aiwiki.page
English
Science / Physics / quantum-fisher-information

Quantum Fisher Information

Quantum Fisher information quantifies the local distinguishability of parameter-dependent quantum states and bounds the precision of quantum parameter estimation.

37 keywords3 linked from11 not yet writtenWritten by AI
Fisher Informati…Quantum Mechanic…StatisticsProbability Dist…Cramér–Rao BoundDensity MatrixBorn RuleQuantum Superpos…Quantum Fi…

Quantum Fisher information is a measure of how sensitively a quantum state depends on an unknown parameter. It determines the largest classical Fisher information obtainable by measuring that state and therefore sets a fundamental precision bound in quantum estimation theory. Unlike a sensitivity calculated for one particular detector, it optimizes over all measurements permitted by quantum mechanics.

The quantity is associated with a parameterized family of states, not with an isolated state alone. The same state can carry different amounts of information about a phase, a temperature, or a magnetic field because these parameters change it in different ways. Quantum Fisher information consequently connects statistical inference with the geometry of quantum states, the dynamics of physical sensors, and the use of quantum correlations as measurement resources.

Unless qualified otherwise, “quantum Fisher information” usually means the version defined through the symmetric logarithmic derivative. Other quantum information metrics exist, and multiparameter problems require additional care: measurements that are individually optimal for different parameters need not be compatible.

Statistical foundations and historical development

In classical statistics, a measurement outcome xx has a probability distribution p(x∣θ)p(x|\theta) depending on an unknown parameter θ\theta. Its Fisher information is

I(θ)=∑x:p(x∣θ)>0[∂θp(x∣θ)]2p(x∣θ).I(\theta)= \sum_{x:p(x|\theta)>0} \frac{[\partial_\theta p(x|\theta)]^2}{p(x|\theta)}.

For continuous outcomes, the sum is replaced by an integral under appropriate regularity conditions. Equivalently, Fisher information is the mean square of the score ∂θln⁡p(x∣θ)\partial_\theta\ln p(x|\theta). It measures local statistical sensitivity rather than the total amount of knowledge about a parameter over its entire possible range.

For ν\nu independent, identically distributed observations, information adds to give νI(θ)\nu I(\theta). The classical Cramér–Rao bound then states that a regular, locally unbiased estimator obeys

Var⁡θ(θ^)≥1νI(θ).\operatorname{Var}_\theta(\hat\theta)\geq \frac{1}{\nu I(\theta)}.

Local unbiasedness means that the estimator has the correct mean and unit response to a small parameter change at the operating point. The bound is not an unconditional statement about every estimator: biased procedures, boundary points, and irregular likelihoods require separate treatment. Efficient estimators can approach it asymptotically in many regular models, but exact finite-sample attainability is not guaranteed.

A quantum statistical model replaces the classical distribution by a parameter-dependent density operator ρθ\rho_\theta. A general measurement is described by a positive operator-valued measure, with effects Ex≥0E_x\geq0 satisfying ∑xEx=I\sum_xE_x=\mathbb I. The Born rule gives

p(x∣θ)=Tr⁡(ρθEx).p(x|\theta)=\operatorname{Tr}(\rho_\theta E_x).

Different measurements generally produce different classical Fisher information. Quantum Fisher information is the pointwise optimum:

FQ(θ)=sup⁡{Ex}I(θ).F_Q(\theta)=\sup_{\{E_x\}} I(\theta).

The optimization is local: a measurement attaining the supremum at one value of θ\theta need not be optimal elsewhere.

Ronald A. Fisher developed the classical information concept in the 1920s. Quantum estimation theory took shape through work by Carl W. Helstrom in the 1960s and 1970s and by Alexander S. Holevo on quantum statistical decision theory and multiparameter estimation. Samuel L. Braunstein and Carlton M. Caves established an influential geometric formulation in 1994, relating optimal statistical distinguishability to a metric on quantum states.

These developments distinguish two questions that can otherwise be conflated. One concerns how informative a chosen experimental readout is; the other concerns how informative the quantum state could be under an optimal readout. Quantum Fisher information answers the second question. An experiment may prepare a highly informative state yet extract only a small fraction of its available information because its measurement is poorly matched to the parameter.

Symmetric logarithmic derivative and explicit formulas

The symmetric logarithmic derivative, usually abbreviated SLD, is a Hermitian operator LθL_\theta defined by

∂θρθ=12(ρθLθ+Lθρθ).\partial_\theta\rho_\theta =\frac12\left(\rho_\theta L_\theta+ L_\theta\rho_\theta\right).

The corresponding quantum Fisher information is

FQ(θ)=Tr⁡(ρθLθ2).F_Q(\theta)=\operatorname{Tr}(\rho_\theta L_\theta^2).

The symmetrized equation is necessary because density operators and their derivatives generally do not commute. When they do commute, the definition reduces to the ordinary classical expression for the eigenvalue distribution.

Let

ρθ=∑aλa∣a⟩⟨a∣.\rho_\theta=\sum_a\lambda_a|a\rangle\langle a|.

In this eigenbasis,

⟨a∣Lθ∣b⟩=2⟨a∣∂θρθ∣b⟩λa+λb,λa+λb>0,\langle a|L_\theta|b\rangle = \frac{2\langle a|\partial_\theta\rho_\theta|b\rangle} {\lambda_a+\lambda_b}, \qquad \lambda_a+\lambda_b>0,

and therefore

FQ(θ)=2∑a,b:λa+λb>0∣⟨a∣∂θρθ∣b⟩∣2λa+λb.F_Q(\theta)= 2\sum_{a,b:\lambda_a+\lambda_b>0} \frac{|\langle a|\partial_\theta\rho_\theta|b\rangle|^2} {\lambda_a+\lambda_b}.

For a full-rank state, the SLD is unique. For a rank-deficient state, its action entirely within the kernel is not fixed, but that freedom does not change FQF_Q.

Where a differentiable spectral representation is available, the same information can be separated into changes of eigenvalues and changes of eigenvectors:

FQ=∑λa>0(∂θλa)2λa+2∑a≠b:λa+λb>0(λa−λb)2λa+λb∣⟨a∣∂θb⟩∣2.F_Q= \sum_{\lambda_a>0}\frac{(\partial_\theta\lambda_a)^2}{\lambda_a} + 2\sum_{a\ne b:\lambda_a+\lambda_b>0} \frac{(\lambda_a-\lambda_b)^2}{\lambda_a+\lambda_b} |\langle a|\partial_\theta b\rangle|^2.

The first term is classical information in the populations of the instantaneous eigenstates. The second records changes in their orientation in Hilbert space. At degeneracies, individual eigenvectors are not unique, so the basis-independent derivative formula is generally safer to use.

For a normalized pure state ∣ψθ⟩|\psi_\theta\rangle, the formula simplifies to

FQ=4(⟨∂θψ∣∂θψ⟩−∣⟨ψ∣∂θψ⟩∣2).F_Q= 4\left( \langle\partial_\theta\psi|\partial_\theta\psi\rangle - |\langle\psi|\partial_\theta\psi\rangle|^2 \right).

The subtraction removes changes consisting only of a global phase, which cannot affect measurement probabilities. Thus a rapidly changing mathematical representative of a state need not imply high physical distinguishability.

A simple example is the mixed qubit family

ρθ=12[I+r(cos⁡θ σx+sin⁡θ σy)],0≤r≤1,\rho_\theta= \frac12\left[ \mathbb I+r(\cos\theta\,\sigma_x+\sin\theta\,\sigma_y) \right], \qquad 0\leq r\leq1,

where σx,σy\sigma_x,\sigma_y are Pauli operators and rr is fixed. Its phase information is FQ=r2F_Q=r^2. A pure equatorial qubit has unit phase information, whereas the maximally mixed state has none. This example shows directly how a reduction of coherence can suppress sensitivity without changing the form of the parameter encoding.

Geometry, fidelity, and distinguishability

Quantum Fisher information supplies a local metric on a family of quantum states. A parameter displacement dθd\theta produces a statistically distinguishable change only to the extent that ρθ+dθ\rho_{\theta+d\theta} differs physically from ρθ\rho_\theta. This geometric interpretation explains why information is nonnegative and why redundant parameterizations can produce zero-information directions.

A central connection uses quantum fidelity. With the squared-fidelity convention,

F(ρ,σ)=[Tr⁡ρσρ]2.\mathcal F(\rho,\sigma)= \left[ \operatorname{Tr} \sqrt{\sqrt{\rho}\sigma\sqrt{\rho}} \right]^2.

The Bures distance is

DB2(ρ,σ)=2(1−F(ρ,σ)).D_B^2(\rho,\sigma)= 2\left(1-\sqrt{\mathcal F(\rho,\sigma)}\right).

For regular, fixed-rank models, its infinitesimal form is

dsB2=14FQ(θ) dθ2.ds_B^2=\frac14F_Q(\theta)\,d\theta^2.

Large quantum Fisher information therefore means that a small parameter displacement moves the state a relatively large distance in this statistical geometry. At points where the rank changes, the infinitesimal Bures expression and the pointwise SLD definition can require distinct limiting treatments.

Under a smooth reparameterization θ=θ(ϕ)\theta=\theta(\phi),

FQ(ϕ)=FQ(θ)(dθdϕ)2.F_Q^{(\phi)} = F_Q^{(\theta)} \left(\frac{d\theta}{d\phi}\right)^2.

The line element is unchanged. Quantum Fisher information itself is therefore not generally dimensionless: if θ\theta is time, it has units of inverse time squared; if θ\theta is magnetic field, it has units of inverse field squared. Numerical comparisons are meaningful only when the parameter and its physical units are specified.

For pure states, the metric is four times the Fubini–Study metric when expressed as quantum Fisher information. More generally, the quantum geometric tensor combines a metric part with an antisymmetric part related to Berry curvature. These structures connect parameter estimation with adiabatic evolution and geometric phases, although their physical roles differ: distinguishability is governed by the metric, while geometric phase accumulation involves curvature.

Quantum Fisher information also obeys a data-processing inequality. For any parameter-independent quantum channel Λ\Lambda,

FQ[Λ(ρθ)]≤FQ[ρθ].F_Q[\Lambda(\rho_\theta)]\leq F_Q[\rho_\theta].

Discarding a subsystem, applying uncontrolled noise, or replacing a quantum system by a classical measurement record cannot increase the available information. This statement does not apply without qualification when the channel itself depends on θ\theta, because the channel may then encode additional information.

Convexity gives a related restriction:

FQ ⁣[∑kpkρθ,k]≤∑kpkFQ[ρθ,k]F_Q\!\left[\sum_k p_k\rho_{\theta,k}\right] \leq \sum_kp_kF_Q[\rho_{\theta,k}]

when the mixing probabilities pkp_k are parameter independent. Forgetting which component was prepared cannot improve sensitivity. If the probabilities depend on the unknown parameter, they can carry their own classical Fisher information, and that contribution must be included.

Unitary encoding and precision scaling

Many sensors encode a parameter through unitary evolution,

ρθ=e−iθGρ0eiθG,\rho_\theta=e^{-i\theta G}\rho_0e^{i\theta G},

where GG is a Hermitian generator. For a pure input state,

FQ=4Var⁡(G).F_Q=4\operatorname{Var}(G).

Sensitivity is therefore controlled by the spread of generator eigenvalues represented in the state. A generator eigenstate acquires only an unobservable global phase and has zero information, whereas a superposition of widely separated eigenvalues can be highly sensitive.

For a mixed input with eigenvalues λa\lambda_a,

FQ=2∑a,b:λa+λb>0(λa−λb)2λa+λb∣⟨a∣G∣b⟩∣2.F_Q= 2\sum_{a,b:\lambda_a+\lambda_b>0} \frac{(\lambda_a-\lambda_b)^2}{\lambda_a+\lambda_b} |\langle a|G|b\rangle|^2.

It follows that

FQ≤4Var⁡ρ0(G),F_Q\leq4\operatorname{Var}_{\rho_0}(G),

with equality for pure states but not generally for mixed states. Classical uncertainty in the generator can produce a large variance without providing useful phase coherence. A mixed state commuting with GG has no sensitivity to this encoding even if its generator variance is nonzero.

In quantum metrology, a common model uses NN two-level probes with

Jz=12∑n=1Nσz(n).J_z=\frac12\sum_{n=1}^{N}\sigma_z^{(n)}.

An equatorial product state has FQ=NF_Q=N for a phase generated by JzJ_z. The associated local uncertainty scales as N−1/2N^{-1/2}, conventionally called shot-noise or standard-quantum-limit scaling. By contrast, the Greenberger–Horne–Zeilinger state

∣GHZ⟩=∣0⟩⊗N+∣1⟩⊗N2|\mathrm{GHZ}\rangle= \frac{|0\rangle^{\otimes N}+|1\rangle^{\otimes N}}{\sqrt2}

has FQ=N2F_Q=N^2, giving an ideal local uncertainty proportional to N−1N^{-1}.

This latter behavior is often called Heisenberg scaling. Its meaning depends on resource accounting. Probe number, number of channel uses, interrogation time, total experimental duration, and available generator strength are different resources. For a fixed generator with extreme eigenvalues gmax⁡g_{\max} and gmin⁡g_{\min}, the maximal quantum Fisher information is

FQmax⁡=(gmax⁡−gmin⁡)2,F_Q^{\max}=(g_{\max}-g_{\min})^2,

attained by an equal superposition of eigenstates at those extremes.

Time estimation illustrates the physical dimensions. With evolution generated by a Hamiltonian HH,

∣ψt⟩=e−iHt/ℏ∣ψ0⟩,FQ(t)=4(ΔH)2ℏ2.|\psi_t\rangle=e^{-iHt/\hbar}|\psi_0\rangle, \qquad F_Q(t)=\frac{4(\Delta H)^2}{\hbar^2}.

Here ℏ\hbar is the reduced Planck constant. Related formulas underlie bounds on the speed of quantum evolution.

High local sensitivity is not equivalent to unambiguous global estimation. A GHZ state accumulates a relative phase NθN\theta, producing closely spaced interference fringes. Without sufficient prior localization or a multiscale measurement strategy, different parameter values can generate the same readout statistics. Quantum Fisher information captures the local steepness of those fringes, not the resolution of their global ambiguity.

Multiple parameters and measurement incompatibility

For a parameter vector θ=(θ1,…,θd)\boldsymbol{\theta}=(\theta_1,\ldots,\theta_d), each component has an SLD LiL_i satisfying

∂iρ=12(ρLi+Liρ).\partial_i\rho= \frac12(\rho L_i+L_i\rho).

The quantum Fisher information matrix is

(FQ)ij=12Tr⁡ ⁣[ρ(LiLj+LjLi)]=Re⁡Tr⁡(ρLiLj).(F_Q)_{ij} = \frac12\operatorname{Tr}\!\left[ \rho(L_iL_j+L_jL_i) \right] = \operatorname{Re}\operatorname{Tr}(\rho L_iL_j).

It describes a quadratic form on parameter displacements. A direction with zero quadratic form is locally unidentifiable: moving along it does not change the state to first order in a statistically detectable way.

For ν\nu independent copies and a regular, locally unbiased vector estimator, the covariance matrix satisfies

Cov⁡(θ^)⪰1νFQ−1\operatorname{Cov}(\hat{\boldsymbol{\theta}}) \succeq \frac1{\nu}F_Q^{-1}

when FQF_Q is invertible. The ordering means that the difference is positive semidefinite. Singular models require restriction to identifiable combinations or a suitable generalized-inverse formulation. Even in nonsingular models, this matrix inequality need not be simultaneously attainable.

The obstacle is specifically quantum. An eigenbasis measurement of one SLD is locally optimal for its associated single parameter, but another parameter may require a different, incompatible measurement. Pairwise commuting SLDs provide a strong sufficient condition for a common optimal measurement. A weaker condition,

Tr⁡ρ[Li,Lj]=0,\operatorname{Tr}\rho[L_i,L_j]=0,

characterizes compatibility in important regular asymptotic settings. It is not equivalent to operator commutativity, and the distinction matters when collective measurements on many copies are allowed.

A pure qubit provides a concrete example:

∣ψϑ,φ⟩=cos⁡ϑ2∣0⟩+eiφsin⁡ϑ2∣1⟩.|\psi_{\vartheta,\varphi}\rangle = \cos\frac{\vartheta}{2}|0\rangle + e^{i\varphi}\sin\frac{\vartheta}{2}|1\rangle.

Its information matrix is

FQ=(100sin⁡2ϑ).F_Q= \begin{pmatrix} 1&0\\ 0&\sin^2\vartheta \end{pmatrix}.

Away from the poles, both coordinates are locally identifiable, but the measurements optimal for polar and azimuthal changes cannot generally attain both individual SLD limits simultaneously. At the poles, the vanishing azimuthal entry reflects the fact that changing φ\varphi does not change the physical state there.

The Holevo Cramér–Rao bound incorporates such trade-offs through a weighted estimation cost. In regular finite-dimensional independent-copy models, it gives the relevant asymptotically attainable collective-measurement limit. The SLD matrix remains valuable, but its inverse alone does not fully characterize a multiparameter experiment.

Unknown nuisance parameters create an additional statistical issue even before incompatibility is considered. If a desired parameter is correlated with unknown loss, contrast, or frequency drift, the relevant component of the inverse full information matrix generally differs from the inverse of its diagonal entry. Treating a nuisance parameter as known can therefore overstate precision, even when the underlying quantum Fisher information calculation is otherwise correct.

Entanglement, spin squeezing, and many-body physics

Quantum Fisher information provides operational tests of entanglement when the parameter generator is a sum of bounded local operators. For NN spin-12\tfrac12 particles and

Jn=12∑k=1Nσn(k),J_{\boldsymbol n}=\frac12\sum_{k=1}^{N} \sigma_{\boldsymbol n}^{(k)},

every fully separable state obeys

FQ[ρ,Jn]≤N.F_Q[\rho,J_{\boldsymbol n}]\leq N.

For pure product states, local variances add and each spin contributes at most one unit of information. Convexity extends the bound to mixtures of product states. Consequently, FQ>NF_Q>N certifies entanglement useful for improving sensitivity under this collective rotation.

Stronger bounds quantify entanglement depth. A kk-producible state is a mixture of states whose entangled groups contain no more than kk particles. Writing N=sk+rN=sk+r, with 0≤r<k0\leq r<k, gives

FQ≤sk2+r2F_Q\leq sk^2+r^2

for the same generator normalization. Exceeding this threshold certifies that at least k+1k+1 particles must belong to an entangled group. The statement depends on the specified tensor-product partition and local generator ranges; it is not a generator-independent characterization of entanglement.

Spin squeezing supplies an experimentally accessible route to enhanced information. For a state polarized along one axis, a common metrological squeezing parameter is

ξR2=N(ΔJ⊥)2∣⟨J⟩∣2,\xi_R^2= \frac{N(\Delta J_\perp)^2}{|\langle\boldsymbol J\rangle|^2},

where J⊥J_\perp is the measured transverse collective-spin component. For the corresponding small rotation and readout,

FQ≥NξR2.F_Q\geq\frac{N}{\xi_R^2}.

Thus ξR2<1\xi_R^2<1 certifies sensitivity beyond the separable bound. The converse does not hold: states with vanishing mean spin, including GHZ states, can have very large quantum Fisher information while conventional polarization-based squeezing parameters are uninformative.

For a pure many-body state, the relation FQ=4Var⁡(G)F_Q=4\operatorname{Var}(G) connects information with correlations. If G=∑igiG=\sum_i g_i, its variance contains both local fluctuations and connected two-point correlations between distinct sites. Correlations extending over many particles can therefore produce faster-than-linear growth with system size. This relation makes quantum Fisher information useful in condensed-matter physics, where enhanced fluctuations often accompany collective ordering or critical behavior.

A different construction studies the sensitivity of the ground state to a Hamiltonian parameter. For a nondegenerate ground state ∣0⟩|0\rangle of H(λ)H(\lambda), perturbation theory gives

FQ(λ)=4∑n>0∣⟨n∣∂λH∣0⟩∣2(En−E0)2.F_Q(\lambda)= 4\sum_{n>0} \frac{|\langle n|\partial_\lambda H|0\rangle|^2} {(E_n-E_0)^2}.

Small excitation gaps can enhance this quantity near a quantum phase transition. This ground-state sensitivity is distinct from phase estimation under a fixed generator. Finite size, symmetry restrictions, and matrix-element selection rules determine whether a particular gap closing produces a large response.

Noise, open dynamics, and error correction

Real probes interact with uncontrolled degrees of freedom, producing decoherence, loss, and other imperfections. A parameter-independent noisy channel applied after encoding cannot increase quantum Fisher information. When signal accumulation and noise occur simultaneously, however, both are part of the parameterized dynamical model, and their competition must be calculated explicitly. Information may increase initially as a signal accumulates and later decrease as coherence is lost.

Consider frequency estimation with

Uω(t)=e−iωtσz/2.U_\omega(t)=e^{-i\omega t\sigma_z/2}.

If an initially equatorial qubit undergoes phase damping with coherence amplitude e−γte^{-\gamma t}, then

FQ(ω;t)=t2e−2γt.F_Q(\omega;t)=t^2e^{-2\gamma t}.

For NN independently prepared probes, this becomes Nt2e−2γtNt^2e^{-2\gamma t}. If each experiment lasts tt, a total duration TT permits approximately T/tT/t repetitions when preparation and readout overheads are neglected. Total information is then NTt e−2γtNTt\,e^{-2\gamma t}, maximized at t=1/(2γ)t=1/(2\gamma).

An NN-particle GHZ state under independent noise of the same kind has coherence e−Nγte^{-N\gamma t} and

FQGHZ(ω;t)=N2t2e−2Nγt.F_Q^{\mathrm{GHZ}}(\omega;t) = N^2t^2e^{-2N\gamma t}.

Its enhanced phase accumulation is accompanied by proportionally faster decoherence. Optimizing the interrogation time at fixed total duration gives the same linear scaling with NN as the product-probe strategy in this particular model. This calculation explains why an ideal noiseless scaling advantage need not survive realistic resource accounting.

The result is not a universal prohibition on quantum enhancement. Correlated noise, noise with a preferred spatial direction, non-Markovian dynamics, auxiliary systems, and access to environmental records can change attainable precision. Even where asymptotic scaling is restricted to linear information growth, quantum strategies can improve constant factors. Such gains can be important in finite experiments, where preparation time and detector noise also matter.

Quantum error correction offers another route. Under standard assumptions of Markovian noise, noiseless ancillas, and arbitrarily fast controls, whether the signal Hamiltonian lies outside the real linear span generated by the identity and the Hermitian combinations of noise operators and their products determines whether error-corrected Heisenberg scaling is possible. The physical requirement is that errors can be removed without also erasing the signal. This condition is commonly called the Hamiltonian-not-in-Lindblad-span condition.

Postselection requires similarly careful accounting. A selected successful branch can have greater quantum Fisher information per retained event than the original average. Its success probability, any parameter dependence of that probability, and information in discarded outcomes must all be counted. For parameter-independent filtering, the complete record obeys data processing; enhanced sensitivity in a rare branch does not by itself establish an overall precision advantage.

Physical applications

Atomic clocks and field sensors are direct applications of quantum Fisher information. In a Ramsey sequence, a first pulse creates an atomic superposition, free evolution encodes a detuning-dependent phase, and a final pulse converts that phase into populations. For NN independent ideal two-level atoms interrogated for time tt, the frequency information is Nt2Nt^2. Entangled inputs can increase it in ideal conditions, while finite coherence, oscillator noise, and imperfect detection determine how much is accessible. The same framework applies to magnetic-field sensing when a Zeeman interaction converts field strength into a spin rotation.

Optical interferometry provides another important setting. A relative phase between two modes can be generated by

G=na−nb2,G=\frac{n_a-n_b}{2},

where nan_a and nbn_b are photon number operators. Quantum Fisher information distinguishes the sensitivities available from coherent illumination, squeezed states, and entangled fixed-photon-number states. Photon loss is especially consequential for superpositions whose useful coherence depends on preserving every photon. Comparisons must specify whether resources are counted by incident photons, detected photons, optical energy, or total measurement time.

Quantum thermometry estimates temperature from a system’s thermal state. For a temperature-independent Hamiltonian,

ρT=e−H/(kBT)Z,\rho_T=\frac{e^{-H/(k_BT)}}{Z},

where ZZ is the partition function, all members of the family commute. Their quantum Fisher information is therefore classical information in the energy populations:

FQ(T)=Var⁡(H)kB2T4=CkBT2.F_Q(T) = \frac{\operatorname{Var}(H)}{k_B^2T^4} = \frac{C}{k_BT^2}.

Here kBk_B is the Boltzmann constant and C=∂T⟨H⟩C=\partial_T\langle H\rangle is the heat capacity. Energy measurement is optimal for this equilibrium model. The identity assumes a fixed Hamiltonian and does not automatically apply to strongly coupled probes described by a temperature-dependent effective Hamiltonian.

Quantum imaging shows why optimizing the measurement can change an apparent resolution limit. For two equally bright, mutually incoherent point sources imaged through a symmetric diffraction-limited system, direct image-plane detection can yield Fisher information about their separation that vanishes as the separation becomes small. The quantum Fisher information has a nonzero small-separation limit in standard models. Spatial-mode measurements can access information missed by direct detection. Source imbalance, unknown centroid, background, and behavior at exactly coincident sources require separate analysis.

Quantum Fisher information also characterizes estimation of damping rates, coupling constants, spectral parameters, and weak forces. In spectroscopy, for example, a parameter can alter both the populations and coherences of a probe, making a population-only readout suboptimal. In each application, the quantum value is a benchmark rather than a complete experimental prediction. Restricted detector access, finite calibration accuracy, drift, and nuisance parameters determine whether a proposed measurement can approach it.

Calculation, experimental certification, and limitations

For a finite-dimensional model with known ρθ\rho_\theta and ∂θρθ\partial_\theta\rho_\theta, the spectral formula provides a direct numerical method. An alternative solves the linear matrix equation

ρL+Lρ=2∂θρ\rho L+L\rho=2\partial_\theta\rho

and evaluates Tr⁡(ρL2)\operatorname{Tr}(\rho L^2). This is a Sylvester equation. Rank deficiency requires support-aware treatment because the equation does not determine the kernel-to-kernel block of LL. Small eigenvalues can also make calculations sensitive to numerical errors in the state or its derivative.

For a full-rank qubit written as

ρθ=12(I+rθ⋅σ),∣rθ∣<1,\rho_\theta=\frac12\left(\mathbb I+\boldsymbol r_\theta\cdot\boldsymbol\sigma\right), \qquad |\boldsymbol r_\theta|<1,

an explicit formula is

FQ=∣∂θr∣2+(r⋅∂θr)21−∣r∣2.F_Q= |\partial_\theta\boldsymbol r|^2+ \frac{(\boldsymbol r\cdot\partial_\theta\boldsymbol r)^2} {1-|\boldsymbol r|^2}.

For a smooth pure-state family on the Bloch sphere,

FQ=∣∂θr∣2.F_Q=|\partial_\theta\boldsymbol r|^2.

The full-rank formula should not be used by blindly substituting ∣r∣=1|\boldsymbol r|=1, because its denominator vanishes. A family staying on the pure-state surface and one approaching it through changing eigenvalues can have different local behavior.

Large many-body systems often make full density-matrix reconstruction impractical. A useful lower bound follows from measuring a parameter-independent observable MM:

FQ≥∣∂θ⟨M⟩∣2Var⁡(M)F_Q\geq \frac{|\partial_\theta\langle M\rangle|^2} {\operatorname{Var}(M)}

when the variance is nonzero and regularity conditions hold. This connects error propagation to optimal information. It certifies how much information is accessible through a particular signal and noise measurement, but equality is not automatic. Full outcome statistics can contain more information than the mean and variance alone reveal.

For a regular single-parameter problem, measuring the SLD eigenbasis at the true operating point attains the quantum Fisher information locally. Since that point is unknown, practical schemes can first obtain a rough estimate and then adjust subsequent measurements. Such adaptive procedures can approach the quantum Cramér–Rao limit asymptotically in identifiable models. Their finite-sample performance still depends on initialization, parameter range, available controls, and possible ambiguities in the likelihood.

Rank-changing families expose an important limitation of purely pointwise calculations. Consider

ρθ=(1−θ2)∣0⟩⟨0∣+θ2∣1⟩⟨1∣,∣θ∣<1.\rho_\theta=(1-\theta^2)|0\rangle\langle0| +\theta^2|1\rangle\langle1|, \qquad |\theta|<1.

For θ≠0\theta\ne0,

FQ(θ)=41−θ2.F_Q(\theta)=\frac{4}{1-\theta^2}.

At θ=0\theta=0, however, ∂θρθ=0\partial_\theta\rho_\theta=0, so the pointwise SLD information is zero. The limit from neighboring points is four, and the local Bures distance also detects a nonzero second-order separation. The model additionally cannot distinguish θ\theta from −θ-\theta. At such singular points, ordinary local unbiasedness and smooth-metric arguments cannot be transferred without examining the statistical model itself.

These distinctions also separate quantum Fisher information from Bayesian error measures. A prior distribution, a finite parameter interval, or a nonquadratic estimation cost can make global bounds more appropriate than a local Cramér–Rao expression. Quantum Fisher information remains a precise description of local state sensitivity, but an operational precision claim must also identify the estimator, measurement class, repetitions, and resource constraints.