Quantum Local Asymptotic Normality (QLAN)

• Quantum Local Asymptotic Normality (QLAN) is a quantum statistical theory that approximates a parametrized quantum statistical model through Gaussian limits with a Hermitian positive semidefinite matrix $J$ that decomposes into a covariance matrix $V$ and a noncommutative component $S$.
• Applications of QLAN include finite-dimensional quantum models (i.i.d. and Markov chains), continuous-time processes, and infinite-dimensional pure-state ensembles. Strong QLAN, particularly, involves trace-norm equivalence of experiments.
• Its primary use case is to elucidate the inference and estimation properties of quantum statistical models.

Quantum local asymptotic normality (QLAN) is an asymptotic theory in which a parametrized quantum statistical model, under local perturbations of order $n^{-1/2}$, is approximated by a quantum Gaussian shift experiment. Its classical analogue replaces a likelihood ratio by a Gaussian likelihood expansion; QLAN must additionally encode noncommutativity through canonical commutation relations (CCR), complex covariance matrices, and quantum measurements. Depending on the model, QLAN may be formulated weakly through quantum convergence in distribution and likelihood expansions, or strongly through quantum channels whose trace-norm errors vanish. It applies to finite-dimensional i.i.d. models, nonfaithful and rank-changing states, quantum Markov chains, continuous-time input-output processes, and certain infinite-dimensional pure-state ensembles.

1. Conceptual framework and local Gaussianization

For a sequence of quantum statistical models

$$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$

QLAN is studied at a reference point $\theta_0$ under the localization

$$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$

The central objects are quantum central sequences $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$ satisfying a quantum central limit theorem,

$\Delta^{(n)}\Longrightarrow N(0,J),$

where $J$ is Hermitian positive semidefinite and $\operatorname{Re}J>0$.

The matrix $J$ decomposes as

$J=V+iS,$

where $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$0 is the covariance matrix and $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$1 specifies the noncommutative component of the limiting experiment. The limiting observables $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$2 satisfy CCR relations of the form

$$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$3

When $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$4, the Gaussian limit is classical. When $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$5, the limiting variables cannot generally be measured simultaneously with arbitrary precision.

Quantum convergence in distribution is defined through quasi-characteristic functions of products of Weyl operators, rather than through ordinary scalar characteristic functions. A quantum Gaussian state $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$6 is characterized by

$$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$7

with the corresponding product formula determined by $$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$8.

The local Gaussianization mechanism has several recurring components:

1. a likelihood or square-root likelihood expansion;
2. a quantum central limit theorem for normalized observables;
3. a quantum analogue of Le Cam’s third lemma;
4. transfer of estimation or testing procedures to the limiting Gaussian model.

The operator-level form developed by Yamagata, Fujiwara, and Gill is

$$S^{(n)}=\{\rho_\theta^{(n)}:\theta\in\Theta\subset\mathbb R^d\},$$9

where the remainder is infinitesimal in the joint quantum-Gaussian sense rather than merely convergent to zero as a scalar random variable (Yamagata et al., 2012). This parallels classical LAN while retaining the imaginary part of $\theta_0$0, which records incompatibility of the limiting observables.

2. Quantum likelihood ratios, absolute continuity, and singularity

For noncommuting density operators, the scalar identity $\theta_0$1 has no direct operator analogue. The symmetric quantum log-likelihood ratio is defined by

$\theta_0$2

When $\theta_0$3 and $\theta_0$4 are faithful, $\theta_0$5 is unique and can be expressed through the operator geometric mean. The quantity

$\theta_0$6

is a fidelity between $\theta_0$7 and $\theta_0$8.

For nonfaithful states, mutual absolute continuity is not adequately described by ordinary support inclusion. A noncommutative Lebesgue decomposition separates $\theta_0$9 into an absolutely continuous component and a singular component: $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$0 The absolutely continuous part admits a square-root likelihood representation

$$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$1

where $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$2 is a square-root likelihood ratio. The singular part is excluded from the likelihood expansion and is required to be asymptotically negligible.

This construction extends QLAN to models with mixed ranks, changing supports, rank-deficient reference states, and local alternatives that are not mutually absolutely continuous at every finite sample size. The corresponding weak-Q-LAN expansion is formulated for the square-root likelihood ratio: $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$3 Here $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$4 denotes infinitesimality in quantum distribution and $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$5 denotes convergence to zero in the reference-state $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$6 sense (Fujiwara et al., 2017).

Quantum contiguity provides the measure-theoretic basis for transferring limiting distributions between reference states and local alternatives. If $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$7 and the square-root likelihood ratios converge jointly with observables, the limiting alternative state is obtained by noncommutative sandwiching: $$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$8 In the Gaussian case, this yields the local shift

$$\theta=\theta_0+\frac{h}{\sqrt n}, \qquad h\in\mathbb R^d.$$9

when the joint reference-state limit has covariance blocks

$$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$0

(1804.03510).

3. Weak and strong formulations

Two principal formulations of QLAN are distinguished.

Weak QLAN concerns convergence of normalized observables and likelihood objects. It does not necessarily provide channels between the original and Gaussian experiments. Its characteristic conclusion is the transfer of Gaussian limits under local alternatives: $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$1 Weak QLAN is sufficiently general to include nonfaithful states and asymptotically negligible singular components. Its limitations include the absence, in general, of a quantum representation theorem, convolution theorem, and local asymptotic minimax theorem.

Strong QLAN is expressed through quantum Le Cam distance. For models $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$2 and $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$3, strong convergence means that there are quantum channels in both directions such that

$$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$4

Equivalently, the original and Gaussian experiments can simulate each other with uniformly vanishing trace-norm error over the local parameter set. This transfers arbitrary measurements and decision procedures asymptotically.

For finite-dimensional i.i.d. models, strong QLAN was developed for broad classes of faithful models through Gaussian approximations. Strong convergence has also been established for pure-state ensembles in infinite-dimensional Hilbert spaces, where nearby tensor powers are equivalent to multimode coherent-state experiments: $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$5 The equivalence holds in quantum Le Cam distance on local neighborhoods $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$6 with $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$7 (Butucea et al., 2017).

The distinction between weak and strong formulations is consequential. Weak QLAN identifies limiting distributions of observables, whereas strong QLAN identifies the asymptotic statistical experiment itself. A representation theorem for weak QLAN was established under an additional $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$8-extendibility condition: every sequence of POVMs with a weak local limit is represented by a POVM on an appropriate quantum Gaussian shift experiment (Fujiwara et al., 2022). In the i.i.d. case, $$\Delta^{(n)}=(\Delta_1^{(n)},\ldots,\Delta_d^{(n)})$$9-extendibility follows from an invariant extension of the SLD tangent space under Holevo’s commutation operator.

4. i.i.d. quantum models and asymptotic estimation

For a smooth finite-dimensional model $\Delta^{(n)}\Longrightarrow N(0,J),$0, consider the tensor-power experiment

$\Delta^{(n)}\Longrightarrow N(0,J),$1

Under suitable differentiability and absolute-continuity conditions, the central sequence is constructed from symmetric logarithmic derivatives (SLDs): $\Delta^{(n)}\Longrightarrow N(0,J),$2 The limiting covariance is

$\Delta^{(n)}\Longrightarrow N(0,J),$3

and the likelihood expansion has the QLAN form

$\Delta^{(n)}\Longrightarrow N(0,J),$4

For additional centered observables $\Delta^{(n)}\Longrightarrow N(0,J),$5, define

$\Delta^{(n)}\Longrightarrow N(0,J),$6

With

$\Delta^{(n)}\Longrightarrow N(0,J),$7

the local alternative shifts the limiting observable by

$\Delta^{(n)}\Longrightarrow N(0,J),$8

The limiting Gaussian model supplies the asymptotic estimation problem. For a positive weight matrix $\Delta^{(n)}\Longrightarrow N(0,J),$9, the Holevo bound is the infimum of

$J$0

over locally unbiased estimators, with the noncommutative penalty determined by the imaginary part of the relevant covariance. The first term corresponds to ordinary information, while the second reflects incompatibility of quantum observables.

The QLAN framework establishes asymptotic achievability of the Holevo bound for broad smooth models, including pure-state and degenerate models. In the construction, an optimal Gaussian measurement is transported back to the original $J$1-copy model. The result is generally a local-shift achievability statement, often first proved on a countable dense subset of local parameters. The stronger asymptotic representation theorem gives a universal lower bound for regular estimators and proves its attainability under $J$2-extendibility (Fujiwara et al., 2022).

A pure-state qubit model illustrates the role of the imaginary covariance. At the reference point,

$J$3

so the two local coordinates converge to a canonical pair. The limiting experiment is a coherent-state Gaussian model, and asymptotically optimal estimators may be chosen separable in the model described in (Yamagata et al., 2012).

QLAN also clarifies the distinction between the SLD, RLD, and Holevo bounds. In multiparameter problems, noncommuting score directions prevent simultaneous attainment of all individual single-parameter bounds. The Gaussian limit converts this issue into a finite-dimensional quantum Gaussian decision problem, with homodyne-type measurements for compatible quadratures and heterodyne-type procedures for joint estimation of noncommuting coordinates.

5. Quantum Markov dynamics

QLAN extends beyond independent copies to correlated quantum processes.

Discrete-time quantum Markov chains

A discrete-time quantum Markov chain is generated by an isometry

$J$4

or equivalently by Kraus operators $J$5 satisfying

$J$6

The associated channel is

$J$7

Under primitivity, the channel has a unique full-rank stationary state and correlations decay exponentially. The asymptotic Fisher information is governed by a Markov covariance involving the resolvent

$J$8

on the zero-mean subspace. For a generator

$J$9

the quantum Fisher information per interaction is

$\operatorname{Re}J>0$0

The second term in the Markov covariance records temporal correlations and distinguishes dynamical information from a one-step variance (Guta et al., 2014).

For a scalar parameter, the local stationary output model converges to a one-mode coherent-state family

$\operatorname{Re}J>0$1

The parameter is localized as

$\operatorname{Re}J>0$2

Strong convergence of the stationary output to the coherent-state model implies that asymptotically optimal Gaussian measurements can be transferred to the output process.

The same theory identifies non-identifiable dynamical directions. A phase change or a unitary coordinate transformation of the hidden memory leaves stationary output states unchanged. The Fisher information vanishes along such directions: $\operatorname{Re}J>0$3 Thus the statistically meaningful parameter space is a quotient by hidden-coordinate symmetries.

More recent work treats irreducible, not necessarily primitive, quantum Markov chains. The identifiable parameter space is then an orbifold, obtained as a quotient by a compact group of state-preserving symmetries. At periodic points, the joint system-output model still has a strong QLAN limit given by a coherent-state shift, whereas the stationary output converges to a product of a Gaussian shift and a stabilizer-averaged Gaussian mixture: $\operatorname{Re}J>0$4 At primitive points the stabilizer is trivial and the mixture reduces to a single Gaussian shift. On a fixed periodic stratum, QLAN is recovered along the stabilizer-fixed tangent directions (Girotti et al., 21 Mar 2026).

Continuous-time quantum Markov processes

In the input-output formalism, a finite-dimensional system interacts with Bosonic fields according to a Hudson–Parthasarathy quantum stochastic differential equation. The reduced system evolution is generated by a Lindblad operator $\operatorname{Re}J>0$5. Under irreducibility, there is a unique full-rank stationary state and the reduced resolvent $\operatorname{Re}J>0$6 governs long-time correlations.

For a scalar dynamical parameter, the local scale is

$\operatorname{Re}J>0$7

The full system-output state converges weakly to a one-mode coherent-state model with quantum Fisher information rate $\operatorname{Re}J>0$8. The rate contains direct contributions from derivatives of the coupling operators and Hamiltonian, together with dynamical response terms obtained by solving a Poisson-type equation involving $\operatorname{Re}J>0$9.

The same process admits classical LAN reductions for particular measurements. Centered counting statistics and integrated homodyne currents converge to ordinary Gaussian shifts: $J$0 with classical Fisher informations

$J$1

These inequalities express information loss from restricting the output to a total count or an integrated quadrature. The full quantum output retains temporal and quantum correlations that may be absent from those summary statistics (Catana et al., 2014).

For multiparameter continuous-time dynamics, the quotient by Hamiltonian phase shifts and system-unitary conjugations forms a principal bundle. The real part of the Markov covariance defines a Fisher metric on the identifiable horizontal space, while its imaginary part defines the symplectic form of the limiting CCR system. The reduced output experiment converges strongly to a multimode coherent-state Gaussian model (Guta et al., 2016).

6. Scope, limitations, and related notions

QLAN should be distinguished from several related constructions.

Classical LAN for high-dimensional statistics has the same local scaling and Gaussian-shift architecture but concerns commuting observations. For example, in supercritical spiked $J$2-ratio models, the full eigenvalue experiment can reduce to the largest outlier and then to a classical Gaussian shift (Dharmawansa et al., 2014). No quantum likelihood, CCR algebra, measurement optimization, or quantum Fisher information is involved.

Metric-projected local asymptotic normality (MP-LAN) addresses positivity constraints in quantum state tomography after a fixed POVM has produced classical data. At rank-deficient states, the unconstrained Gaussian estimator is projected onto a Fisher-metric tangent cone. This explains non-Gaussian maximum-likelihood behavior and the failure of the usual Wilks theorem. MP-LAN is a constrained classical LAN construction, not standard QLAN (Scholten et al., 2016).

Strong QLAN is stronger than weak QLAN because it supplies asymptotically inverse quantum channels and trace-norm equivalence of experiments. Weak QLAN can apply in settings where strong equivalence has not been established, particularly models with changing supports or singular local components.

Important limitations remain. General quantum convergence in distribution is not available in the same form as classical convergence in distribution. Full representation theorems require additional structural conditions such as $J$3-extendibility. Quantum convolution and local asymptotic minimax theorems are not available for every weak-Q-LAN model. Adaptive implementation is also nontrivial because the reference point $J$4 is treated as known in the local theory; proposed procedures use a preliminary estimator followed by a locally optimal measurement.

For continuous-time processes, the multidimensional theory was initially restricted, and explicit optimal measurements on the original output were not universally constructed. For Markov chains, primitivity yields a single Gaussian shift, while periodicity creates orbifold singularities and Gaussian mixtures. For infinite-dimensional models, the strongest equivalence results currently concern i.i.d. pure-state ensembles, with mixed states and general dependent preparations requiring additional control of spectral and covariance structures.

The central content of QLAN is therefore the identification of a local quantum Gaussian experiment that preserves the statistically relevant information at the $J$5 scale. Its real covariance determines distinguishability and Fisher geometry; its imaginary covariance determines the CCR structure and measurement incompatibility. In i.i.d., Markovian, continuous-time, and selected infinite-dimensional settings, QLAN reduces asymptotic quantum inference to a Gaussian problem while retaining the noncommutative structure that has no classical counterpart.