Why Generalized Measurements Are Needed

Generalized measurements are needed because most laboratory measurements are not ideal projective measurements on the system Hilbert space alone.

This does not make projective measurement obsolete. Projective measurements remain the sharp ideal limit and the right model for many textbook calculations. The point is narrower and more practical:

$$\text{realistic outcome statistics} \quad \text{and} \quad \text{realistic post-measurement states} \quad \text{often need more data than a PVM.}$$

The generalized-measurement hierarchy keeps those data separate:

Object	What It Answers	What It Does Not Answer Alone
PVM	sharp projective probabilities and ideal projective update	detector inefficiency, weak readout, indirect coupling
POVM	outcome probabilities for general effects	post-measurement state
Kraus operators	one representation of a state transformation	unique physical mechanism
instrument	outcome probabilities and state updates	interpretation beyond the operational model
channel	nonselective evolution after ignoring outcomes	which outcome occurred

The rest of this chapter gives the formal pages: POVMs, Kraus Operators, Quantum Instruments, Naimark Dilation, and Stinespring Dilation. This page explains why those objects are unavoidable.

Projective Measurements Are an Ideal Limit

An ideal discrete projective measurement is described by projectors $\{P_a\}$ satisfying

$$P_aP_b=\delta_{ab}P_a, \qquad \sum_aP_a=I.$$

For input state $\rho$, the outcome probabilities are

$$p(a)=\operatorname{Tr}(P_a\rho).$$

In the Lüders idealization, the selective state is

$$\rho_a = \frac{P_a\rho P_a} {\operatorname{Tr}(P_a\rho)}.$$

This model is powerful because it is sharp, repeatable, and algebraically simple. But each of those words is an idealization. A real detector can be inefficient, noisy, weak, indirect, destructive, coarse grained, finite-resolution, or mediated by an environment. Then the outcome probabilities may still be well defined, but not by orthogonal projectors on the original system alone.

Inefficient Detectors

A detector may fail to click even when the system is in the state it is designed to detect. Suppose a qubit detector is intended to click on $\lvert1\rangle$ with efficiency $\eta$, where $0\le\eta\le1$. A simple probability model is

$$F_{\mathrm{click}} = \eta\lvert1\rangle\langle1\rvert,$$

and

$$F_{\mathrm{no}} = \lvert0\rangle\langle0\rvert + (1-\eta)\lvert1\rangle\langle1\rvert.$$

These effects are positive and complete:

$$F_{\mathrm{click}}+F_{\mathrm{no}}=I.$$

For $0\lt\eta\lt1$, the click effect is not a projector because

$$F_{\mathrm{click}}^2 = \eta^2\lvert1\rangle\langle1\rvert \ne F_{\mathrm{click}}.$$

Thus inefficient detection is naturally a POVM, not a PVM. A separate instrument is still needed to specify what state remains after a click or no-click event.

Noisy Readout and Mislabeling

Even when a detector always returns a label, the label may be wrong with some probability. A noisy two-outcome $\sigma_z$ readout can be represented by effects

$$F_0 = (1-\epsilon)\lvert0\rangle\langle0\rvert + \epsilon\lvert1\rangle\langle1\rvert,$$

and

$$F_1 = \epsilon\lvert0\rangle\langle0\rvert + (1-\epsilon)\lvert1\rangle\langle1\rvert, \qquad 0\le\epsilon\le\frac{1}{2}.$$

The probabilities include classical readout confusion. The detector may still strongly dephase the qubit, weakly disturb it, reset it, or leave it nearly unchanged. The POVM only describes the reported labels. The backaction belongs to the instrument.

This is the first major reason for the generalized framework:

$$\text{the same effects can have different physical state updates.}$$

Coarse-Grained Outcomes

An apparatus may have fine-grained microscopic outcomes but report only a coarse label. If fine outcomes are $(m,\alpha)$ with effects $F_{m\alpha}$, the reported effect is

$$F_m = \sum_\alpha F_{m\alpha}.$$

The probability for the coarse outcome is

$$p(m)=\operatorname{Tr}(\rho F_m).$$

Coarse graining is common:

• a detector bins a continuous pointer into finite windows;
• several decay channels are grouped as a single “click”;
• a spectrometer reports an energy range rather than a sharp eigenvalue;
• a degenerate eigenspace is reported as one outcome while hidden labels are partly resolved.

The coarse-grained probability is easy to write. The coarse-grained post-measurement state can be more subtle because different microscopic alternatives may leave different states.

Indirect Measurements

Many measurements do not couple a detector directly to the target observable. Instead, the system interacts with an ancilla, probe pulse, resonator, pointer, or environment, and that auxiliary degree of freedom is measured.

A schematic indirect measurement starts with a system $S$ and apparatus $A$:

$$\rho_S\otimes\eta_A \quad \xrightarrow{\ U\ } \quad U(\rho_S\otimes\eta_A)U^\dagger.$$

Then a projective pointer measurement $\{Q_m\}$ is made on the apparatus. The system outcome probability can be written as

$$p(m) = \operatorname{Tr}_S(\rho_S F_m)$$

for some positive system effect $F_m$. Even when the final pointer measurement is projective on $A$, the induced measurement on $S$ is generally a POVM.

This is the physical content behind Naimark Dilation: a generalized measurement on a system can be represented as a projective measurement on a larger Hilbert space.

Weak and Unsharp Measurements

A weak measurement extracts only partial information in one shot. A simple weak qubit readout in the $\sigma_z$ basis can have effects

$$F_\pm = \frac{1}{2} \left( I\pm\epsilon\sigma_z \right), \qquad 0\le\epsilon\le1.$$

When $\epsilon=1$, this becomes a sharp projective measurement. When $\epsilon$ is small, each outcome only weakly biases the state assignment. Many such weak updates can accumulate into a strong measurement record.

Weak measurements are not merely projective measurements with bad notation. Their state update depends on the instrument, and their continuous-time limits lead to stochastic master equations and quantum trajectories.

Measurements Through an Environment

Sometimes no laboratory pointer is read directly. Instead, the system leaks information into an environment:

• photons leaving a cavity carry information about an atom or superconducting qubit;
• phonons carry which-path or energy information;
• scattered particles reveal position information;
• an unmonitored bath produces decoherence and dissipation.

If the environmental record is observed, the system follows a conditioned trajectory. If the record is ignored, the system follows a channel or master equation. The same physical coupling can therefore appear as a measurement, a nonselective channel, decoherence, or open-system dynamics depending on which degrees of freedom are retained.

This is why generalized measurement theory sits next to quantum channels and open systems.

More Outcomes Than Dimensions

A projective measurement on a two-dimensional Hilbert space can have at most two nonzero rank-one orthogonal outcomes. But a qubit POVM can have three, four, or more outcomes.

The trine POVM is a standard example: three effects symmetrically placed in the equatorial plane of the Bloch sphere. Such POVMs are useful in state discrimination and tomography because they ask questions that are not equivalent to choosing one orthonormal basis.

Generalized measurements therefore expand not only the realism of detector modeling, but also the design space for information extraction.

State Discrimination

Nonorthogonal quantum states cannot be perfectly distinguished with no possibility of error. Generalized measurements let one choose tradeoffs that projective measurements on the original system cannot realize as simply.

For example, unambiguous state discrimination introduces an inconclusive outcome:

$$\{F_1,F_2,F_?\}, \qquad F_1+F_2+F_?=I.$$

The conclusive outcomes can be arranged to avoid wrong identifications, but the price is that $F_?$ occurs with nonzero probability. This is a genuinely quantum measurement-design tradeoff, not just detector imperfection.

Destructive and Output-Changing Measurements

Projective measurement formulas often assume the output system lives in the same Hilbert space as the input. Some detectors destroy, absorb, reset, or replace the system.

Photon absorption is the simplest warning. A detector click can be modeled by an operation proportional to

$$a\rho a^\dagger,$$

which lowers photon number in the remaining field. The click event is a useful measurement record, but it is not a nondestructive projection that leaves the original photon available for immediate remeasurement.

More generally, a measurement operation may map input states on $\mathcal H_{\mathrm{in}}$ to output states on $\mathcal H_{\mathrm{out}}$. The instrument language allows this; the basic projective postulate does not cover it cleanly.

The Operational Hierarchy

The clean way to choose the right object is to ask what prediction is needed.

Prediction Needed	Use
sharp ideal outcomes and Lüders update	PVM plus projective update
probabilities for noisy, weak, indirect, or coarse outcomes	POVM
one explicit operator representation of an operation	Kraus operators
probabilities and conditional states for each outcome	quantum instrument
state after the measurement record is ignored	nonselective channel
state conditioned on a time-continuous record	stochastic trajectory

The hierarchy is not a ladder where higher objects make lower ones wrong. It is a bookkeeping discipline: use the least object that answers the physical question, but no less.

What Generalization Does Not Mean

Generalized measurement theory does not mean every measurement is arbitrary. The same constraints that make projective measurement probabilistic still apply in broadened form:

• effects must be positive;
• complete outcome sets must normalize probabilities;
• state transformations must be completely positive on systems entangled with references;
• complete nonselective operations must preserve trace;
• outcome-conditioned states must be normalized only after the outcome probability is accounted for.

These constraints are why POVMs, Kraus maps, and instruments form a stable formalism rather than a list of detector-specific hacks.

Common Mistakes

Treating POVMs as optional decoration

POVMs are not only for exotic protocols. Inefficiency, noise, finite resolution, indirect measurement, and coarse graining already require them.

Forgetting the state update

A POVM effect gives a probability. It does not say what state is left for later predictions.

Assuming every generalized measurement is weak

Some generalized measurements are weak, but others are strong, destructive, informationally complete, or optimized for state discrimination.

Treating Kraus operators as unique

Kraus representations are not unique. The physical content is the operation or instrument, plus whatever apparatus model justifies a particular representation.

Dropping no-click outcomes

If a detector can fail, the no-click or inconclusive branch is part of the complete measurement. Omitting it usually produces missing probability.

Calling every environment interaction a measured outcome

An environment may carry information even when no record is observed. The unobserved case is a channel or master equation, not a selective measurement record.

Exercises

Inefficient Click Is Not a Projector

For

$$F_{\mathrm{click}} = \eta\lvert1\rangle\langle1\rvert,$$

show that $F_{\mathrm{click}}$ is an effect for $0\le\eta\le1$ but not a projector unless $\eta=0$ or $\eta=1$.

Solution

For every vector $\lvert\psi\rangle=\alpha\lvert0\rangle+\beta\lvert1\rangle$,

$$\langle\psi\rvert F_{\mathrm{click}}\lvert\psi\rangle = \eta|\beta|^2\ge0,$$

so the operator is positive when $\eta\ge0$. Also $F_{\mathrm{click}}\le I$ when $\eta\le1$, so it can be part of a complete two-outcome POVM.

It is a projector only if $F_{\mathrm{click}}^2=F_{\mathrm{click}}$. But

$$F_{\mathrm{click}}^2 = \eta^2\lvert1\rangle\langle1\rvert.$$

Thus $F_{\mathrm{click}}^2=F_{\mathrm{click}}$ requires $\eta^2=\eta$, so $\eta=0$ or $\eta=1$.

Same POVM, Different Backaction

Let $F_0=F_1=I/2$. Give two different instruments with this same POVM.

Solution

One instrument flips a fair classical coin and leaves the system unchanged:

$$\mathcal I_0(\rho)=\frac{1}{2}\rho, \qquad \mathcal I_1(\rho)=\frac{1}{2}\rho.$$

Another flips the same fair label but also fully dephases the qubit in the $\sigma_z$ basis:

$$\mathcal J_0(\rho) = \frac{1}{2} \sum_{k=0}^1 \lvert k\rangle\langle k\rvert \rho \lvert k\rangle\langle k\rvert,$$

and

$$\mathcal J_1(\rho) = \frac{1}{2} \sum_{k=0}^1 \lvert k\rangle\langle k\rvert \rho \lvert k\rangle\langle k\rvert.$$

Both instruments have effects $I/2$ for outcomes $0$ and $1$, but the second destroys $\sigma_z$-basis coherence while the first does not.

Choosing the Object

For each task, name the minimal object needed: outcome probabilities only; conditional post-measurement states; state after ignoring the record; continuous conditioned record.

Solution

Outcome probabilities only require a POVM. Conditional post-measurement states require an instrument, or at least a specified outcome operation. The state after ignoring the record is the nonselective channel $\sum_m\mathcal I_m$. A continuous conditioned record requires a continuous-time instrument limit, usually written as a stochastic master equation or quantum trajectory.

Cross-Links

• Measurement as an Operation
• Projective Measurements
• State Update Rules
• Measurement Backaction
• POVMs
• Kraus Operators
• Quantum Instruments
• Naimark Dilation
• Stinespring Dilation
• Unsharp Measurements
• Weak Measurements
• Compatible, Incompatible, and Sequential Measurements
• Completely Positive Maps

References

• E. B. Davies and J. T. Lewis, “An operational approach to quantum probability,” Communications in Mathematical Physics 17, 239–260 (1970).
• C. W. Helstrom, Quantum Detection and Estimation Theory, Academic Press (1976).
• K. Kraus, States, Effects, and Operations: Fundamental Notions of Quantum Theory, Springer (1983).
• A. S. Holevo, Probabilistic and Statistical Aspects of Quantum Theory, North-Holland (1982); Edizioni della Normale (2011).
• P. Busch, P. Lahti, and P. Mittelstaedt, The Quantum Theory of Measurement, 2nd ed., Springer (1996).
• P. Busch, P. Lahti, J.-P. Pellonpää, and K. Ylinen, Quantum Measurement, Springer (2016).
• M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, Cambridge University Press (2010).