Observables and Hermitian Operators

Everything a quantum state can tell you is locked inside a vector, yet an experiment returns a number — a click, a voltage, a spin projection of $+\tfrac{1}{2}$ . The bridge between the vector and the number is an observable, and quantum mechanics is astonishingly restrictive about what one is allowed to be: every measurable quantity — energy, momentum, a spin component, the parity of a code block — is a Hermitian (self-adjoint) operator $A=A^{\dagger}$ on the system's Hilbert space $\mathcal{H}$ . That single condition is not decoration: it is exactly what forces measurement outcomes to be real, forces distinct outcomes to be perfectly distinguishable, and forces their probabilities to sum to one. The engine that turns the abstract condition into physics is the spectral theorem: any Hermitian operator on a finite-dimensional space decomposes as $A=\sum_a a\,P_a$ , where the $a$ are its real eigenvalues and the $P_a$ are mutually orthogonal projectors summing to the identity. The eigenvalues are the only numbers a measurement can produce; the eigenvectors form the measurement basis. This one decomposition is the seed from which the measurement postulate, the Born rule, projective measurement, expectation values, and the Pauli toolkit all grow — every downstream article is, in some sense, a reading of $A=\sum_a a\,P_a$ . The same object wears a second face: exponentiated as $U=e^{-iAt}$ it becomes a unitary, the generator of time evolution — which is why energy is at once the quantity you measure and the clock that moves the state. On a qubit the theory becomes visual: $\{I,X,Y,Z\}$ span every $2\times2$ observable, and the Bloch sphere turns "measuring an observable" into "asking which axis the state points along."

Key takeaways

A physical observable is a Hermitian operator $A=A^{\dagger}$ on the Hilbert space $\mathcal{H}$ ; the condition $A=A^{\dagger}$ is what makes it measurable, not a technicality.
Hermiticity forces real eigenvalues (the only allowed outcomes) and an orthonormal eigenbasis (the measurement basis) — this is the spectral theorem, and both facts drop straight out of $A=A^{\dagger}$ .
The spectral decomposition $A=\sum_a a\,P_a$ with $P_aP_b=\delta_{ab}P_a$ and $\sum_a P_a=I$ packages the outcomes and their projectors into one object that drives every measurement rule downstream.
The Born rule reads probabilities off the projectors, $\Pr(a)=\langle\psi|P_a|\psi\rangle$ , and the mean is $\langle A\rangle=\langle\psi|A|\psi\rangle=\sum_a a\,\Pr(a)$ ; a state is sharp for $A$ exactly when it is an eigenvector, where $\mathrm{Var}(A)=0$ .
Hermitian operators are precisely the generators of unitary evolution through $U=e^{-iAt}$ , tying observables to dynamics and symmetry.
On a qubit, $\{I,X,Y,Z\}$ span all $2\times2$ Hermitian matrices, so every observable is $A=c_0I+\vec c\cdot\vec\sigma$ with eigenvalues $c_0\pm|\vec c|$ — the Bloch-sphere picture of measurement.
Key idea: an observable is a Hermitian operator $A=A^{\dagger}$ ; its spectral decomposition $A=\sum_a a\,P_a$ supplies the real outcomes $a$ , the projectors $P_a$ onto the measurement basis, and — with the Born rule $\Pr(a)=\langle\psi|P_a|\psi\rangle$ — everything about an ideal measurement of that quantity.

The core insight

An observable is a Hermitian operator $A=A^{\dagger}$ ; its spectral decomposition $A=\sum_a a\,P_a$ supplies the real outcomes $a$ , the projectors $P_a$ onto the measurement basis, and — together with the Born rule $\Pr(a)=\langle\psi|P_a|\psi\rangle$ — everything about an ideal measurement of that quantity.

Where this lives today

The observable-as-Hermitian-operator postulate is the daily arithmetic of quantum technology. In superconducting and trapped-ion processors, a computational-basis readout is literally a measurement of $Z$ on each qubit, and rotating the basis before readout implements $X$ , $Y$ , or a general $\hat n\cdot\vec\sigma$ — the primitive at the end of every circuit. In quantum error correction, the surface code repeatedly measures commuting Hermitian stabilizers (products of Paulis): because they commute they share an eigenbasis and can be measured jointly without disturbing the encoded logical information, and their $\pm1$ eigenvalues form the syndrome a decoder turns into a correction. In the variational workloads that dominate today's hardware — VQE for molecular ground states, QAOA for optimization — the objective is an expectation value $\langle\psi(\vec\theta)|H|\psi(\vec\theta)\rangle$ , estimated by decomposing $H=\sum_k c_k P_k$ into Pauli terms, sampling each $\langle P_k\rangle$ , and recombining with the real coefficients $c_k$ . And in metrology, the variance identity of Derivation B is the resource being fought over: squeezed and entangled probes push $\mathrm{Var}(A)$ below the standard quantum limit to sharpen atomic clocks and gravitational-wave detectors. Every one of these is equation (2) plus the Born rule, run at scale.

The physical picture

Picture an observable not as a matrix but as a labeled partition of Hilbert space. A Hermitian operator carves $\mathcal{H}$ into orthogonal eigenspaces — the distinguishable answers — and stamps a real eigenvalue onto each; a measurement asks "which piece is the state in?", and the readout is that piece's label. Because the pieces are orthogonal the alternatives never bleed together; because the labels are real the dial shows honest numbers; because the projectors are complete ( $\sum_a P_a=I$ ) the possibilities are exhaustive and their probabilities sum to one. The first figure makes this concrete: an observable with outcomes $\pm1$ splits the plane into two orthogonal rays, and a tilted state projects onto them with Born weights that fix the expectation value.

Figure

An observable with eigenvalues plus and minus one splits Hilbert space into two orthogonal eigenspaces; a state tilted 30° in the plane projects onto them with Born probabilities 0.75 and 0.25, so the expectation value is ⟨A⟩ = 0.50

The second picture turns a qubit observable into geometry. Every single-qubit observable is a real combination of $I,X,Y,Z$ , and each Pauli's eigenvectors define one of the three orthogonal Bloch axes: $Z$ the poles $|0\rangle,|1\rangle$ , $X$ the pair $|+\rangle,|-\rangle$ , $Y$ the pair $|{+}i\rangle,|{-}i\rangle$ . Measuring $\hat n\cdot\vec\sigma$ asks "does the Bloch vector point up or down along $\hat n$ ?", and the expectation value is the projection of the state's Bloch vector onto that axis. The figure shows a state at $\theta=60^\circ$ whose $Z$ -expectation is just $\cos\theta=0.50$ .

Figure

The three Pauli observables as the orthogonal Bloch axes: Z has eigenstates at the poles, X and Y at the equator, each with eigenvalues plus and minus one; a state at polar angle 60° has expectation value ⟨Z⟩ = cos 60° = 0.50

The core equations

The starting point is the defining condition. An operator is Hermitian — self-adjoint — when it equals its own adjoint, which in inner-product language means it moves freely from one side of the bracket to the other. This is the algebraic content of "observable."

\langle\phi|A\psi\rangle=\langle A\phi|\psi\rangle\quad\Longleftrightarrow\quad A=A^{\dagger} \qquad (1)

where: $A^{\dagger}$ — the adjoint (conjugate transpose in matrix form), defined by $\langle\phi|A\psi\rangle=\langle A^{\dagger}\phi|\psi\rangle$ ; $|\psi\rangle,|\phi\rangle$ — arbitrary states in $\mathcal{H}$ ; $\langle\cdot|\cdot\rangle$ — the inner product. Reading: $A$ acts the same way to the left as to the right, and that symmetry is what forces real eigenvalues below. The spectral theorem repackages a Hermitian operator into a menu of outcomes and the projectors that pick them out.

A=\sum_a a\,P_a,\qquad P_aP_b=\delta_{ab}P_a,\qquad \sum_a P_a=I \qquad (2)

where: $a$ — a real eigenvalue, i.e. a possible outcome; $P_a$ — the orthogonal projector onto the eigenspace of eigenvalue $a$ , so $P_a=P_a^{\dagger}=P_a^2$ ; $\delta_{ab}$ — the Kronecker delta enforcing $P_aP_b=0$ for $a\neq b$ ; $I$ — the identity, with $\sum_a P_a=I$ expressing completeness. Reading: the outcomes $a$ are the labels, the $P_a$ are the "which piece" filters, and completeness guarantees the state lands somewhere. Given the projectors, the Born rule assigns each outcome a probability equal to the squared length of the projected state.

\Pr(a)=\langle\psi|P_a|\psi\rangle=\lVert P_a|\psi\rangle\rVert^2 \qquad (3)

where: $\Pr(a)$ — probability of outcome $a$ ; $|\psi\rangle$ — the normalized pre-measurement state, $\langle\psi|\psi\rangle=1$ ; $\lVert\cdot\rVert$ — the vector norm. Reading: because $P_a$ is a projector, $\langle\psi|P_a|\psi\rangle=\lVert P_a|\psi\rangle\rVert^2\ge0$ , and $\sum_a\Pr(a)=\langle\psi|I|\psi\rangle=1$ automatically. Averaging the eigenvalues against their Born weights gives the expectation value, computable directly from the operator without finding the eigenvectors.

\langle A\rangle=\langle\psi|A|\psi\rangle=\sum_a a\,\Pr(a) \qquad (4)

where: $\langle A\rangle$ — the mean outcome over many measurements on identically prepared copies; the two expressions agree by inserting $A=\sum_a a\,P_a$ . Reading: the sandwich $\langle\psi|A|\psi\rangle$ returns a real number precisely because $A$ is Hermitian — the workhorse quantity of variational quantum algorithms. A measurement also updates the state: conditioned on outcome $a$ , it jumps to its normalized projection into that eigenspace — the projection postulate, or "collapse."

|\psi\rangle\;\longmapsto\;\frac{P_a|\psi\rangle}{\lVert P_a|\psi\rangle\rVert} \qquad (5)

where: $P_a|\psi\rangle$ — the unnormalized component inside eigenspace $a$ ; $\lVert P_a|\psi\rangle\rVert=\sqrt{\Pr(a)}$ — the norm restoring unit length. Reading: repeating the measurement returns $a$ with certainty, because the post-measurement state is already an eigenstate of $A$ . The same operator that reads out numbers also generates motion: exponentiating it gives a unitary whose eigenphases are the eigenvalues, so observables and dynamical generators are one class of object.

U=e^{-iAt}=\sum_a e^{-iat}P_a,\qquad U^{\dagger}U=I \qquad (6)

where: $U$ — the unitary generated by $A$ (with $A=H/\hbar$ this is Schrödinger evolution); $t$ — a real parameter (time); $e^{-iat}$ — a unit-modulus phase per eigenvalue. Reading: real eigenvalues $a$ become pure phases of modulus one, and unit-modulus eigenvalues are exactly what makes $U$ unitary — the Hermitian/unitary bridge in one line. Finally, on a qubit the four matrices $\{I,X,Y,Z\}$ exhaust the Hermitian $2\times2$ world, so any observable is a real four-parameter combination whose spectrum reads off the Bloch vector $\vec c$ .

A=c_0 I+\vec c\cdot\vec\sigma,\qquad \vec\sigma=(X,Y,Z),\qquad \text{eigenvalues } c_0\pm|\vec c| \qquad (7)

where: $c_0\in\mathbb{R}$ — the identity coefficient, fixing $\mathrm{Tr}\,A=2c_0$ ; $\vec c\in\mathbb{R}^3$ — the Pauli coefficient vector, whose direction $\hat n=\vec c/|\vec c|$ is the measurement axis; $\vec\sigma$ — the Pauli vector. Reading: a qubit observable is a point in $\mathbb{R}^4$ , its two eigenvalues are $c_0\pm|\vec c|$ , and its eigenstates are the $\pm\hat n$ poles on the Bloch sphere.

Derivation AHermitian operators have real eigenvalues and orthogonal eigenvectors

The goal is to show that the two properties an observable must have — real outcomes, and perfectly distinguishable (orthogonal) states for distinct outcomes — are not extra assumptions but forced consequences of the single condition $A=A^{\dagger}$ . We take an eigenvalue equation, pair it against a bra through the inner product, and let Hermiticity do all the work; nothing beyond the definition of the adjoint is used. This is the load-bearing lemma under the entire spectral theorem.

Step 1. Fix two eigenpairs of $A$ with unit-norm eigenvectors, $A|a\rangle=a|a\rangle$ and $A|b\rangle=b|b\rangle$ , allowing the possibility $a=b$ . Why: reality is a statement about one eigenpair, orthogonality a statement about two; setting up both at once lets a single algebraic identity deliver both.

Step 2. For reality, sandwich $A$ in its own eigenstate: $a\langle a|a\rangle=\langle a|A|a\rangle$ . Because $A=A^{\dagger}$ , this scalar equals its conjugate, $\langle a|A|a\rangle=\overline{\langle a|A^{\dagger}|a\rangle}=\bar a\langle a|a\rangle$ . Since $\langle a|a\rangle>0$ , $a=\bar a$ , so $a\in\mathbb{R}$ . Why: a self-adjoint sandwich is always real, and an eigenvalue is such a sandwich over a positive norm — reality is immediate.

Step 3. For orthogonality, evaluate $\langle a|A|b\rangle$ two ways. To the right it gives $b\langle a|b\rangle$ ; to the left, using $A=A^{\dagger}$ and the reality just proven, $\langle Aa|b\rangle=\bar a\langle a|b\rangle=a\langle a|b\rangle$ . Why: Hermiticity is the license to slide $A$ across the inner product, and the two evaluations of the same bracket must agree.

Step 4. Subtract the two evaluations: $(b-a)\langle a|b\rangle=0$ . If $a\neq b$ then necessarily $\langle a|b\rangle=0$ , so eigenvectors for distinct eigenvalues are orthogonal; within a degenerate eigenspace (where $a=b$ ) an orthonormal basis can always be built by Gram–Schmidt. Together these give the orthonormal eigenbasis that equation (2) rests on. Why: distinct outcomes must be perfectly distinguishable for probabilities to make sense, and orthogonality is exactly perfect distinguishability.

Go deeper · Derivation B · 1 minSharpness, variance, and what a definite value means

Derivation BSharpness, variance, and what a definite value means

The aim is to make precise the intuition that a system "has a definite value" of an observable only in special states, and to identify those states exactly. We define the variance of $A$ as a genuine spread of outcomes, rewrite it as the squared length of a vector so non-negativity is manifest, then ask when that length can vanish. The payoff is a clean iff-statement — zero uncertainty exactly when the state is an eigenstate — the quantitative backbone of the uncertainty principle and of a "sharp" measurement.

Step 1. Define the variance by centering the observable on its own mean, $\mathrm{Var}(A)=\big\langle(A-\langle A\rangle I)^2\big\rangle$ , where $\langle A\rangle=\langle\psi|A|\psi\rangle$ is a real number by Derivation A. Why: variance should measure spread about the mean, so subtracting $\langle A\rangle I$ first is the honest definition; expanding it recovers the familiar $\langle A^2\rangle-\langle A\rangle^2$ .

Step 2. Because $A-\langle A\rangle I$ is Hermitian, the expectation of its square folds into a norm: $\mathrm{Var}(A)=\lVert(A-\langle A\rangle I)|\psi\rangle\rVert^2\ge0$ . Why: writing the second moment as $\lVert M|\psi\rangle\rVert^2$ makes non-negativity automatic — a variance can never come out negative, as a physical spread should not.

Step 3. A norm vanishes only for the zero vector, so $\mathrm{Var}(A)=0$ holds if and only if $(A-\langle A\rangle I)|\psi\rangle=0$ , i.e. $A|\psi\rangle=\langle A\rangle|\psi\rangle$ . The state is then an eigenvector, and $\langle A\rangle$ is its eigenvalue. Why: this pins the meaning of "definite value" — a measurement is deterministic exactly on eigenstates, and everywhere else there is irreducible spread.

Step 4. Read off the physics: on an eigenstate, outcome $a$ occurs with probability one and equation (5) leaves the state unchanged; off an eigenstate, $\mathrm{Var}(A)>0$ quantifies genuine indeterminacy no better preparation of that observable can remove. Why: the bridge from algebra to the lab — the variance is what repeated shots on a real device scatter by.

Go deeper · Derivation C · 1 minPauli operators as the complete basis of qubit observables

Derivation CPauli operators as the complete basis of qubit observables

The goal is to prove that on a qubit there is nothing more to know than the four numbers $(c_0,\vec c)$ : every Hermitian $2\times2$ matrix is a unique real combination of $I,X,Y,Z$ , each Pauli is a $\pm1$ -valued observable, and the eigenstates of $X,Y,Z$ are the three orthogonal Bloch axes. We count the real degrees of freedom of a Hermitian matrix, exhibit the bijection to the Pauli set, and read off spectra from two one-line algebraic facts. This is what makes the Bloch sphere the complete measurement theory of a qubit.

Step 1. Write the general Hermitian $2\times2$ matrix. Hermiticity forces a real diagonal and conjugate off-diagonals, $A=\begin{pmatrix}\alpha&\beta\\ \bar\beta&\gamma\end{pmatrix}$ with $\alpha,\gamma\in\mathbb{R}$ , $\beta\in\mathbb{C}$ — four real parameters. Why: the count is the whole point — a four-dimensional real space needs exactly four basis matrices to fill it.

Step 2. Expand in the Pauli set: $A=\tfrac{\alpha+\gamma}{2}I+\mathrm{Re}(\beta)\,X-\mathrm{Im}(\beta)\,Y+\tfrac{\alpha-\gamma}{2}Z$ . The four real coefficients map bijectively onto $\{I,X,Y,Z\}$ , which are linearly independent, so they form a basis of the real vector space of qubit observables — this is equation (7). Why: bijectivity means "specify a qubit observable" and "specify $(c_0,\vec c)\in\mathbb{R}^4$ " are the same act, with no observable left out and none double-counted.

Step 3. Read the spectra from $X^2=Y^2=Z^2=I$ together with $\mathrm{Tr}\,X=\mathrm{Tr}\,Y=\mathrm{Tr}\,Z=0$ . Squaring to the identity forces every eigenvalue to satisfy $\lambda^2=1$ , so $\lambda\in\{+1,-1\}$ ; tracelessness forces one of each. Hence each Pauli is a genuine $\pm1$ observable. Why: these two facts — involution and tracelessness — are all you need to know a Pauli's spectrum without ever diagonalizing.

Step 4. Identify the measurement bases as the three orthogonal Bloch axes: $Z\to\{|0\rangle,|1\rangle\}$ , $X\to\{|+\rangle,|-\rangle\}$ , $Y\to\{|{+}i\rangle,|{-}i\rangle\}$ . A general observable $c_0I+\vec c\cdot\vec\sigma$ has eigenvalues $c_0\pm|\vec c|$ with eigenstates at the poles $\pm\hat n$ , $\hat n=\vec c/|\vec c|$ . Why: the geometric punchline — measuring any qubit observable is choosing an axis $\hat n$ and asking which pole the state favors, with $\langle A\rangle=c_0+\vec c\cdot\langle\vec\sigma\rangle$ .

Limiting cases

Special structure in $A$ collapses the general machinery to a familiar case; the table is a fast way to internalize what the spectral decomposition is doing.

Case	Result	Meaning
Non-degenerate spectrum	$A=\sum_a a\,\vert a\rangle\langle a\vert$	Every eigenvalue distinct, so each $P_a=\vert a\rangle\langle a\vert$ is rank one and the measurement basis is unique up to phases — the cleanest complete measurement.
Degenerate eigenvalue	$P_a=\sum_{k}\vert a,k\rangle\langle a,k\vert$	Several eigenvectors share an eigenvalue, so $P_a$ has rank $>1$ ; the measurement projects onto the subspace but cannot resolve within it.
Projector as observable	$A=P,\ \ P^2=P=P^{\dagger}$	Eigenvalues are just $0$ and $1$ — a yes/no question — and $\langle\psi\vert P\vert\psi\rangle$ is directly the probability of "yes."
Eigenstate (sharp) input	$A\vert a\rangle=a\vert a\rangle\Rightarrow\mathrm{Var}(A)=0$	The state already lives in one eigenspace, so outcome $a$ is certain and the state is undisturbed by measuring $A$ .
$A^2=I$ (involution)	eigenvalues $\pm1$ , and $A$ is both Hermitian and unitary	The Pauli operators and reflections; here observable and dynamical generator coincide in a single matrix.

Common trapHermitian and unitary are different jobs that happen to overlap for the Paulis

The most common early error blurs the two great matrix classes. Hermitian operators ( $A=A^{\dagger}$ , real eigenvalues) are observables — the things you measure; unitary operators ( $U^{\dagger}U=I$ , unit-modulus eigenvalues) are gates — the things you do. A generic Hamiltonian is Hermitian but not unitary; a generic gate is unitary but not Hermitian. They overlap only when $A^2=I$ , which is why the Paulis $X,Y,Z$ are simultaneously observables and gates — a coincidence that tempts beginners to treat the classes as interchangeable. A subtler trap: the eigenvalues, not the eigenvectors, the name, or the matrix entries, are the outcomes. Matrix entries are basis-dependent and can be complex; eigenvalues are basis-invariant and, by Derivation A, guaranteed real.

Worked examples

1 · Measuring a tilted qubit (fully worked, with numbers)

A qubit is prepared in $|\psi\rangle=\cos(\theta/2)|0\rangle+\sin(\theta/2)|1\rangle$ at Bloch polar angle $\theta=60^\circ$ , so $|\psi\rangle=0.8660\,|0\rangle+0.5000\,|1\rangle$ . Measure $Z$ (eigenvalues $\pm1$ , projectors $P_{+1}=|0\rangle\langle0|$ , $P_{-1}=|1\rangle\langle1|$ ). Find the probabilities, expectation value, variance and standard deviation, and post-measurement state.

Step 1. Probabilities by the Born rule (equation 3), with rank-one projectors: $\Pr(+1)=|\langle0|\psi\rangle|^2=(0.8660)^2=\mathbf{0.75}$ and $\Pr(-1)=|\langle1|\psi\rangle|^2=(0.5000)^2=\mathbf{0.25}$ , summing to one as completeness demands.

Step 2. Expectation value (equation 4): $\langle Z\rangle=(+1)(0.75)+(-1)(0.25)=\mathbf{0.50}$ . This is exactly $\cos\theta=\cos60^\circ$ , the projection of the Bloch vector onto the $z$ -axis — the geometric reading of the second figure.

Step 3. Spread. Since $Z^2=I$ , $\langle Z^2\rangle=1$ , so $\mathrm{Var}(Z)=1-(0.50)^2=\mathbf{0.75}$ and $\Delta Z=\sqrt{0.75}=\mathbf{0.8660}$ . Note $\Delta Z=|\sin\theta|=\sin60^\circ$ : uncertainty is largest on the equator, zero at the poles, matching Derivation B.

Step 4. Collapse (equation 5): outcome $+1$ sends $|\psi\rangle\mapsto|0\rangle$ , outcome $-1$ sends it to $|1\rangle$ , and a repeat $Z$ -measurement returns the same value with certainty. Summary: Pr = (0.75, 0.25), ⟨Z⟩ = 0.50, Var = 0.75, ΔZ = 0.866, collapse to |0⟩ or |1⟩.

2 · A general observable and its spectral decomposition

Consider $A=2I+X+Z=\begin{pmatrix}3&1\\1&1\end{pmatrix}$ . Find its eigenvalues via the Bloch form, then $\langle A\rangle$ , the probabilities, and $\mathrm{Var}(A)$ for the input state $|0\rangle$ .

Step 1. Read the spectrum from equation (7): $c_0=2$ , $\vec c=(1,0,1)$ , $|\vec c|=\sqrt2$ , so the eigenvalues are $c_0\pm|\vec c|=2\pm\sqrt2$ , i.e. $a_+=\mathbf{3.414}$ , $a_-=\mathbf{0.586}$ . The measurement axis $\hat n=(1,0,1)/\sqrt2$ is tilted $45^\circ$ from $z$ in the $x$ – $z$ plane.

Step 2. Expectation the fast way, $\langle A\rangle=c_0+\vec c\cdot\langle\vec\sigma\rangle$ . On $|0\rangle$ , $\langle\vec\sigma\rangle=(0,0,1)$ , so $\langle A\rangle=2+1=\mathbf{3.00}$ — equivalently the top-left entry $\langle0|A|0\rangle=3$ .

Step 3. Probabilities. The eigenstate $|a_+\rangle$ points along $\hat n$ (polar angle $45^\circ$ ), so $\Pr(a_+)=|\langle a_+|0\rangle|^2=\cos^2(22.5^\circ)=\mathbf{0.8536}$ and $\Pr(a_-)=\sin^2(22.5^\circ)=\mathbf{0.1464}$ . Check: $3.414(0.8536)+0.586(0.1464)=\mathbf{3.00}$ , matching Step 2.

Step 4. Variance. $\langle A^2\rangle=a_+^2\Pr(a_+)+a_-^2\Pr(a_-)=10.00$ , so $\mathrm{Var}(A)=10.00-3.00^2=\mathbf{1.00}$ , which also equals $|\vec c|^2(1-(\hat n\cdot\hat z)^2)=2(1-\tfrac12)=1$ . Summary: eigenvalues 2±√2, ⟨A⟩ = 3.00, Pr = (0.854, 0.146), Var = 1.00.

Physics of the portrait

The idea that a physical quantity is a self-adjoint operator, and its spectrum is the set of measurable values, was the achievement of one of the century's deepest mathematicians, who gave quantum theory its rigorous Hilbert-space skeleton.

Origins · 1 minJohn von Neumann and how this was discovered

John von Neumann

1903–1957 · the mathematical foundations of quantum mechanics

When von Neumann arrived at Göttingen in the late 1920s, quantum mechanics existed in two rival dialects — Heisenberg's matrices and Schrödinger's waves — held together by Dirac's brilliant but mathematically informal transformation theory, which leaned on "improper" objects like the delta function that no existing analysis could justify. Von Neumann's response, culminating in his 1932 Mathematische Grundlagen der Quantenmechanik, was to rebuild the theory on the abstract Hilbert space and the spectral theory of self-adjoint operators, which he extended to the unbounded case. His decisive move was to identify observables with self-adjoint operators and to prove that the spectral theorem supplies, for each, a projection-valued measure — exactly the projectors $P_a$ of equation (2) — whose statistics reproduce measured values. From this vantage he derived $\langle A\rangle=\mathrm{Tr}(\rho A)$ , introduced the density operator for statistical mixtures, and formulated the projection postulate. He built on Hermite's 1850s observation that symmetric matrices have real eigenvalues, on Hilbert's spectral theory, and on Born's 1926 probabilistic reading of the wavefunction, welding them into the axioms every quantum-information course now opens with. That an observable is a Hermitian operator and its outcomes are eigenvalues is, in a real sense, von Neumann's sentence.

History

1855

— Hermite shows that symmetric (self-adjoint) matrices have real eigenvalues

1926

— Born gives the probabilistic interpretation of the wavefunction

1927

— Heisenberg states the uncertainty principle for conjugate observables

1930

— Dirac's Principles of Quantum Mechanics systematizes bra-ket transformation theory

1932

— von Neumann founds QM on self-adjoint operators and the spectral theorem

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.1.5, §2.2.3–2.2.5 · Preskill, Lecture Notes on Quantum Computation, Ch. 2–3 · Watrous, The Theory of Quantum Information, §1.1 · Wilde, Quantum Information Theory, Ch. 3 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §8 · Mermin, Quantum Computer Science, Ch. 1 · von Neumann, Mathematical Foundations of Quantum Mechanics, Ch. III–IV. All equations are standard. Key identities to keep in mind: reality and orthogonality follow from $A=A^{\dagger}$ ; the spectral decomposition satisfies $\sum_aP_a=I$ ; the variance obeys $\mathrm{Var}(A)=\langle A^2\rangle-\langle A\rangle^2\ge0$ with equality if and only if the state is an eigenstate; and the Paulis satisfy $X^2=Y^2=Z^2=I$ , are traceless, have eigenvalues $\pm1$ , and together with $I$ form a four-matrix basis for all $2\times2$ Hermitian matrices. The worked examples yield: Example 1, $\Pr=(0.75,0.25)$ , $\langle Z\rangle=0.50$ , $\mathrm{Var}=0.75$ ; Example 2, eigenvalues $2\pm\sqrt2$ , $\langle A\rangle=3.00$ , $\Pr=(0.8536,0.1464)$ , $\mathrm{Var}=1.00$ .

Prerequisites. Inner Product, Norm, and Normalization · Dirac Bra-Ket Notation · The Qubit and Its State Space. Leads to. The Measurement Postulate · Born Rule · Projective Measurements · Expectation Values and Variance · Pauli Operators.

Ask the tutor: Want to see the spectral decomposition of any observable you name, worked out with its projectors, Born probabilities, and Bloch-sphere picture — just give me the operator and a state.

Why does a physical observable have to be Hermitian in the first place?

Because the three things a measurement must do map exactly onto the three things $A=A^{\dagger}$ guarantees. Outcomes are dial readings, so they must be real — Hermiticity forces real eigenvalues. Distinct outcomes must be perfectly distinguishable, so their states must be orthogonal — Hermiticity forces an orthonormal eigenbasis. The outcome list must be exhaustive, so the projectors must sum to the identity — the spectral theorem gives $\sum_aP_a=I$ . Drop self-adjointness and at least one fails: a non-normal operator can have complex "outcomes" and non-orthogonal eigenbases. So Hermiticity is not a modelling convenience but the minimal condition under which "measurement" is coherent.

Is every Hermitian operator an observable, and is every observable unitary?

Every Hermitian operator is a legitimate observable in principle — though whether a given one is easy to measure in the lab is a separate engineering question. The converse pairing is the real trap: observables and unitaries are different classes. A unitary can have complex eigenvalues $e^{i\theta}$ , which are not valid real outcomes, so a generic gate is not an observable. A generic Hermitian operator — say a Hamiltonian with eigenvalues other than $\pm1$ — is not unitary. The two classes intersect precisely when $A^2=I$ , which forces eigenvalues $\pm1$ ; the Pauli operators and reflections live in that intersection and are simultaneously observables and gates. Everywhere else, "the thing you measure" and "the thing you do" are distinct objects.

How does the expectation value work for a mixed state, and why is it still real?

Replace the pure-state sandwich by a trace against the density operator: $\langle A\rangle=\mathrm{Tr}(\rho A)$ , with $\Pr(a)=\mathrm{Tr}(\rho P_a)$ . This reduces to $\langle\psi|A|\psi\rangle$ when $\rho=|\psi\rangle\langle\psi|$ and correctly averages over an ensemble when $\rho$ is mixed. Reality survives because both $\rho$ and $A$ are Hermitian: $\mathrm{Tr}(\rho A)^{*}=\mathrm{Tr}(A^{\dagger}\rho^{\dagger})=\mathrm{Tr}(A\rho)=\mathrm{Tr}(\rho A)$ , so the value equals its own conjugate. This is the trace form von Neumann built the statistical theory on, and it is what noisy near-term hardware actually estimates.

Why is one and the same operator both the energy observable and the generator of time evolution?

Because exponentiating a Hermitian operator produces a unitary whose eigenphases are its eigenvalues: $U=e^{-iHt/\hbar}=\sum_a e^{-ia t/\hbar}P_a$ . The Hamiltonian $H=\sum_a a\,P_a$ is a yardstick — its eigenvalues are the measurable energy levels, its eigenvectors the stationary states. The same $H$ read as $e^{-iHt/\hbar}$ is a clock that moves the state forward in time. The real eigenvalues that make $H$ a valid observable are exactly the unit-modulus phases that make $U$ unitary, so the two roles are two readings of one spectral decomposition — the deepest reason energy conservation and time-translation symmetry are the same statement (Noether's theorem, quantum form).

What changes when an eigenvalue is degenerate?

Degeneracy means several orthogonal states share one eigenvalue $a$ , so $P_a$ has rank greater than one and projects onto a whole subspace rather than a single ray. Three things follow. First, the measurement tells you the state landed in that subspace but cannot resolve where inside it — equation (5) returns the normalized projection, which may still be a superposition. Second, the eigenbasis is no longer unique: any orthonormal basis of the subspace works, and you fix one by measuring an additional commuting observable that breaks the degeneracy (a complete set of commuting observables). Third, the Born probability $\Pr(a)=\langle\psi|P_a|\psi\rangle$ still holds verbatim — written with projectors precisely so degeneracy needs no special treatment.

Can two observables be measured together, and how do I tell?

Two observables can be measured simultaneously — assigned joint definite values — exactly when they commute, $[A,B]=AB-BA=0$ . Commuting Hermitian operators share an eigenbasis, so one measurement in that basis reads out both; this is why stabilizer codes measure many commuting Paulis in parallel. When $[A,B]\neq0$ there is no shared eigenbasis, measuring one disturbs the other, and their spreads obey $\mathrm{Var}(A)\,\mathrm{Var}(B)\ge\tfrac14|\langle[A,B]\rangle|^2$ — the Robertson relation that specializes to Heisenberg's $\Delta x\,\Delta p\ge\hbar/2$ . The commutator is the single diagnostic: zero means jointly measurable, non-zero means complementary.