma-z_\pi. V j n t κ ( h ) = z γ − z π .
Consequently, the sample sizes n j ( t ) n_j(t) n j ( t ) prescribed by the limiting Gaussian power satisfy
n j ( t ) ∼ V j ( z γ − z π ) 2 t 2 κ ( h ) 2 , t ↓ 0. (161) n_j(t)\sim\frac{V_j(z_\gamma-z_\pi)^2}{t^2\kappa(h)^2},\qquad t\downarrow0. \tag*{(161)} n j ( t ) ∼ t 2 κ ( h ) 2 V j ( z γ − z π ) 2 , t ↓ 0. (161)
Taking the quotient in the orientation used by the Pitman efficiency of the first test relative to the second gives
ARE P ( T 1 , T 2 ) = lim t ↓ 0 n 2 ( t ) n 1 ( t ) = V W S S V e f f . \operatorname{ARE}_{\mathrm{P}}(T_1,T_2)=\lim_{t\downarrow0}\frac{n_2(t)}{n_1(t)}=\frac{V_{\mathrm{WSS}}}{V_{\mathrm{eff}}}. ARE P ( T 1 , T 2 ) = t ↓ 0 lim n 1 ( t ) n 2 ( t ) = V eff V WSS .
This proves (59) .
Proof of Corollary 6.3. The difference V W S S − V s p V_{\mathrm{WSS}}-V_{\mathrm{sp}} V WSS − V sp
Let D 0 : = diag ( α 0 ) D_0 := \operatorname{diag}(\alpha_0) D 0 := diag ( α 0 ) . Then it is easy to check that w 0 : = D 0 − 1 A 0 I s p , 0 + c 0 w_0 := D_0^{-1}A_0I_{\mathrm{sp},0}^{+}c_0 w 0 := D 0 − 1 A 0 I sp , 0 + c 0 is a feasible solution to the optimization problem (61) , and that w 0 T D 0 w 0 = V s p w_0^T D_0w_0 = V_{\mathrm{sp}} w 0 T D 0 w 0 = V sp . For any other feasible w w w , write w = w 0 + r w=w_0+r w = w 0 + r . Since w w w and w 0 w_0 w 0 satisfy the first constraint in (61) , we get A 0 T r = 0 A_0^T r=0 A 0 T r = 0 , and hence w 0 T D 0 r = c 0 T I s p , 0 + A 0 T r = 0 w_0^T D_0r=c_0^T I_{\mathrm{sp},0}^{+}A_0^T r=0 w 0 T D 0 r = c 0 T I sp , 0 + A 0 T r = 0 . Expanding the quadratic form w T D 0 w w^T D_0w w T D 0 w gives
w T D 0 w = w 0 T D 0 w 0 + r T D 0 r = V s p + r T D 0 r ≥ V s p . w^T D_0w=w_0^T D_0w_0+r^T D_0r=V_{\mathrm{sp}}+r^T D_0r\ge V_{\mathrm{sp}}. w T D 0 w = w 0 T D 0 w 0 + r T D 0 r = V sp + r T D 0 r ≥ V sp .
Since D 0 ≻ 0 D_0\succ0 D 0 ≻ 0 , equality with V s p V_{\mathrm{sp}} V sp occurs precisely when r = 0 r=0 r = 0 , or w = w 0 w=w_0 w = w 0 . This establishes the uniqueness of the minimizer. Now, coming to V W S S V_{\mathrm{WSS}} V WSS , define
b ˉ 0 = b 0 − ( α 0 T b 0 ) 1 \bar{b}_0=b_0-(\alpha_0^T b_0)1 b ˉ 0 = b 0 − ( α 0 T b 0 ) 1
It is easy to check that b ˉ 0 \bar{b}_0 b ˉ 0 is a feasible solution to the optimization problem (61) . Observe that
b ˉ 0 T D 0 b ˉ 0 = b 0 T C 0 b 0 = V W S S . \bar{b}_0^T D_0\bar{b}_0=b_0^T C_0b_0=V_{\mathrm{WSS}}. b ˉ 0 T D 0 b ˉ 0 = b 0 T C 0 b 0 = V WSS .
This implies that V W S S ≥ V s p V_{\mathrm{WSS}}\ge V_{\mathrm{sp}} V WSS ≥ V sp , with equality if and only if b ˉ 0 = w 0 \bar{b}_0=w_0 b ˉ 0 = w 0 . Now note that
b ˉ 0 = w 0 ⟹ b ˉ 0 ∈ Ran ( D 0 − 1 A 0 ) . \bar{b}_0=w_0\Longrightarrow\bar{b}_0\in\operatorname{Ran}(D_0^{-1}A_0). b ˉ 0 = w 0 ⟹ b ˉ 0 ∈ Ran ( D 0 − 1 A 0 ) .
We show the converse. Suppose b ˉ 0 = D 0 − 1 A 0 v \bar{b}_0=D_0^{-1}A_0v b ˉ 0 = D 0 − 1 A 0 v . The constraint A 0 T b ˉ 0 = c 0 A_0^T\bar{b}_0=c_0 A 0 T b ˉ 0 = c 0 gives I s p , 0 v = c 0 I_{\mathrm{sp},0}v=c_0 I sp , 0 v = c 0 . This implies
v − I s p , 0 + c 0 ∈ ker ( I s p , 0 ) = ker ( A 0 ) v-I_{\mathrm{sp},0}^{+}c_0\in\ker(I_{\mathrm{sp},0})=\ker(A_0) v − I sp , 0 + c 0 ∈ ker ( I sp , 0 ) = ker ( A 0 )
and hence D 0 − 1 A 0 ( v − I s p , 0 + c 0 ) = 0 ⟹ b ˉ 0 = w 0 D_0^{-1}A_0(v-I_{\mathrm{sp},0}^{+}c_0)=0\Longrightarrow\bar{b}_0=w_0 D 0 − 1 A 0 ( v − I sp , 0 + c 0 ) = 0 ⟹ b ˉ 0 = w 0 . Therefore,
b ˉ 0 = w 0 ⟺ b ˉ 0 ∈ Ran ( D 0 − 1 A 0 ) (162) \bar{b}_0=w_0\Longleftrightarrow\bar{b}_0\in\operatorname{Ran}(D_0^{-1}A_0) \tag*{(162)} b ˉ 0 = w 0 ⟺ b ˉ 0 ∈ Ran ( D 0 − 1 A 0 ) (162)
Now observe that D 0 b ˉ 0 = D 0 b 0 − α 0 ( α 0 T b 0 ) = C 0 b 0 D_0\bar{b}_0=D_0b_0-\alpha_0(\alpha_0^T b_0)=C_0b_0 D 0 b ˉ 0 = D 0 b 0 − α 0 ( α 0 T b 0 ) = C 0 b 0 . Hence the condition (162) becomes
C 0 b 0 ∈ Ran ( A 0 ) . C_0b_0\in\operatorname{Ran}(A_0). C 0 b 0 ∈ Ran ( A 0 ) .
Thus equality holds if and only if the covariance-weighted derivative C 0 b 0 C_0b_0 C 0 b 0 lies in the space of first-order spectral changes Ran ( A 0 ) \operatorname{Ran}(A_0) Ran ( A 0 ) . This proves (63) .
The difference V s p − V e f f V_{\mathrm{sp}}-V_{\mathrm{eff}} V sp − V eff
Consider the Hilbert space
H 0 : = { H = H ∗ : Tr ( ρ 0 H ) = 0 } , (163) \mathcal{H}_0:=\{H=H^*: \operatorname{Tr}(\rho_0H)=0\}, \tag*{(163)} H 0 := { H = H ∗ : Tr ( ρ 0 H ) = 0 } , (163)
with the inner product
⟨ H , K ⟩ ρ 0 = 1 2 Tr { ρ 0 ( H K + K H ) } . \langle H,K\rangle_{\rho_0}=\frac{1}{2}\operatorname{Tr}\{\rho_0(HK+KH)\}. ⟨ H , K ⟩ ρ 0 = 2 1 Tr { ρ 0 ( H K + K H )} .
Its associated squared norm is ∥ H ∥ ρ 0 2 = Tr ( ρ 0 H 2 ) \lVert H\rVert_{\rho_0}^2=\operatorname{Tr}(\rho_0H^2) ∥ H ∥ ρ 0 2 = Tr ( ρ 0 H 2 ) . Then note that T 0 = span R { L 0 , 1 , … , L 0 , d } \mathcal{T}_0=\operatorname{span}_{\mathbb{R}}\{L_{0,1},\ldots,L_{0,d}\} T 0 = span R { L 0 , 1 , … , L 0 , d } is a subspace of this space. Recall the SLD information matrix ( G 0 ) j k = ⟨ L 0 , j , L 0 , k ⟩ ρ 0 (G_0)_{jk}=\langle L_{0,j},L_{0,k}\rangle_{\rho_0} ( G 0 ) j k = ⟨ L 0 , j , L 0 , k ⟩ ρ 0 .
We now show that V e f f V_{\mathrm{eff}} V eff is the squared norm of an observable in (163) . Define Y 0 = ∑ j = 1 d ( v 0 ) j L 0 , j Y_0=\sum_{j=1}^{d}(v_0)_jL_{0,j} Y 0 = ∑ j = 1 d ( v 0 ) j L 0 , j , where v 0 = G 0 − 1 c 0 v_0=G_0^{-1}c_0 v 0 = G 0 − 1 c 0 . Consequently,
⟨ L 0 , j , Y 0 ⟩ ρ 0 = ( c 0 ) j , ∥ Y 0 ∥ ρ 0 2 = v 0 T G 0 v 0 = c 0 T G 0 − 1 c 0 = V e f f . \langle L_{0,j},Y_0\rangle_{\rho_0}=(c_0)_j,\qquad\|Y_0\|_{\rho_0}^2=v_0^{T}G_0v_0=c_0^{T}G_0^{-1}c_0=V_{\mathrm{eff}}. ⟨ L 0 , j , Y 0 ⟩ ρ 0 = ( c 0 ) j , ∥ Y 0 ∥ ρ 0 2 = v 0 T G 0 v 0 = c 0 T G 0 − 1 c 0 = V eff .
Similarly, V s p V_{\mathrm{sp}} V sp is the squared norm of an observable in (163) . Define H s p , 0 = U 0 diag ( w 0 ) U 0 ∗ H_{\mathrm{sp},0}=U_0\operatorname{diag}(w_0)U_0^* H sp , 0 = U 0 diag ( w 0 ) U 0 ∗ . Then observe that Tr ( ρ 0 H s p , 0 ) = α 0 T w 0 = 0 \operatorname{Tr}(\rho_0H_{\mathrm{sp},0})=\alpha_0^Tw_0=0 Tr ( ρ 0 H sp , 0 ) = α 0 T w 0 = 0 , so H s p , 0 ∈ H 0 H_{\mathrm{sp},0}\in\mathcal{H}_0 H sp , 0 ∈ H 0 . Furthermore,
∥ H s p , 0 ∥ ρ 0 2 = Tr ( ρ 0 H s p , 0 2 ) = Tr ( U 0 D 0 U 0 ∗ U 0 diag ( w 0 ) 2 U 0 ∗ ) = Tr ( U 0 D 0 diag ( w 0 ) 2 U 0 ∗ ) = Tr ( D 0 w 0 w 0 T ) = w 0 T D 0 w 0 = V s p . \begin{aligned}
\|H_{\mathrm{sp},0}\|_{\rho_0}^2
&=\operatorname{Tr}(\rho_0H_{\mathrm{sp},0}^2)\\
&=\operatorname{Tr}(U_0D_0U_0^*U_0\operatorname{diag}(w_0)^2U_0^*)\\
&=\operatorname{Tr}(U_0D_0\operatorname{diag}(w_0)^2U_0^*)\\
&=\operatorname{Tr}(D_0w_0w_0^T)\\
&=w_0^TD_0w_0\\
&=V_{\mathrm{sp}}.
\end{aligned} ∥ H sp , 0 ∥ ρ 0 2 = Tr ( ρ 0 H sp , 0 2 ) = Tr ( U 0 D 0 U 0 ∗ U 0 diag ( w 0 ) 2 U 0 ∗ ) = Tr ( U 0 D 0 diag ( w 0 ) 2 U 0 ∗ ) = Tr ( D 0 w 0 w 0 T ) = w 0 T D 0 w 0 = V sp .
We now show that H s p , 0 − Y 0 H_{\mathrm{sp},0}-Y_0 H sp , 0 − Y 0 and Y 0 Y_0 Y 0 are orthogonal. Indeed, observe that
( U 0 ∗ L 0 , j U 0 ) i i = ( A 0 ) i j α 0 , i . (U_0^*L_{0,j}U_0)_{ii}=\frac{(A_0)_{ij}}{\alpha_{0,i}}. ( U 0 ∗ L 0 , j U 0 ) ii = α 0 , i ( A 0 ) ij .
Therefore,
⟨ L 0 , j , H s p , 0 ⟩ ρ 0 = ∑ i = 1 r α 0 , i ( U 0 ∗ L 0 , j U 0 ) i i ( w 0 ) i = ∑ i = 1 r ( A 0 ) i j ( w 0 ) i = ( A 0 T w 0 ) j = ( c 0 ) j = ⟨ L 0 , j , Y 0 ⟩ ρ 0 . \begin{aligned}
\langle L_{0,j},H_{\mathrm{sp},0}\rangle_{\rho_0}
&=\sum_{i=1}^{r}\alpha_{0,i}(U_0^*L_{0,j}U_0)_{ii}(w_0)_i\\
&=\sum_{i=1}^{r}(A_0)_{ij}(w_0)_i\\
&=(A_0^Tw_0)_j\\
&=(c_0)_j\\
&=\langle L_{0,j},Y_0\rangle_{\rho_0}.
\end{aligned} ⟨ L 0 , j , H sp , 0 ⟩ ρ 0 = i = 1 ∑ r α 0 , i ( U 0 ∗ L 0 , j U 0 ) ii ( w 0 ) i = i = 1 ∑ r ( A 0 ) ij ( w 0 ) i = ( A 0 T w 0 ) j = ( c 0 ) j = ⟨ L 0 , j , Y 0 ⟩ ρ 0 .
Hence
⟨ L 0 , j , H s p , 0 − Y 0 ⟩ ρ 0 = 0 , j = 1 , … , d . (164) \langle L_{0,j},H_{\mathrm{sp},0}-Y_0\rangle_{\rho_0}=0,\qquad j=1,\ldots,d. \tag*{(164)} ⟨ L 0 , j , H sp , 0 − Y 0 ⟩ ρ 0 = 0 , j = 1 , … , d . (164)
Since Y 0 = ∑ j = 1 d ( v 0 ) j L 0 , j Y_0=\sum_{j=1}^{d}(v_0)_jL_{0,j} Y 0 = ∑ j = 1 d ( v 0 ) j L 0 , j ,
⟨ H s p , 0 − Y 0 , Y 0 ⟩ ρ 0 = ∑ j = 1 d ( v 0 ) j ⟨ H s p , 0 − Y 0 , L 0 , j ⟩ ρ 0 = 0. \begin{aligned}
\langle H_{\mathrm{sp},0}-Y_0,Y_0\rangle_{\rho_0}
&=\sum_{j=1}^{d}(v_0)_j\langle H_{\mathrm{sp},0}-Y_0,L_{0,j}\rangle_{\rho_0}\\
&=0.
\end{aligned} ⟨ H sp , 0 − Y 0 , Y 0 ⟩ ρ 0 = j = 1 ∑ d ( v 0 ) j ⟨ H sp , 0 − Y 0 , L 0 , j ⟩ ρ 0 = 0.
Then V s p − V e f f = ∥ H s p , 0 − Y 0 ∥ ρ 0 2 ≥ 0 V_{\mathrm{sp}}-V_{\mathrm{eff}}=\|H_{\mathrm{sp},0}-Y_0\|_{\rho_0}^2\geq0 V sp − V eff = ∥ H sp , 0 − Y 0 ∥ ρ 0 2 ≥ 0 . Indeed,
∥ H s p , 0 − Y 0 ∥ ρ 0 2 = ∥ H s p , 0 ∥ ρ 0 2 − ∥ Y 0 ∥ ρ 0 2 = V s p − V e f f ≥ 0. (165) \begin{aligned}
\|H_{\mathrm{sp},0}-Y_0\|_{\rho_0}^2
&=\|H_{\mathrm{sp},0}\|_{\rho_0}^2-\|Y_0\|_{\rho_0}^2\\
&=V_{\mathrm{sp}}-V_{\mathrm{eff}}\geq0. \tag*{(165)}
\end{aligned} ∥ H sp , 0 − Y 0 ∥ ρ 0 2 = ∥ H sp , 0 ∥ ρ 0 2 − ∥ Y 0 ∥ ρ 0 2 = V sp − V eff ≥ 0. (165)
Equality holds if and only if H s p , 0 = Y 0 H_{\mathrm{sp},0}=Y_0 H sp , 0 = Y 0 . Since H s p , 0 H_{\mathrm{sp},0} H sp , 0 commutes with ρ 0 \rho_0 ρ 0 , we have
H s p , 0 = Y 0 ⟹ [ Y 0 , ρ 0 ] = 0. H_{\mathrm{sp},0}=Y_0\quad\Longrightarrow\quad[Y_0,\rho_0]=0. H sp , 0 = Y 0 ⟹ [ Y 0 , ρ 0 ] = 0.
We prove the converse. If Y 0 Y_0 Y 0 commutes with ρ 0 \rho_0 ρ 0 , then since ρ 0 \rho_0 ρ 0 has simple spectrum, Y 0 Y_0 Y 0 is diagonal in the eigenbasis of ρ 0 \rho_0 ρ 0 , so that
Y 0 = U 0 diag ( y ) U 0 ∗ Y_0 = U_0 \operatorname{diag}(y) U_0^* Y 0 = U 0 diag ( y ) U 0 ∗
for some y ∈ R r y \in\mathbb{R}^r y ∈ R r . We show that y y y is feasible under the optimization problem (61) . First,
α 0 T y = Tr ( ρ 0 Y 0 ) = 0. \alpha_0^T y = \operatorname{Tr}(\rho_0Y_0)
= 0. α 0 T y = Tr ( ρ 0 Y 0 ) = 0.
Also,
( A 0 T y ) j = ∑ i = 1 r ( A 0 ) i j y i = ∑ i = 1 r α 0 , i ( U 0 ∗ L 0 , j U 0 ) i i y i = ⟨ L 0 , j , Y 0 ⟩ ρ 0 = ( c 0 ) j . \begin{aligned}
(A_0^T y)_j
&= \sum_{i=1}^{r}(A_0)_{ij}y_i \\
&= \sum_{i=1}^{r}\alpha_{0,i}(U_0^*L_{0,j}U_0)_{ii}y_i \\
&= \langle L_{0,j},Y_0\rangle_{\rho_0} \\
&= (c_0)_j.
\end{aligned} ( A 0 T y ) j = i = 1 ∑ r ( A 0 ) ij y i = i = 1 ∑ r α 0 , i ( U 0 ∗ L 0 , j U 0 ) ii y i = ⟨ L 0 , j , Y 0 ⟩ ρ 0 = ( c 0 ) j .
Thus A 0 T y = c 0 A_0^T y = c_0 A 0 T y = c 0 , and y y y is feasible. But the value of the objective function at y y y is
y T D 0 y = Tr ( ρ 0 Y 0 2 ) = ∥ Y 0 ∥ ρ 0 2 = V e f f . \begin{aligned}
y^T D_0 y &= \operatorname{Tr}(\rho_0Y_0^2) \\
&= \lVert Y_0\rVert_{\rho_0}^2 \\
&= V_{\mathrm{eff}}.
\end{aligned} y T D 0 y = Tr ( ρ 0 Y 0 2 ) = ∥ Y 0 ∥ ρ 0 2 = V eff .
Hence
V e f f ≥ V s p . V_{\mathrm{eff}} \ge V_{\mathrm{sp}}. V eff ≥ V sp .
We also had earlier, from (165) , that V s p ≥ V e f f V_{\mathrm{sp}} \ge V_{\mathrm{eff}} V sp ≥ V eff . Hence V s p = V e f f V_{\mathrm{sp}} = V_{\mathrm{eff}} V sp = V eff . By the uniqueness of the minimizer, y = w 0 y = w_0 y = w 0 . Conjugating with U 0 U_0 U 0 implies Y 0 = H s p , 0 Y_0 = H_{\mathrm{sp},0} Y 0 = H sp , 0 . Hence we have proved
H s p , 0 = Y 0 ⟺ [ Y 0 , ρ 0 ] = 0. H_{\mathrm{sp},0} = Y_0 \Longleftrightarrow[Y_0,\rho_0] = 0. H sp , 0 = Y 0 ⟺ [ Y 0 , ρ 0 ] = 0.
This proves that V s p = V e f f ⟺ [ Y 0 , ρ 0 ] = 0 V_{\mathrm{sp}} = V_{\mathrm{eff}} \Longleftrightarrow[Y_0,\rho_0] = 0 V sp = V eff ⟺ [ Y 0 , ρ 0 ] = 0 , which is (64) . On the finite-dimensional real Hilbert space of centered Hermitian matrices, the commutation operator satisfies
⟨ D ρ 0 H , K ⟩ ρ 0 = − ⟨ H , D ρ 0 K ⟩ ρ 0 . (166) \langle D_{\rho_0}H,K\rangle_{\rho_0} = -\langle H,D_{\rho_0}K\rangle_{\rho_0}. \tag*{(166)} ⟨ D ρ 0 H , K ⟩ ρ 0 = − ⟨ H , D ρ 0 K ⟩ ρ 0 . (166)
This follows directly in the reference eigenbasis: on each real plane formed by the real and imaginary parts of H i j H_{ij} H ij , the map is a scalar multiple of a ninety-degree rotation, while the SLD inner product is a positive multiple of the Euclidean inner product. It vanishes on diagonal matrices. Since the spectrum is simple,
ker D ρ 0 = { H = H ∗ : [ H , ρ 0 ] = 0 } . \ker D_{\rho_0} = \{H = H^*: [H,\rho_0] = 0\}. ker D ρ 0 = { H = H ∗ : [ H , ρ 0 ] = 0 } .
If T 0 \mathcal{T}_0 T 0 is invariant under D ρ 0 D_{\rho_0} D ρ 0 , (166) implies that T 0 ⊥ \mathcal{T}_0^\perp T 0 ⊥ is invariant as well. Hence its orthogonal projection commutes with the operator:
D ρ 0 Π T 0 = Π T 0 D ρ 0 . D_{\rho_0}\Pi_{\mathcal{T}_0} = \Pi_{\mathcal{T}_0}D_{\rho_0}. D ρ 0 Π T 0 = Π T 0 D ρ 0 .
Now note that
Therefore
D ρ 0 Y 0 = D ρ 0 Π T 0 H s p , 0 = Π T 0 D ρ 0 H s p , 0 = 0. D_{\rho_0}Y_0=D_{\rho_0}\Pi_{T_0}H_{\mathrm{sp},0}=\Pi_{T_0}D_{\rho_0}H_{\mathrm{sp},0}=0. D ρ 0 Y 0 = D ρ 0 Π T 0 H sp , 0 = Π T 0 D ρ 0 H sp , 0 = 0.
It follows that [ Y 0 , ρ 0 ] = 0 [Y_0,\rho_0]=0 [ Y 0 , ρ 0 ] = 0 . The equality V s p = V e f f V_{\mathrm{sp}}=V_{\mathrm{eff}} V sp = V eff now follows. Finally, if rank ( A 0 ) = r − 1 \operatorname{rank}(A_0)=r-1 rank ( A 0 ) = r − 1 , then since 1 T A 0 = 0 1^{T}A_0=0 1 T A 0 = 0 ,
Ran ( A 0 ) = { x ∈ R r : 1 T x = 0 } . \operatorname{Ran}(A_0)=\{x\in\mathbb{R}^{r}:1^{T}x=0\}. Ran ( A 0 ) = { x ∈ R r : 1 T x = 0 } .
Since 1 T C 0 = 0 1^{T}C_0=0 1 T C 0 = 0 , we have C 0 b 0 ∈ Ran ( A 0 ) C_0b_0\in\operatorname{Ran}(A_0) C 0 b 0 ∈ Ran ( A 0 ) . Thus V W S S = V s p V_{\mathrm{WSS}}=V_{\mathrm{sp}} V WSS = V sp by (63) . If the SLD space is also D ρ 0 D_{\rho_0} D ρ 0 -invariant, then V s p = V e f f V_{\mathrm{sp}}=V_{\mathrm{eff}} V sp = V eff , and hence
V W S S = V s p = V e f f . V_{\mathrm{WSS}}=V_{\mathrm{sp}}=V_{\mathrm{eff}}. V WSS = V sp = V eff .
□ \square □