Quantum Measurement as Estimation

A quantum measurement does not normally reveal an unknown parameter directly. It produces a random classical outcome whose distribution depends on that parameter. Estimation is the subsequent statistical task of converting one or more outcomes into an estimate, interval, decision, or control action.

The basic chain is

$$\theta \longrightarrow \rho_\theta \longrightarrow x \sim p(x\mid\theta) \longrightarrow \widehat\theta(x).$$

The quantum model determines which outcome distributions are possible. Statistical inference then connects a finite record to the unknown parameter. This page develops that connection, including detector errors, calibration, phase ambiguity, and uncertainty. Quantum Sensing explains the accompanying decoherence and measurement backaction.

Required background. Use density operators to describe an encoded probe, POVMs to compute outcome probabilities, and Fisher information to recognize a local sensitivity measure.

Define the Estimand First

The estimand is the quantity the analysis promises to recover. It need not be an eigenvalue of an observable.

Task	Estimand	Typical data	Appropriate output
phase sensing	periodic phase $\theta\in[0,2\pi)$	binary outcomes or photon counts	circular estimate and interval
spectroscopy	frequency, detuning, or line center	counts versus interrogation setting	estimate with calibration and line-shape uncertainty
magnetometry	field amplitude, component, or waveform	spin readout records	scalar, vector, or time series
loss sensing	channel transmissivity $T$	click and no-click events	bounded estimate or one-sided interval
thermometry	temperature $T_{\mathrm{temp}}>0$	energies, populations, or trajectories	positive estimate and uncertainty
hypothesis testing	one of several states or channel models	measurement outcomes	decision and type-I/type-II errors

Several superficially similar targets can require different analyses. Estimating an instantaneous field value, a time average, a Fourier amplitude, and a model coupling are different tasks even when the same sensor produces the records.

Identifiability

A parameter is identifiable only if distinct allowed values lead to distinguishable data models. If

$$p(D\mid\theta,\lambda) = p(D\mid\theta',\lambda')$$

for every possible data set $D$, then the experiment cannot distinguish $(\theta,\lambda)$ from $(\theta',\lambda')$. More data from the same design do not cure structural non-identifiability. One needs an additional setting, calibration, prior restriction, or different measurement.

Identifiability is global. In a regular model with strictly positive outcome probabilities, a vanishing probability derivative gives zero Fisher information in that direction. At zero-probability outcomes or changing support, this argument needs care: the information defined at a singular point can differ from its nearby limit. A singular Fisher matrix also need not imply global non-identifiability; higher-order parameter dependence can remain.

From a Quantum Model to a Likelihood

Let $u$ denote preparation and control settings, $\theta$ the target parameter, and $\lambda$ nuisance parameters. A general parameter-dependent output state can be written

$$\rho_{\theta,\lambda}^{\,u} = \mathcal E_{\theta,\lambda}^{\,u} \left( \rho_0^{\,u} \right),$$

where $\mathcal E_{\theta,\lambda}^{\,u}$ is a quantum channel. A fixed environment initially uncorrelated with the probe is a sufficient condition for this description across preparations. Preparation-dependent initial correlations may require an enlarged model. Dissipation, control error, and loss can be included; loss requires an explicit lost or vacuum outcome if the output space otherwise contains only the surviving probes.

For a POVM $\{M_x^{\,v}\}$ selected by a known measurement setting $v$, independent of the unknown parameter,

$$M_x^{\,v}\succeq0, \qquad \sum_x M_x^{\,v}=I,$$

the Born rule gives the likelihood for one outcome:

$$p(x\mid\theta,\lambda,u,v) = \operatorname{Tr} \left[ M_x^{\,v} \rho_{\theta,\lambda}^{\,u} \right].$$

Once the state, channel, and POVM are fixed, this is an ordinary classical statistical model. Quantum theory constrains which likelihoods can be generated and which measurements are physically available; the estimator itself acts on classical data.

Measurement instruments and sequential data

A POVM gives outcome probabilities but not the conditional state required for later measurements. A quantum instrument $\{\mathcal I_x\}$ supplies both:

$$p(x\mid\rho) = \operatorname{Tr} \left[ \mathcal I_x(\rho) \right],$$ $$\rho_x = \frac{ \mathcal I_x(\rho) }{ p(x\mid\rho) }.$$

Sequential likelihoods depend on these state updates. If the next control depends on the observed history $x_{<k}=(x_1,\ldots,x_{k-1})$, then

$$\begin{aligned} p(x_{1:n}\mid\theta) = \prod_{k=1}^{n} p\left( x_k \middle| x_{<k}, \theta, u_k(x_{<k}) \right). \end{aligned}$$

Adaptive data are not identically distributed, but they still have a well-defined likelihood when the policy is specified. Ignoring the conditioning can double-count information or assign the wrong uncertainty.

Detector and Readout Models

The ideal POVM outcome $x$ may not equal the recorded symbol $y$. A classical readout channel $C(y\mid x)$ produces

$$q(y\mid\theta) = \sum_x C(y\mid x) p(x\mid\theta).$$

For a binary detector with false-positive probability $\alpha$ and false-negative probability $\beta$,

$$q(1\mid\theta) ={} (1-\beta) p(1\mid\theta) + \alpha \left[ 1-p(1\mid\theta) \right].$$

The analysis should fit or otherwise use $q$, not pretend the observed counts came from $p$. Algebraically inverting a confusion matrix can create negative corrected frequencies and unstable uncertainties. A forward likelihood respects probability constraints when its parameters are restricted to their physical domains. It propagates calibration uncertainty only if $\alpha$ and $\beta$ are themselves included in a joint fit, profile analysis, or posterior; fixing their estimated values does not do this automatically. Measurement Error Mitigation compares forward inference with response-matrix inversion.

Loss and postselection are also part of the outcome model. A no-click event may carry information and cannot be discarded merely because it is inconvenient. If acceptance depends on $\theta$, accepted records follow a conditional likelihood. Using the unconditional model on them can bias the estimate. Correct conditioning can support valid inference, but discarding acceptance counts loses their information and does not remove the cost of rejected trials.

Encoding Models

Unitary encoding

For a scalar phase,

$$\rho_\theta = e^{-i\theta G} \rho_0 e^{i\theta G},$$

where $\theta$ and the Hermitian generator $G$ are dimensionless. The phase may include a physical coupling and time, such as

$$\theta = \gamma BT.$$

Estimating $B$, $\gamma$, or $T$ is not the same task unless the other factors are known. Calibration uncertainty in $\gamma$ or $T$ becomes a nuisance parameter.

Channel encoding

Many targets change a channel rather than a closed-system phase:

$$\rho_\theta = \mathcal E_\theta(\rho_0).$$

Examples include transmissivity, dephasing rate, relaxation time, temperature-dependent dissipation, and an unknown force acting during open evolution. Ancillas and controls may improve channel discrimination, but the output likelihood must still include every retained mode and loss boundary.

Waveform encoding

A time-dependent signal $\theta(t)$ may enter a Hamiltonian

$$H(t) = H_0 + \theta(t)A.$$

Here $A$ has units of energy divided by the signal units; it is not necessarily the dimensionless phase generator above. The estimand might be the entire waveform, a filtered component, or coefficients in a basis expansion,

$$\theta(t) = \sum_{j=1}^{m} c_j f_j(t).$$

An infinite-dimensional waveform cannot be recovered from finite data without smoothness, bandwidth, sparsity, prior, or model assumptions. The reported resolution is therefore a property of both the sensor and the function class.

Estimators, Decisions, and Loss

An estimator is a function of the observed data:

$$\widehat\theta = \delta(D).$$

There is no universally best estimator independent of the loss, parameter range, and sampling model.

Frequentist risk

For loss $L(\widehat\theta,\theta)$, the frequentist risk is

$$R(\theta,\delta) = \mathbb E_\theta \left[ L( \delta(D), \theta ) \right].$$

With squared-error loss,

$$L( \widehat\theta,\theta ) = ( \widehat\theta-\theta )^2,$$

the risk is mean-square error. If

$$b(\theta) = \mathbb E_\theta[ \widehat\theta ] - \theta$$

is the bias, then

$$\operatorname{MSE}_\theta( \widehat\theta ) = \operatorname{Var}_\theta( \widehat\theta ) + b(\theta)^2.$$

A low-variance but strongly biased estimator may be worse than a noisier unbiased one.

Likelihood methods

The maximum-likelihood estimator is

$$\widehat\theta_{\mathrm{ML}} \in \operatorname*{arg\,max}_{\theta} p(D\mid\theta).$$

MLEs are often asymptotically efficient in regular identifiable models, but they can be biased at finite sample size, nonunique in periodic models, and pinned to boundaries. A likelihood value is not by itself a confidence probability for $\theta$.

Bayesian methods

Given a prior density $\pi(\theta)$,

$$\pi(\theta\mid D) = \frac{ p(D\mid\theta)\pi(\theta) }{ \int p(D\mid\vartheta) \pi(\vartheta) d\vartheta }.$$

A Bayes action minimizes posterior expected loss:

$$\widehat\theta_{\mathrm{B}} \in \operatorname*{arg\,min}_{a} \int L(a,\theta) \pi(\theta\mid D) d\theta.$$

For a real scalar, squared loss gives the posterior mean, absolute loss gives a posterior median, and zero–one loss in a discrete problem gives a maximum-posterior-probability decision. The prior is part of the model and should be justified, varied in a robustness analysis, and kept distinct from calibration data.

Minimax and hypothesis testing

A minimax rule controls worst-case risk,

$$\inf_\delta \sup_{\theta\in\Theta} R(\theta,\delta).$$

In hypothesis testing, the output is a decision rather than a point estimate. Performance is described by type-I and type-II errors, receiver-operating curves, or a cost-weighted decision risk. Estimation and testing can use the same measurement data but answer different questions.

Choose a Loss That Respects the Parameter

Ordinary squared error does not suit every parameter space.

For a phase $\theta\in[0,2\pi)$, values near $0$ and $2\pi$ are close. A circular loss such as

$$L_{\mathrm{circ}} \left( \widehat\theta,\theta \right) = 4 \sin^2 \left( \frac{ \widehat\theta-\theta }{2} \right)$$

respects periodicity.

For a vector parameter, a weighted quadratic loss

$$L = ( \widehat{\boldsymbol\theta} - \boldsymbol\theta )^{\mathsf T} W ( \widehat{\boldsymbol\theta} - \boldsymbol\theta )$$

states which combinations matter. The units and scale in $W$ are substantive choices, not mathematical decoration.

For rare-event detection, missing a true signal and raising a false alarm may have very different costs. A credible sensing claim identifies the decision loss actually relevant to the application.

Local Information Versus Global Estimation

Classical and Quantum Fisher Information describes local curvature of a measured likelihood and its optimization over quantum measurements. Neither quantity alone solves a global estimation problem.

Consider a qubit prepared in

$$|+\rangle = \frac{ |0\rangle+|1\rangle }{\sqrt2}$$

and encoded by

$$U_\theta = e^{-i\theta\sigma_z/2}.$$

Measuring $X$ gives

$$p_X(+\mid\theta) = \frac{ 1+\cos\theta }{2}.$$

The data cannot distinguish $\theta$ from $-\theta$ or from $\theta+2\pi k$. This is a global ambiguity even where the local Fisher information is nonzero.

Measuring $Y$ instead gives

$$p_Y(+\mid\theta) = \frac{ 1+\sin\theta }{2},$$

which resolves the sign near $\theta=0$ but remains periodic globally. A known phase offset, several settings, an adaptive policy, or a prior interval can select a branch.

The quantum Cramér–Rao bound can be locally tight while global mean-square error remains large because of branch mistakes. Conversely, a globally robust measurement may sacrifice some local Fisher information to increase dynamic range.

Worked Example: Finite-Count Ramsey Estimation

Suppose a Ramsey-type experiment has contrast $C$, phase offset $\phi_0$, and plus-outcome probability

$$p_+( \theta,C,\phi_0 ) = \frac{ 1+ C\sin( \theta+\phi_0 ) }{2}.$$

From $n$ independent shots with $k$ plus outcomes,

$$p( k\mid \theta,C,\phi_0 ) = \binom nk p_+^k ( 1-p_+ )^{n-k}.$$

Assume known $0<C\leq1$ and $\phi_0$, and choose the increasing branch $\theta+\phi_0\in[-\pi/2,\pi/2]$. When the observed fraction gives an arcsine argument in $[-1,1]$, the MLE is

$$\widehat\theta_{\mathrm{ML}} = \arcsin \left[ \frac{ 2k/n-1 }{C} \right] - \phi_0.$$

For every finite record, the binomial likelihood is maximized by projecting $k/n$ onto the attainable probability interval $[(1-C)/2,(1+C)/2]$. Equivalently, clip $(2k/n-1)/C$ to $[-1,1]$ before taking the arcsine. An out-of-range fraction therefore selects a branch endpoint; it does not make the data invalid. A narrower allowed phase interval requires its own attainable probability interval.

The inverse must match the chosen branch. On the decreasing branch $\theta+\phi_0\in[\pi/2,3\pi/2]$, the inverse for $\theta+\phi_0$ is $\pi-\arcsin$ of the same clipped argument. For example, $C=0.8$, $\phi_0=0$, and $k/n=0.7$ give $\widehat\theta=5\pi/6$ on that branch, while the principal arcsine gives $\pi/6$ outside it. If both branches are allowed, these two phases have identical likelihoods: more shots at this setting cannot select between them. At $C=0$, the likelihood is phase independent and no phase MLE is unique.

If $C$ is also unknown, one setting identifies only the combination

$$C\sin( \theta+\phi_0 ).$$

Varying the known analysis phase, measuring another quadrature, or adding calibration data is required to separate phase from contrast. If the offset itself is unknown, the data also confound it with $\theta$. Treating fitted nuisance parameters as exact generally understates uncertainty; the penalty can vanish at a locally orthogonal operating point.

Nuisance Parameters and Calibration

Write the full likelihood as

$$p(D\mid\theta,\boldsymbol\lambda),$$

where $\boldsymbol\lambda$ may include contrast, detection efficiency, background rate, phase offset, drift, temperature, coupling strength, or timing error.

For a regular local model with Fisher matrix

$$F = \begin{pmatrix} F_{\theta\theta} & F_{\theta\lambda} \\ F_{\lambda\theta} & F_{\lambda\lambda} \end{pmatrix},$$

the effective information about $\theta$ after allowing $\lambda$ to be unknown is the Schur complement

$$F_{\mathrm{eff}} = F_{\theta\theta} - F_{\theta\lambda} F_{\lambda\lambda}^{-1} F_{\lambda\theta},$$

when the inverse exists. Since the subtracted term is positive semidefinite, nuisance uncertainty cannot improve the local bound without additional information.

Calibration data can be included through a joint likelihood,

$$p( D_{\mathrm{science}}, D_{\mathrm{cal}} \mid \theta,\lambda ) = p( D_{\mathrm{science}} \mid \theta,\lambda ) p( D_{\mathrm{cal}} \mid \lambda ),$$

when the two data sets are conditionally independent. Treating an estimated calibration constant as exact discards its uncertainty.

Transmission and efficiency degeneracy

If a single photon is transmitted with probability $T$ and detected with efficiency $\eta$, an idealized click probability is

$$p_{\mathrm{click}} = \eta T.$$

From click/no-click data at one setting, only the product $\eta T$ is identifiable. Increasing the number of trials estimates that product more precisely but does not separate $T$ from $\eta$. A calibrated efficiency, a reference path, or another design is necessary.

Multiparameter Quantum Estimation

For several parameters, the classical Fisher information becomes a matrix and the estimator has a covariance matrix. Quantum optimization is subtler than maximizing each diagonal element independently.

Measurements optimal for different parameters may be incompatible. A quantum Fisher matrix can bound each local direction yet fail to be jointly attainable by one measurement on one copy. Collective measurements, weighted costs, Holevo-type bounds, and asymptotic protocols may be needed.

Classical and Quantum Fisher Information develops the SLD quantum Fisher matrix and measurement compatibility. Cramér–Rao Bounds explains when these local metrics bound an estimator’s error. The cost and physically available measurements determine which bound is relevant.

Continuous Records and Correlated Data

A continuously monitored system produces a time series rather than independent shots. The record might be a photocurrent, homodyne trace, sequence of quantum jumps, or thresholded event stream.

The likelihood is determined by the conditional dynamics. Schematically,

$$p( D_{0:T} \mid \theta ) = \prod_k p( dD_k \mid D_{<k}, \theta ).$$

A filter estimates the current conditional state or parameter using past data. A smoother may estimate a past signal using both earlier and later records. These are different causal tasks.

Correlations invalidate counting time samples as independent repetitions. Positive autocorrelation often increases the variance of a sample mean, but correlation does not universally decrease parameter information: correlations can themselves encode the parameter. Detector bandwidth, discretization, unobserved channels, and model mismatch enter the trajectory likelihood. Bayesian Quantum Measurement develops conditional state inference from records.

A state estimator reconstructs $\rho_t$ or its sufficient coordinates. A parameter estimator infers a fixed or time-varying model parameter. The two may be coupled, but they should not be conflated.

Experimental Design and Adaptation

Controls $u$ and measurement settings $v$ determine the likelihood. Design is therefore part of estimation.

Possible design criteria include:

• expected Fisher information near a nominal parameter;
• expected posterior variance or entropy;
• worst-case risk over an interval;
• probability of resolving competing hypotheses;
• robustness to nuisance parameters and drift;
• information gained per unit time, energy, or sample exposure.

An adaptive experiment chooses the next setting from earlier outcomes:

$$u_{k+1} = f( x_1,\ldots,x_k ).$$

This can keep an interferometer near a sensitive fringe, resolve phase branches, or concentrate data where competing models differ. The policy must be included in the likelihood and resource count.

Using the same data to choose a model, tune a stopping rule, and report an unadjusted interval can invalidate nominal coverage. Predeclared policies, sequential-analysis corrections, simulation-based calibration, or held-out validation data address different parts of this problem.

Uncertainty Is Part of the Result

A point estimate without uncertainty is incomplete.

Confidence regions

A $1-\alpha$ confidence procedure is designed so that, under repeated data generated at the true parameter,

$$\Pr_\theta[ \theta \in C(D) ] \ge 1-\alpha$$

over the declared model class. The probability statement concerns the procedure’s repeated-sampling coverage.

Credible regions

A Bayesian credible region $B(D)$ satisfies

$$\int_{B(D)} \pi( \theta\mid D ) d\theta = 1-\alpha.$$

This is a posterior probability statement conditional on the model and prior. A credible region need not have frequentist coverage, though some procedures can have both properties approximately or by design.

Systematic and model uncertainty

Repeated quantum shots reduce statistical uncertainty. They do not average away:

• a wrong Hamiltonian or channel model;
• an unknown calibration offset shared by every trial;
• drift outside the modeled stochastic process;
• selection bias;
• waveform components outside the assumed basis;
• detector nonlinearities absent from the likelihood.

Robustness checks should vary plausible model classes, nuisance priors, calibration assumptions, and data partitions. A narrow interval from a misspecified model is precise, not trustworthy.

Checking an inferred parameter

Compare predictions with counts at settings not used to fit the model, inspect residuals over time, and fit simulated records with known parameters to test bias and interval coverage. Signal injections can check calibration independently. A good fit alone does not establish identifiability: different parameter pairs may predict the same records.

An expectation value is a model prediction, not a finite-data estimate. Likewise, QFI optimizes local measurement sensitivity but supplies neither a detector nor a global branch-selection rule. Quantum Phase Estimation is a particular coherent eigenphase algorithm, distinct from the broader statistical task discussed here.

Exercises

1. Find a phase ambiguity

A qubit phase experiment measures $X$ and has

$$p(+\mid\theta) = \frac{1+\cos\theta}{2}.$$

Name all parameter symmetries that leave the likelihood unchanged. Suggest one local remedy near $\theta=0$.

Solution

The likelihood is invariant under

$$\theta \mapsto -\theta$$

and under

$$\theta \mapsto \theta+2\pi k, \qquad k\in\mathbb Z.$$

Measuring $Y$ gives $(1+\sin\theta)/2$, which distinguishes the sign locally near zero. Equivalently, apply a known $\pi/2$ analysis phase before the $X$ measurement. Global $2\pi$ periodicity remains and requires a restricted interval, multiple scales, or a circular decision rule.

2. Include readout confusion

An ideal binary model has $p(1\mid\theta)=\theta$, with $0\le\theta\le1$. The detector has false-positive probability $\alpha=0.02$ and false-negative probability $\beta=0.08$. Find the observed success probability $q(1\mid\theta)$.

Solution

The observed one can arise from a correctly detected ideal one or a false positive:

$$\begin{aligned} q(1\mid\theta) &= (1-\beta)\theta + \alpha(1-\theta) \\ &= 0.92\theta + 0.02 - 0.02\theta \\ &= 0.02+0.90\theta. \end{aligned}$$

Fitting the ideal model directly to observed frequencies would produce a biased estimate of $\theta$.

3. Bayesian Bernoulli update

For $k$ successes in $n$ Bernoulli trials with success probability $\theta$, use a uniform prior on $[0,1]$. Find the posterior distribution and the Bayes estimator under squared-error loss.

Solution

The likelihood is proportional to

$$\theta^k (1-\theta)^{n-k}.$$

Multiplying by the uniform prior gives

$$\theta\mid D \sim \operatorname{Beta} \left( k+1, n-k+1 \right).$$

Squared-error loss selects the posterior mean:

$$\widehat\theta_{\mathrm B} = \frac{k+1}{n+2}.$$

This differs from the MLE $k/n$ and is generally biased in the frequentist sense at finite $n$.

4. Use circular loss

Two equally probable posterior phase values are $\epsilon$ and $2\pi-\epsilon$, with small $\epsilon>0$. What does an ordinary linear posterior mean give? What circular estimate is physically sensible?

Solution

Treating the interval as a line gives

$$\frac{ \epsilon+ (2\pi-\epsilon) }{2} = \pi,$$

which lies opposite both probable phases.

On the circle, both values lie near zero. Their unit phasors are complex conjugates with positive real part, so the circular mean direction is

$$\widehat\theta_{\mathrm{circ}} = 0 \pmod{2\pi}.$$

The loss and estimator must respect periodic geometry.

5. Nuisance-parameter penalty

A two-parameter Fisher matrix is

$$F = \begin{pmatrix} 100&30\\ 30&25 \end{pmatrix},$$

where the first parameter is the target. Find the effective information when the second parameter is unknown.

Solution

The Schur complement is

$$\begin{aligned} F_{\mathrm{eff}} &= 100 - 30 \frac{1}{25} 30 \\ &= 100-36 \\ &= 64. \end{aligned}$$

The local variance bound is therefore $1/64$ rather than the misleading known-nuisance value $1/100$ for one use of this total Fisher matrix.

6. Diagnose transmission non-identifiability

A click probability is $p_{\mathrm{click}}=\eta T$. Alice observes enough data to determine $p_{\mathrm{click}}=0.42$ almost exactly. List three pairs $(\eta,T)$ consistent with the result and name one way to identify $T$.

Solution

Examples include

$$(\eta,T) \in \{ (0.6,0.7), (0.7,0.6), (0.84,0.5) \}.$$

All produce $0.42$. Alice can identify $T$ by independently calibrating $\eta$, adding a reference measurement with known transmission, or changing the design so $\eta$ and $T$ affect different recorded statistics.

7. Factor an adaptive likelihood

The first binary measurement uses setting $u_1$. If $x_1=1$, the second setting is $u_A$; otherwise it is $u_B$. Write the two-step likelihood.

Solution

The joint likelihood is

$$p( x_1,x_2 \mid \theta ) ={} p( x_1 \mid \theta,u_1 ) \times p( x_2 \mid x_1,\theta,u_2(x_1) ),$$

where

$$u_2(x_1) = \begin{cases} u_A,&x_1=1,\\ u_B,&x_1=0. \end{cases}$$

The second factor cannot be replaced by an unconditional identical-trial distribution because its setting and input state may depend on $x_1$.

References

• O. E. Barndorff-Nielsen, R. D. Gill, and P. E. Jupp, “On quantum statistical inference,” Journal of the Royal Statistical Society: Series B 65, 775–816 (2003), doi:10.1111/1467-9868.00415.
• S. L. Braunstein and C. M. Caves, “Statistical distance and the geometry of quantum states,” Physical Review Letters 72, 3439–3443 (1994), doi:10.1103/PhysRevLett.72.3439.
• S. Gammelmark and K. Mølmer, “Bayesian parameter inference from continuously monitored quantum systems,” Physical Review A 87, 032115 (2013), doi:10.1103/PhysRevA.87.032115.
• C. W. Helstrom, Quantum Detection and Estimation Theory, Academic Press (1976), doi:10.1016/C2013-0-10310-8.
• A. S. Holevo, Probabilistic and Statistical Aspects of Quantum Theory, 2nd ed., Edizioni della Normale (2011), doi:10.1007/978-88-7642-378-9.
• E. L. Lehmann and G. Casella, Theory of Point Estimation, 2nd ed., Springer (1998), doi:10.1007/b98854.
• M. G. A. Paris, “Quantum estimation for quantum technology,” International Journal of Quantum Information 7, 125–137 (2009), doi:10.1142/S0219749909004839.
• M. Tsang, H. M. Wiseman, and C. M. Caves, “Fundamental quantum limit to waveform estimation,” Physical Review Letters 106, 090401 (2011), doi:10.1103/PhysRevLett.106.090401.
• A. Wald, Statistical Decision Functions, Wiley (1950), doi:10.1002/9780470316511.