Hermitian Operators and Observables

Measurements return real numbers. A detector reads 3.7 eV, never $3.7 + 0.2i$ eV. That single laboratory banality, pushed through the formalism you have built over the last four articles, selects one family of operators to represent everything measurable — and this article proves why. The chosen family is the Hermitian operators: those equal to their own conjugate-mirror, $\langle f|\hat A g\rangle = \langle\hat A f|g\rangle$ for all states. Three theorems, each proved here in a few honest lines, deliver exactly the package a theory of measurement needs: Hermitian operators have real eigenvalues (the possible readings), orthogonal eigenstates for distinct readings (outcomes are perfectly distinguishable, as the inner-product article said they must be), and real expectation values in every state (averages of real readings are real). The proofs are small; the checking is where the craft lives. Is the momentum operator Hermitian? Yes — because of the $i$ : integration by parts shows bare $d/dx$ is anti-Hermitian, flipping sign under the mirror, and the factor $-i\hbar$ flips it back. Is $\hat x\hat p$ Hermitian? No — and the failure is computable, exactly one commutator's worth — while the symmetrized combination $\tfrac12(\hat x\hat p + \hat p\hat x)$ passes. Every quantization recipe, every "write down the operator for this classical quantity" exercise, runs through checks like these, and this article drills them on the operators you own. One honesty clause completes the picture. For the unbounded operators of real physics, "Hermitian" comes with domain fine print — boundary terms in the integration by parts that vanish only for suitable states, a distinction (symmetric versus truly self-adjoint) flagged here with a computable half-line example and settled at the chapter's end. The wave-register story of what observables are was told in the previous chapter's Operators and Observables; this article is where the formalism proves that story had to be true.

Key takeaways

An operator is Hermitian when $\langle f|\hat A g\rangle = \langle\hat A f|g\rangle$ for all states — it can hop across the inner product without change.
Theorem 1: Hermitian eigenvalues are real — the right property for measurement readings.
Theorem 2: eigenstates with distinct eigenvalues are orthogonal — different outcomes are perfectly distinguishable.
Theorem 3: $\langle\hat A\rangle$ is real in every state — and conversely (on complex spaces), all-real expectations force Hermiticity.
Credentials must be checked: $\hat x$ ✓; $\hat p = -i\hbar\,d/dx$ ✓ (the $i$ is load-bearing — bare $d/dx$ is anti-Hermitian); $\hat x\hat p$ ✗ (defect $= i\hbar\langle f|g\rangle$ ); $\tfrac12(\hat x\hat p + \hat p\hat x)$ ✓.
Degenerate eigenvalues: eigenvectors within one eigenspace can always be orthogonalized (Gram–Schmidt) — orthogonality across the spectrum survives.
Unbounded operators carry domain fine print (symmetric vs self-adjoint): boundary terms are physics — flagged here, resolved at Rigged Hilbert Space.
Key idea: Hermiticity is exactly the property that forces real eigenvalues, orthogonal eigenstates, and real averages — precisely the ingredients quantum mechanics needs to assign real, probabilistically consistent values to measurements.

The core insight

Self-adjointness is exactly the property that forces real eigenvalues, an orthogonal eigenbasis, and real expectation values — precisely the ingredients quantum mechanics needs to assign real, probabilistically consistent outcomes to measurements.

Where this lives today

"Is it Hermitian?" is a working checklist item across physics and engineering. Quantum-chemistry and materials codes diagonalize explicitly Hermitian Hamiltonians with algorithms (`eigh`, LAPACK's Hermitian drivers) that assume the symmetry and would silently produce garbage without it — re-Hermitizing after numerical assembly is standard hygiene. Quantum computing measures Hermitian operators by construction: every readout is an eigenvalue of a Pauli string, and variational algorithms estimate $\langle\hat H\rangle$ as real-by-theorem sums of Pauli expectations. The fine print earns its keep too: engineered non-Hermitian systems — gain-loss optics, PT-symmetric lattices — are a thriving industry precisely because they break these theorems in controlled ways (exceptional points, with no Hermitian analogue, now enhance sensors). And the boundary-term physics of equation (5) is daily bread in nanoscience, where self-adjoint extensions of $\hat p^2$ on wires and rings are the menu of physically distinct contacts.

The physical picture

Picture the inner product as a wall between two rooms — bra state one side, ket state the other. An operator in the ket room can try to walk through the wall: $\langle f|\hat A g\rangle \to \langle \hat A^{?} f|g\rangle$ . Most operators arrive altered — the wall conjugates them, in a sense the adjoint article will formalize. The Hermitian operators pass through as themselves. That self-mirror property looks like idle symmetry until you feed it eigenstates and watch it squeeze out reality and orthogonality — the theorems below are literally the operator meeting its own reflection and cancelling everything complex.

Figure

The mirror test — a Hermitian operator crosses the inner product unchanged

Why should measurability care about this mirror? Trace the requirements backwards. A measurable quantity needs possible values (real numbers) and states in which the value is definite (eigenstates); repeats must re-return the value, and states with different definite values must be tellable-apart with certainty — which the inner-product article showed means orthogonal. Reality plus orthogonality: exactly what the mirror property delivers, and (the converse theorem below) only the mirror property delivers. The postulate "observables are Hermitian" — an axiom in The Postulates of Quantum Mechanics later this chapter — is less a decree than reverse engineering: nothing else would work. You have been using the consequences all along. Every energy eigenvalue of the one-dimensional chapter was real because $\hat H$ is Hermitian; the ISW eigenfunctions' orthogonality is Theorem 2 wearing sines; and the wave-register account of the operator↔observable correspondence (Operators and Observables, previous chapter) is the physical story whose formal engine you are now installing.

The definition, and the credentials of the standard operators

The mirror statement, precisely — for all states $|f\rangle, |g\rangle$ in the relevant domain:

\langle f|\hat A g\rangle = \langle\hat A f|g\rangle \qquad \text{(}\hat A\text{ Hermitian)} \qquad (1)

where: in $L^2$ this reads $\int f^*(\hat A g)\,dx = \int(\hat A f)^*g\,dx$ — the operator shifts from ket to bra without residue. (The equivalent packaging $\hat A = \hat A^\dagger$ is deliberately deferred: the dagger's definitional theory belongs to Hermitian Conjugate and Adjoint Operators later this chapter; definition (1) is self-contained — Griffiths' working form.) In matrices, (1) becomes $A_{mn} = A_{nm}^*$ : conjugate-transpose symmetry, diagonal real. Credentials are checked by doing the integral, and the two canonical checks teach opposite lessons. Position passes trivially — $x$ is real, and multiplying under a conjugate is insensitive to which side it formally sits on. Momentum is the instructive one, and it earns a card.

Derivation AMomentum passes the mirror test — and the i is the repair

We verify that $\hat p = -i\hbar\,d/dx$ satisfies definition (1) on normalizable states of the line, and locate exactly why the imaginary unit is structural rather than decorative. The strategy is a single integration by parts, with the bookkeeping of two sign flips watched closely — this is the template for every Hermiticity audit you will ever run, and its boundary term is the seed of the domain fine print later in the article.

Step 1. Set up the left side of (1) and integrate by parts once:

\int_{-\infty}^{\infty} f^*\Bigl(-i\hbar\frac{dg}{dx}\Bigr)dx = -i\hbar\Bigl[f^*g\Bigr]_{-\infty}^{\infty} + i\hbar\int_{-\infty}^{\infty}\frac{df^*}{dx}\,g\,dx \qquad (2)

Step 2. Kill the boundary term: normalizable states have vanishing tails, so $[f^*g]_{-\infty}^{\infty} = 0$ . Why this step is a live wire, not a formality: on a half-line, or with singular potentials, the bracket survives — equation (5) below computes it — and Hermiticity becomes a statement about the domain.

Step 3. Recognize the survivor: $+i\hbar\int (df/dx)^*g\,dx = \int(-i\hbar\,df/dx)^*\,g\,dx = \langle\hat p f|g\rangle$ — the conjugation flipped the sign of $-i$ , exactly cancelling the minus sign that parts produced. Two flips, net zero: $\hat p$ crosses the mirror unchanged.

Step 4. Run the control: delete the $i$ . Bare $d/dx$ keeps only the parts-flip: $\langle f|g'\rangle = -\langle f'|g\rangle$ — anti-Hermitian, sign reversed on crossing. The imaginary unit is the repair that makes momentum measurable; anti-Hermitian operators get their own (generator) career, but not an observable's. Verified: symbolically (sympy) on Gaussian-damped test pairs — $\langle f|\hat p g\rangle = \langle\hat p f|g\rangle$ exactly, $\langle f|g'\rangle + \langle f'|g\rangle = 0$ exactly, and the difference $\langle f|g'\rangle - \langle f'|g\rangle \ne 0$ .

Products require care: Hermitian × Hermitian is not automatically Hermitian. For $\hat x\hat p$ , walking the pair across the mirror reverses their order (the daggering grammar of the bra-ket article), and un-reversing costs the commutator:

\langle f|\hat x\hat p\,g\rangle - \langle\hat x\hat p\,f|g\rangle = i\hbar\,\langle f|g\rangle \ \ne 0, \qquad \text{but}\quad \tfrac12\bigl(\hat x\hat p + \hat p\hat x\bigr) \ \text{is Hermitian} \qquad (3)

where: the defect is exactly one commutator's worth, state-independent — computed symbolically on test functions in this article's verification. The symmetrized combination distributes the failure evenly and cancels it. This is the symmetrization rule of quantization folklore: classical products like $xp$ have no unique quantum image, and the Hermitian requirement is what forces the symmetric choice. (Deeper orderings and their ambiguities are a story for the dynamics chapter; the credential-checking habit is this article's deliverable.)

Go deeper · Derivation B · 2 minThe three theorems — reality, orthogonality, and the converse

Derivation BThe three theorems — reality, orthogonality, and the converse

We prove the package that makes Hermitian operators the observables: (i) eigenvalues are real; (ii) eigenvectors of distinct eigenvalues are orthogonal; (iii) expectation values are real in every state — with the converse of (iii) sketched honestly. The strategy for (i) and (ii) is the same two-line move: evaluate one cleverly chosen inner product twice, once letting $\hat A$ act right and once letting it hop left by definition (1), and subtract. Watch how little is used — no wave equation, no specific operator, only the mirror property.

Step 1. Reality. Let $\hat A|a\rangle = a|a\rangle$ with $|a\rangle$ normalized. Evaluate $\langle a|\hat A a\rangle$ both ways:

\langle a|\hat A a\rangle = a\,\langle a|a\rangle = a, \qquad \langle\hat A a|a\rangle = a^*\langle a|a\rangle = a^*

Definition (1) equates them: $a = a^*$ — real. Why the star appeared on the left route: the eigenvalue exited the bra slot, and the bra side conjugates scalars (the sesquilinearity of the inner-product article, load-bearing at last).

(a - b)\,\langle b|a\rangle = 0 \quad\Longrightarrow\quad \langle b|a\rangle = 0

Why the hypothesis $a \ne b$ is essential — and what happens without it: inside a degenerate eigenspace the argument says nothing, and eigenvectors there are genuinely free to be non-orthogonal. The repair is standard: any basis of the eigenspace can be orthogonalized (Gram–Schmidt, verified numerically below) while staying inside the eigenspace — so an orthonormal eigenbasis exists across the whole spectrum. The systematic basis theory is Complete Orthonormal Bases'; the fact needed here is only that degeneracy is an organizable freedom, not a breakdown.

Step 3. Real expectations. For any state, $\langle\hat A\rangle = \langle\psi|\hat A\psi\rangle = \langle\hat A\psi|\psi\rangle = \langle\psi|\hat A\psi\rangle^*$ — equal to its own conjugate, hence real. (The middle equality is (1); the last is conjugate symmetry.)

Step 4. The converse — reality is not just necessary but characteristic. Suppose $\langle\psi|\hat A\psi\rangle$ is real for every state of a complex Hilbert space. Applying the assumption to $|\psi\rangle = |u\rangle + |v\rangle$ and to $|u\rangle + i|v\rangle$ and combining (the polarization trick) forces $\langle u|\hat A v\rangle = \langle\hat A u|v\rangle$ for all pairs: Hermiticity. So the operator family with all-real measurement statistics is exactly the Hermitian family — the postulate has no slack. (Note the complex field is essential: on a real space, reality of diagonal elements is free and characterizes nothing — one more reason quantum mechanics is complex.) Verified: numerically on random $6\times6$ complex matrices — Hermitian: eigenvalues real to machine precision, eigenvectors orthonormal, $\mathrm{Im}\langle H\rangle < 10^{-12}$ on random states, spectral reconstruction $H = \sum_i w_i|v_i\rangle\langle v_i|$ exact; non-Hermitian control: complex expectations found immediately. The $2\times2$ case $\bigl(\begin{smallmatrix}1 & 1-i\\ 1+i & 3\end{smallmatrix}\bigr)$ : eigenvalues $2 \mp \sqrt3$ , eigenvectors orthogonal. Gram–Schmidt inside a degenerate eigenspace: orthonormal pair, still eigenvectors.

Assembled, the theorems hand the measurement chapter everything it will postulate: real outcomes $a_n$ , orthogonal outcome states $|a_n\rangle$ , and — once completeness of the eigenbasis is certified (stated here as the spectral fact; its theory belongs to Eigenvalues and Eigenstates, Complete Orthonormal Bases, and, for continuous spectra, Continuous Spectra and Dirac Delta Normalization) — the decomposition

\hat A = \sum_n a_n\,|a_n\rangle\langle a_n| \qquad (4)

where: the operator is its menu of outcomes, weighted by the projectors onto their states. Expectation values then unpack exactly as the wave-mechanics chapter computed them: $\langle\hat A\rangle = \sum_n a_n|\langle a_n|\psi\rangle|^2$ — real numbers weighted by (what the Born-rule article will certify as) probabilities. Even the variance machinery of Expectation Values and Variance sits on Theorem 3: real mean, real spread, honest statistics.

The fine print: symmetric is not quite self-adjoint

For matrices, definition (1) is the whole story. For the unbounded operators of continuum physics, equation (2)'s boundary term is a live wire, and it is the honest way to see a distinction mathematicians insist on. Put the same momentum operator on the half-line $x > 0$ (a particle confined to one side of a wall) and redo the integration by parts on functions that need not vanish at the origin:

\langle f|\hat p\,g\rangle - \langle\hat p f|g\rangle = i\hbar\,f^*(0)\,g(0) \qquad (5)

where: the defect is a pure boundary term — computed exactly for explicit exponentials in the verification ( $f = e^{-x}$ , $g = e^{-2x}$ : defect $= i\hbar$ ). On the half-line, $\hat p$ is Hermitian only on states vanishing at the origin ("symmetric" on that restricted domain), and — the deeper fact — no boundary condition makes it fully self-adjoint there: a particle on a half-line has no legitimate momentum observable, though it has a perfectly good $\hat p^2$ (with appropriate wall conditions). The moral for working physics: boundary conditions are part of an operator's identity — choosing them is choosing which formal expression is actually an observable, and every "which boundary condition do I impose?" moment of the wells chapter was secretly a self-adjointness decision. The systematic resolution (deficiency indices, rigged spaces) is the chapter-end material of Rigged Hilbert Space; the flag planted here is the working physicist's share.

Limiting cases

case	result	meaning
finite dimensions	Hermitian = conjugate-transpose symmetric; all three theorems + spectral form exact	spin systems: zero fine print
$\hat x$ , $\hat V(x)$	Hermitian by reality of the multiplier	potentials are always legal observables
$\hat p$ on the full line	Hermitian on normalizable states — the $i$ repairs the parts-flip	momentum measurable; boundary term dies with the tails
bare $d/dx$	anti-Hermitian: sign flips on crossing	multiply by $\pm i$ to get an observable; the $i$ is load-bearing
products $\hat A\hat B$	Hermitian iff $[\hat A,\hat B] = 0$ ; otherwise symmetrize	quantization's ordering rule
degenerate spectra	orthogonality within an eigenspace is a choice (Gram–Schmidt)	organizable freedom, not failure
half-line $\hat p$	boundary defect $i\hbar f^*(0)g(0)$ ; no self-adjoint version	domains are physics — boundary conditions define observables

Common trapReal eigenvalues are enough — any operator with a real spectrum is an observable

Reality of the spectrum without Hermiticity buys almost nothing: a non-Hermitian matrix can have all-real eigenvalues while its eigenvectors are not orthogonal — outcomes that cannot be reliably distinguished — and while expectation values in superpositions come out complex (the verification exhibits exactly this on a random matrix). The full observable package is reality plus orthogonality plus real statistics in every state, and the converse theorem shows that package is available only from Hermiticity. Guard the direction of the logic too: Hermiticity is checked against definition (1) — by integral or by conjugate-transpose — never by inspecting a spectrum. (The fashionable PT-symmetric Hamiltonians with real spectra earn their measurability only by re-engineering the inner product itself — the exception that proves exactly this rule.)

Worked examples

Example 1 — credential audits (confidence builder). Classify as Hermitian / anti-Hermitian / neither, on normalizable states of the line: (a) $\hat T = \hat p^2/2m$ ; (b) $\hat D = d/dx$ ; (c) $\hat B = \hat x\hat p$ ; (d) $\hat S = \tfrac12(\hat x\hat p + \hat p\hat x)$ .

Step 1. (a): walk $\hat p^2$ across the mirror one factor at a time — each $\hat p$ crosses unchanged (equation 2), so $\hat T$ is Hermitian ✓. Kinetic energy is measurable, and its expectation $\langle\hat T\rangle = \|\hat p\psi\|^2/2m \ge 0$ is not only real but non-negative — the positivity that anchored the zero-point-energy arguments of the previous chapter.

Step 2. (b): one integration by parts, one sign: $\langle f|\hat D g\rangle = -\langle\hat D f|g\rangle$ — anti-Hermitian. Its eigen-"values" along imaginary directions are the tell: $e^{kx}$ solves $\hat D\psi = k\psi$ for real $k$ but is not normalizable; the honest spectrum sits on the imaginary axis, which the factor $-i\hbar$ rotates onto the real line.

Step 3. (c): the mirror reverses the pair; un-reversing costs $[\hat x,\hat p] = i\hbar$ : defect $i\hbar\langle f|g\rangle$ — neither (verified symbolically: the defect equals $+i\hbar\langle f|g\rangle$ exactly on Gaussian-damped test functions).

Step 4. (d): symmetrizing splits the ordering democratically; the two defects cancel — Hermitian ✓. This is the quantum image of the classical quantity $xp$ (the virial), and the audit you just ran is the standard quantization ritual: symmetrize until the mirror test passes.

Example 2 — measurement statistics from the theorems (exam level). A spin-½ particle is prepared in $|{+z}\rangle$ , and $S_y$ is measured. Predict everything measurable, using only this article's package.

Step 1. Credentials: $S_y = \tfrac{\hbar}{2}\bigl(\begin{smallmatrix}0 & -i\\ i & 0\end{smallmatrix}\bigr)$ equals its conjugate transpose ✓ — a legal observable despite the conspicuous $i$ 's (they sit symmetrically; compare the equally imaginary but anti-Hermitian $\bigl(\begin{smallmatrix}0&1\\-1&0\end{smallmatrix}\bigr)$ ).

Step 2. Outcomes = eigenvalues: solving gives $\pm\hbar/2$ — real, as Theorem 1 demanded — with normalized eigenvectors $|{\pm y}\rangle = \tfrac{1}{\sqrt2}(|{+z}\rangle \pm i|{-z}\rangle)$ , orthogonal as Theorem 2 demanded ( $\langle{+y}|{-y}\rangle = \tfrac12(1 + i\cdot i^*\cdot(-1))\ldots = 0$ — mind the conjugations).

Step 3. Statistics: $P(\pm\hbar/2) = |\langle{\pm y}|{+z}\rangle|^2 = \boxed{\tfrac12}$ each. Mean: $\langle S_y\rangle = 0$ — real ✓ (Theorem 3), and computable equally as the matrix sandwich, which vanishes because the diagonal of $S_y$ is zero.

Step 4. Spread: $\langle S_y^2\rangle = \hbar^2/4$ (since $S_y^2 = (\hbar/2)^2\mathbb 1$ ), so $\Delta S_y = \boxed{\hbar/2}$ — maximal: a $z$ -definite state is maximally undecided about $y$ . Every number in this example is machine-verified; and the whole exercise previews the measurement postulates with nothing assumed beyond the three theorems.

The man who made observables rigorous

Origins · 1 minJohn von Neumann and how this was discovered

John von Neumann

1903–1957 · Budapest, Göttingen, Princeton

The adjective is Hermite's — his 1855 reality theorem is why "Hermitian" means measurable, a story told with his band at this article's wave-register twin, Operators and Observables — but the theory here is von Neumann's. Arriving in Göttingen in 1926 as Hilbert's assistant, he watched physicists assert that every symmetric operator had real spectrum and a complete eigenbasis, knew it was false as stated, and spent 1927–1929 building the mathematics that made it true where it is true: the symmetric-versus-self-adjoint distinction (this article's half-line momentum is his textbook casualty), the deficiency-index machinery counting an operator's self-adjoint extensions — the formal version of "boundary conditions are physics" — and the spectral theorem for unbounded self-adjoint operators, finally certifying real spectra and complete eigenbases for the actual operators of quantum mechanics, not just matrices. His 1932 Mathematische Grundlagen installed "observable = self-adjoint operator" as a postulate with proofs behind it. His band at Bound States versus Scattering States tells the spectral-decomposition half of the story — discrete versus continuous spectra; here he is why this article's fine print is fine print rather than quicksand. Game theory, the stored-program computer, and the Manhattan Project fit in the same twenty years.

History

1855

— Hermite proves matrices equal to their conjugate transpose have real eigenvalues

1904

— Hilbert's spectral theory of symmetric forms: orthogonality and expansions in infinite dimensions

1925

— Born, Heisenberg, and Jordan need real-eigenvalue arrays for observables — Hermite's matrices take the job

1927

— von Neumann distinguishes symmetric from self-adjoint and builds the spectral theorem for the latter

1932

— the Mathematische Grundlagen fixes "observable = self-adjoint operator" as the standing postulate

Going further

References. Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., §3.2–3.3 · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §1.3 · Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 1 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. I, Ch. II · Bonneau, Faraut & Valent, Am. J. Phys. 69, 322 (2001) — self-adjoint extensions for physicists.

Provenance. All theorems and equations are standard textbook results. A verifier should: (1) reproduce the reality and orthogonality theorems by direct conjugation (Griffiths §3.3, Sakurai §1.3, Shankar Ch.1); (2) confirm the spectral decomposition follows from the resolution of identity; (3) check the adjoint-in-components identity $(A^\dagger)_{ij}=\overline{A_{ji}}$ ; (4) verify Pauli-matrix eigenvalues $\pm1$ giving spin outcomes. For this edition all four were machine-executed (sympy/numpy), plus the credential audits this article adds: $\hat x$ and $\hat p$ Hermitian on explicit Gaussian-damped pairs (boundary terms evaluated); $d/dx$ anti-Hermitian; the $\hat x\hat p$ defect $= +i\hbar\langle f|g\rangle$ exactly and its symmetrized repair; random-matrix verification of all three theorems with a non-Hermitian control (complex expectations found); spectral reconstruction; the $2\times2$ case ( $2 \mp \sqrt3$ ); Gram–Schmidt within a degenerate eigenspace; Pauli eigenvalues $\pm1$ with the full Example-2 statistics ( $P = \tfrac12$ each, $\langle S_y\rangle = 0$ , $\Delta S_y = \hbar/2$ ); and the half-line boundary defect $i\hbar f^*(0)g(0) = i\hbar$ for $e^{-x}, e^{-2x}$ . Per the chapter ownership map: the adjoint/dagger formalism is owned by Hermitian Conjugate and Adjoint Operators, eigenbasis completeness by Eigenvalues and Eigenstates + Complete Orthonormal Bases (+ CSPEC for continua), the measurement postulates by The Postulates of Quantum Mechanics / Born Rule / Measurement and Collapse, and the wave-register correspondence story by the previous chapter's Operators and Observables — this article recalls it by title and proves its formal engine.

Builds on: Hilbert Space and State Vectors · Inner Product and Norm of States · Dirac Bra-Ket Notation · Linear Operators on Hilbert Space · Eigenvalues and Eigenstates

Leads to: The Postulates of Quantum Mechanics · Born Rule and Measurement Probabilities · Completeness Relation and Resolution of Identity · Projection Operators · Expectation Values and Variance · Measurement and Collapse

Is "Hermitian" the same as "self-adjoint"? Textbooks seem to use them interchangeably.

For matrices and bounded operators, identical. For unbounded operators they split: "symmetric/Hermitian" means definition (1) holds on some domain; "self-adjoint" additionally requires the domain bookkeeping to close exactly, and only self-adjoint operators are guaranteed real spectra, complete eigenbases, and legal $e^{-i\hat At}$ evolutions. Physics texts blur the distinction because the standard full-line operators are safely self-adjoint; the half-line example is this article's controlled exhibit of the gap. When boundary conditions are in play, think "self-adjoint."

Why do observables need orthogonal eigenstates — couldn't nearly-orthogonal work?

Repeatability breaks. Measure, get outcome $a$ : the state is the corresponding eigenstate (collapse postulate ahead), and an immediate re-measurement must return $a$ with certainty — requiring zero amplitude on every other outcome state: exact orthogonality. "Nearly orthogonal" outcomes would let a second look contradict the first, and the inner-product article's distinguishability logic shows such readings never deserved to be called distinct values of one quantity.

The eigenvalues of x̂ and p̂ form continua — do the three theorems still apply?

The content survives; the packaging changes. Continuum "eigenstates" ( $|x\rangle$ , $|p\rangle$ ) are delta-normalized objects outside the Hilbert space, orthogonality becomes $\langle x|x'\rangle = \delta(x-x')$ , and the spectral sum (4) becomes an integral. Making that honest is the program of Continuous Spectra and Dirac Delta Normalization and Rigged Hilbert Space at the chapter's end. Every wave-packet calculation (honest states) inherits the theorems with no asterisks.

Can a Hermitian operator fail to correspond to anything actually measurable?

Yes — Hermiticity is necessary, not sufficient. Superselection rules forbid some Hermitian operators (nothing connects different charges); locality forbids others; practicality forbids most (a random Hermitian matrix on a many-body space has no conceivable apparatus). The formalism is permissive by design; physics — symmetries, locality, hardware — curates the menu. The converse direction is solid: anything measurable must be represented Hermitianly.

Where does the variance/uncertainty machinery fit into this article's theorems?

Directly on top. Theorem 3 makes $\langle\hat A\rangle$ real; then $\Delta A^2 = \langle(\hat A - \langle A\rangle)^2\rangle = \|(\hat A - \langle A\rangle)\psi\|^2$ — real and non-negative by construction. Pair two observables and apply Cauchy–Schwarz to those displaced vectors and the generalized uncertainty principle appears — the derivation the Generalized Uncertainty Principle article executes. Measurement statistics is Hermiticity plus geometry, nothing else.

Are anti-Hermitian operators physically useless, then?

Far from it — they are the generators. Every anti-Hermitian operator is $i$ times a Hermitian one, and their exponentials are exactly the unitary transformations: $e^{-i\hat Ht/\hbar}$ (evolution), $e^{-i\hat pa/\hbar}$ (translation), $e^{-i\hat J_z\phi/\hbar}$ (rotation) — Stone's theorem, delivered in Unitary Operators and Transformations. Division of labour: Hermitian = what you measure; $i\times$ Hermitian, exponentiated = what the system does.