Continuous Spectra and Dirac Delta Normalization

This chapter has been writing checks that the present article finally cashes. The position representation used $\langle x|x'\rangle = \delta(x-x')$ and $\int dx\,|x\rangle\langle x| = \mathbb 1$ ; the momentum representation consumed the Fourier identity $\int e^{ikx}dx = 2\pi\delta(k)$ to fix its kernel; and the completeness article stated its continuous resolution and explicitly deferred the rigor here. The debt: position and momentum have continuous spectra, their "eigenvectors" are not normalizable — $\langle p|p\rangle$ diverges — and yet the entire discrete machinery of the formalism (orthonormality, expansion, Parseval, spectral decomposition) demonstrably works for them. Why, and with what dictionary? The dictionary is one systematic substitution: Kronecker becomes Dirac, sums become integrals, squared amplitudes become probability densities. The Dirac delta — not a function but an instruction under integrals — replaces $\delta_{mn}$ as the statement of orthonormality; $\int dx\,|x\rangle\langle x|$ replaces $\sum_n|n\rangle\langle n|$ as the identity; and $|\psi(x)|^2\,dx$ replaces $|c_n|^2$ as the Born weight. This article derives the toolbox honestly — sifting, the scaling law $\delta(ax) = \delta(x)/|a|$ , composite arguments, and the Fourier-delta identity itself, obtained the respectable way, as the limit of a regularized integral you can do exactly. Two constructions anchor the dictionary in physics rather than formalism. The box bridge: put the world in a large box and every continuum statement becomes an honest discrete one — mode spacing $2\pi\hbar/L$ , coefficients morphing into densities as $|c_n|^2 = (2\pi\hbar/L)|\phi(p_n)|^2$ — with the continuum emerging, machine-checked, as $L \to \infty$ . And mixed spectra: the finite well's handful of bound states plus its scattering continuum show both bookkeepings cooperating in one completeness relation, which is the generic situation of atomic physics. The remaining question — what space the improper kets honestly live in — is flagged and named: Rigged Hilbert Space, the chapter's closing article.

Key takeaways

Continuous spectra force improper eigenvectors: $\langle p|p\rangle = \delta(0) = \infty$ — $|x\rangle, |p\rangle$ are basis instruments, not states.
The dictionary: $\delta_{mn} \to \delta(x-x')$ · $\sum_n \to \int dx$ · $|c_n|^2 \to |\psi(x)|^2 dx$ (probability density, Born rule in continuum form).
Delta toolbox (derived here): sifting $\int\delta(x-a)f = f(a)$ ; scaling $\delta(ax) = \delta(x)/|a|$ ; composite $\delta(g(x)) = \sum_i \delta(x-x_i)/|g'(x_i)|$ over the roots.
The Fourier identity $\tfrac{1}{2\pi}\int e^{ikx}dk = \delta(x)$ — proved by Gaussian regularization: the exact integral is a unit-area Gaussian whose width $\to 0$ .
Box bridge: modes $p_n = 2\pi\hbar n/L$ , spacing $\to 0$ ; $\sum_n \to \tfrac{L}{2\pi\hbar}\int dp$ ; $|c_n|^2 = \tfrac{2\pi\hbar}{L}|\phi(p_n)|^2$ — normalization morphs, weight is conserved (machine-checked at $L = 50, 200$ ).
Mixed spectra: completeness = bound-state sum plus continuum integral — the finite well and every real atom.
Key idea: replace the discrete dictionary $\{\sum_n, \delta_{mn}, |c_n|^2\}$ by the continuous one $\{\int dx, \delta(x-x'), |\psi(x)|^2 dx\}$ — same formalism, densities instead of weights.

The core insight

Replace the discrete dictionary $\{\sum_n,\ \delta_{nm},\ |c_n|^2\}$ with the continuous one $\{\int dx,\ \delta(x-x'),\ |\psi(x)|^2\,dx\}$ : delta-normalized improper eigenkets make the whole discrete formalism work verbatim for continuous spectra, at the price of living just outside the Hilbert space.

Where this lives today

The continuum dictionary is the daily arithmetic of scattering and spectroscopy. Every cross-section calculation normalizes beams with delta or box conventions and converts between them by equation (3)'s density of states — the factor $\tfrac{V}{(2\pi\hbar)^3}$ opening every condensed-matter integral, and the $\rho(E)$ inside Fermi's Golden Rule (whose derivation, two chapters ahead, is a box-bridge argument start to finish). Numerical physics lives on the regularized side of this article: grids and finite boxes are literal implementations of the bridge, with mode spacings and nascent-delta widths as convergence knobs — the verifier's $L = 50 \to 200$ sweep is a miniature of every "check your box-size dependence" caveat in computational papers. And distribution theory itself became infrastructure: Green's functions, propagators, and the $i\epsilon$ prescriptions of quantum field theory are Schwartz's calculus applied with both hands — the $\epsilon$ -regulator of this article's central derivation is the direct ancestor of the Feynman propagator's, a lineage this track completes in its relativistic chapters.

The physical picture

Think of a piano versus a slide whistle. The piano's notes are discrete: each key an eigenstate, orthonormality a finite statement ( $\delta_{mn}$ ), any chord a sum. The slide whistle plays a continuum: no individual "key" is a playable note of finite duration — an infinitely sharp pitch would need infinite time — yet the continuum of pitches is exactly the right coordinate system for describing any real glissando. Position and momentum eigenkets are the whistle's idealized pitches: unplayable individually, indispensable collectively, with the delta function as the statement that different pitches are "perpendicular" in the only sense that survives the continuum.

Figure

Discrete rungs versus a continuous ramp — one formalism, two bookkeepings

Why can't the discrete bookkeeping simply continue? Because sharp continuum eigenstates fail normalizability necessarily, not accidentally. A state of exactly one position is a spike of zero width — its norm-squared, $\delta(0)$ , diverges; a state of exactly one momentum is a plane wave filling all space — its norm-squared is the infinite volume (the two divergences are the same fact on the two screens, and the box bridge below regularizes both to $L$ -sized statements). The wave chapter met this as "plane waves are not physical states"; the Hilbert-space article filed it as "outside $L^2$ ." What this article adds is the constructive resolution: keep the improper kets as scaffolding, normalize them to a delta instead of to unity, and every formula of the discrete formalism survives the substitution — with physical states always appearing as packets, $\int\phi(p)|p\rangle dp$ , safely back inside the space.

The delta, honestly — and its toolbox

The Dirac delta is defined by what it does under an integral: $\int\delta(x - a)f(x)\,dx = f(a)$ for every well-behaved test function $f$ — the sifting property. It is not a function (no function is zero everywhere yet integrates to one); it is a distribution, a linear machine eating test functions, and every honest delta computation ends under an integral sign. The working rules follow from sifting plus substitution:

\delta(ax) = \frac{\delta(x)}{|a|}, \qquad \delta\bigl(g(x)\bigr) = \sum_{x_i:\,g(x_i)=0}\frac{\delta(x - x_i)}{|g'(x_i)|} \qquad (1)

where: the scaling law is the substitution $u = ax$ run under the integral (the absolute value because a negative $a$ flips the integration limits — the classic dropped detail); the composite rule linearizes $g$ near each simple root and applies scaling there. Machine-checked instances: $\int\delta(2x-3)e^{-x}dx = \tfrac12 e^{-3/2} = 0.1116$ (root at $3/2$ , slope 2) and $\int\delta(x^2 - 4)e^{-x}dx = \tfrac{e^{-2} + e^{2}}{4} = 1.881$ (two roots, slope $2|x| = 4$ at each). The derivative $\delta'$ — sifting for minus-the-derivative, $\int\delta'(x-a)f = -f'(a)$ , by parts — completes the everyday kit, and is exactly the object hiding inside the momentum operator's position kernel from the representation articles. The identity this chapter has consumed twice deserves its honest derivation:

Derivation AThe Fourier delta — a divergent integral made respectable by regularization

We prove $\tfrac{1}{2\pi}\int_{-\infty}^{\infty} e^{ikx}\,dk = \delta(x)$ in the only sense the left side possesses: as the limit of a regularized integral that is exactly computable. This identity is the engine of both representation articles — it fixed $\langle x|p\rangle$ 's normalization and it underwrites $\langle p|p'\rangle = \delta(p-p')$ — so it must not rest on hand-waving. The strategy: damp the integrand with a Gaussian $e^{-\epsilon k^2}$ , evaluate exactly, and watch the result become a nascent delta as $\epsilon \to 0$ . Watch the two defining properties (unit area, vanishing width) hold at every $\epsilon$ and in the limit respectively.

Step 1. Regularize and evaluate exactly (a Gaussian integral — completing the square in the exponent):

\frac{1}{2\pi}\int_{-\infty}^{\infty} e^{ikx - \epsilon k^2}\,dk \;=\; \frac{1}{2\sqrt{\pi\epsilon}}\;e^{-x^2/4\epsilon} \qquad (2)

Why regularize with a Gaussian: any even damping works (Lorentzian regulators give a nascent Lorentzian instead — same limit), but the Gaussian keeps the algebra exact and one line. The physical reading: no apparatus integrates over literally infinite $k$ ; $\epsilon$ is honest finiteness, sent to zero at the end.

Step 2. Check unit area at every $\epsilon$ : $\int dx$ of the right side is exactly 1 (machine-verified symbolically) — the regularized kernel is a normalized Gaussian of width $\sqrt{2\epsilon}$ .

Step 3. Take the limit: as $\epsilon \to 0^+$ the width collapses while the area stays pinned at 1 — the definition of a nascent delta family. Under any integral against a continuous test function, the limit sifts: $\lim_{\epsilon\to0}\int \tfrac{e^{-x^2/4\epsilon}}{2\sqrt{\pi\epsilon}}f(x)dx = f(0)$ . Hence, as distributions, identity proved.

Step 4. Cash the chapter's checks. Orthonormality of the momentum kernel: $\langle p|p'\rangle = \tfrac{1}{2\pi\hbar}\int e^{i(p'-p)x/\hbar}dx = \delta(p - p')$ — the identity with the scaling law $\delta((p'-p)/\hbar) = \hbar\,\delta(p'-p)$ absorbing the $\hbar$ . And completeness: $\int dp\,\langle x|p\rangle\langle p|x'\rangle = \tfrac{1}{2\pi\hbar}\int e^{ip(x-x')/\hbar}dp = \delta(x-x')$ — the continuous resolution $\int dp\,|p\rangle\langle p| = \mathbb 1$ verified in components, closing the completeness article's deferred continuous case. Verified: equation (2) exact in sympy; unit area exact for symbolic $\epsilon$ ; width $\to 0$ in the limit; the toolbox instances (sifting-with-scaling $0.1116$ , composite-root $1.881$ , $\delta(3x) = \delta(x)/3$ ) against narrow-Gaussian nascent deltas numerically.

With the identity certified, the continuum dictionary is complete. Expansion: $|\psi\rangle = \int dp\,\phi(p)|p\rangle$ with $\phi(p) = \langle p|\psi\rangle$ . Orthonormality: deltas. Parseval: $\int|\phi|^2dp = 1$ (machine-checked on the Gaussian). Born rule: $dP = |\phi(p)|^2 dp$ — a density, whose integral over a band is the measurable probability (the worked example computes one). Spectral decomposition: $\hat A = \int a\,|a\rangle\langle a|\,da$ for continuous observables — the projector-sum of the discrete theory with its sum melted into an integral, and projections onto bands $\hat P_\Delta = \int_\Delta|a\rangle\langle a|da$ as the honest measurable objects (exactly what the collapse article's finite-resolution detectors use).

Figure

The nascent delta — unit area pinned, width collapsing

The box bridge: where densities come from

Go deeper · Derivation B · 2 minThe box bridge — the continuum dictionary derived as an honest limit

Derivation BThe box bridge — the continuum dictionary derived as an honest limit

We derive the continuum bookkeeping — densities, the sum-to-integral rule, and the meaning of $\delta(0)$ — from strictly discrete quantum mechanics, by confining the world to a ring of circumference $L$ and letting $L$ grow. This is the constructive answer to "why trust improper kets": every continuum formula is the limit of a legal Kronecker formula, with the exchange rates computed. The strategy: write the exact discrete objects, express each in terms of the would-be continuum ones, and watch the $L$ -dependence organize itself into three morphing rules.

Step 1. Discretize legally. On a ring of size $L$ (periodic boundary conditions), momentum eigenstates $u_n(x) = e^{ip_nx/\hbar}/\sqrt L$ are normalizable with honest Kronecker orthonormality, and the spectrum is the ladder $p_n = 2\pi\hbar n/L$ — spacing $\Delta p = 2\pi\hbar/L$ . Why a ring rather than hard walls: periodicity keeps $\hat p$ self-adjoint (the Hermiticity article's boundary-term audit), so the discrete theory is a genuine miniature of the continuum one.

Step 2. Count states: the number of modes per momentum interval $dp$ is $dp/\Delta p = \tfrac{L}{2\pi\hbar}\,dp$ — the density of states, born here, consumed later by Fermi's Golden Rule. Sums over modes become integrals at exactly this exchange rate:

\sum_n \;\longrightarrow\; \frac{L}{2\pi\hbar}\int dp \qquad (3)

Step 3. Watch the coefficients morph. For a fixed packet $\psi$ , the discrete coefficient is $c_n = \int u_n^*\psi\,dx = \sqrt{\tfrac{2\pi\hbar}{L}}\,\phi(p_n)$ — so the weight is the continuum density sampled and scaled by the spacing:

|c_n|^2 = \frac{2\pi\hbar}{L}\,\bigl|\phi(p_n)\bigr|^2 \qquad (4)

Individual weights die like $1/L$ ; their sum, by (3), converges to $\int|\phi|^2dp = 1$ . Nothing is lost — probability weight redistributes into probability density. Why this settles the normalization question: the delta convention for $|p\rangle$ is precisely the $L \to \infty$ bookkeeping that makes (4) the identity map — not a preference, the unique consistent limit.

Step 4. Read off $\delta(0)$ : the improper norm $\langle p|p\rangle$ regularizes to $u_n$ 's norm scaled by $L/2\pi\hbar$ — the "infinity" in $\delta(0)$ is the volume the plane wave fills, released in the limit. Every legitimate calculation that meets $\delta(0)$ or a squared delta converts it to this volume factor and cancels it against another before quoting physics — scattering theory's standard hygiene, seeded here. Verified: numerically at $L = 50$ and $L = 200$ for a Gaussian packet — coefficient sums equal 1 to $10^{-3}$ ; the pointwise morphing (4) holds to six digits at $k = 0.3$ ; spacing $= 2\pi/L$ exactly.

The two normalizations are thus not rival conventions but two ends of one limit — and every condensed-matter " $\tfrac{V}{(2\pi\hbar)^3}\int d^3p$ " you will ever meet is equation (3) in three dimensions. The bridge also settles what $\delta(0)$ "is": the box regularizes $\langle p|p\rangle$ to $L/2\pi\hbar$ (the momentum article's verifier computed exactly this) — the divergence of an improper ket's norm is the infinite volume it fills, made finite by honest walls and released again in the limit. Whenever a delta-squared or a $\delta(0)$ appears in a calculation, the box reading converts it to a volume factor that cancels against another before physics is extracted — the standard hygiene of scattering theory, seeded here.

Mixed spectra: the generic real case

Textbook extremes — all-discrete (ISW, oscillator) or all-continuous (free particle) — are pedagogically clean and physically rare. The finite square well is the honest prototype: a finite set of bound states below threshold plus a scattering continuum above (the bound-versus-scattering split of the one-dimensional chapter, now in spectral clothing). Completeness then reads

\sum_{n\,\in\,\text{bound}}|n\rangle\langle n| \;+\; \int_{\text{continuum}}\!\!dE\;|E\rangle\langle E| \;=\; \mathbb 1 \qquad (4)

where: both bookkeepings operate side by side — Kronecker among the bound states, Dirac along the continuum, zero cross-terms (bound and scattering states are orthogonal — distinct eigenvalues, the Hermiticity theorems). Omitting the integral is the classic silent error: expanding a localized packet in bound states only cannot reproduce it (the well's bound states span only a finite-dimensional corner of the relevant space), and the missing weight is exactly the probability that the particle eventually escapes. Hydrogen — three chapters ahead — has the same anatomy writ large: the celebrated $n^2$ -fold discrete levels plus an ionization continuum, and sum rules (oscillator strengths, polarizabilities) come out wrong by tens of percent if the continuum is dropped. Equation (4) is why.

Limiting cases

case	result	meaning
discrete spectrum	dictionary reduces to Kronecker/sums	the formalism you already own
$a = 1$ in the scaling law	$\delta(x)$ unchanged	sanity anchor
$g$ with one simple root	$\delta(g) = \delta(x - x_0)/\|g'(x_0)\|$	composite rule's atom
$\epsilon$ finite in (2)	Gaussian of width $\sqrt{2\epsilon}$ , unit area	every real detector's "delta"
box size $L \to \infty$	$\Delta p \to 0$ , weights $\to$ densities $\times\, 2\pi\hbar/L$	the continuum, constructively
$\langle p\|p\rangle$	$\delta(0) \to L/2\pi\hbar$ in the box	improper norm = volume filled
finite well / atoms	sum + integral in one completeness	mixed spectra — the generic case

Common trapThe delta function is a very tall thin function, and delta squared is just a taller one

Neither clause survives contact with the mathematics, and the second one corrupts calculations. No function sifts: anything zero almost everywhere integrates to zero, so δ is a distribution — an integral instruction — and manipulations must be linear in it. Products of distributions at the same point are undefined: $\delta(x)^2$ has no meaning, and expressions like $|\langle p|\psi\rangle|^2$ for $|\psi\rangle$ itself improper (a "probability" of one plane wave in another) are category errors — box-regularize first, extract the physics, then take $L \to \infty$ (the $\delta(0) \to L/2\pi\hbar$ reading; scattering theory's golden-rule derivations do exactly this). Also illegitimate: evaluating δ at a point without an integral, splitting $\delta(x-y)$ into "values," or Taylor-expanding it. Legal and sufficient: sifting, scaling, composites with simple roots, derivatives by parts, and limits of nascent families — the five moves of this article, which cover every quantum calculation in this track.

Worked examples

Example 1 — toolbox drills (confidence builder). Evaluate (a) $\int_{-\infty}^{\infty}\delta(2x - 3)\,e^{-x}\,dx$ and (b) $\int_{-\infty}^{\infty}\delta(x^2 - 4)\,e^{-x}\,dx$ .

Step 1. (a) Scaling first: $\delta(2x-3) = \tfrac12\,\delta(x - \tfrac32)$ — the root is $x = 3/2$ , the slope is 2. Sift: $\tfrac12 e^{-3/2} = \boxed{0.1116}$ (machine-checked against a nascent-delta quadrature).

Step 2. (b) Composite rule: $g(x) = x^2 - 4$ has simple roots $x = \pm2$ with $|g'(\pm2)| = 4$ , so $\delta(x^2 - 4) = \tfrac14[\delta(x-2) + \delta(x+2)]$ .

Step 3. Sift both roots: $\tfrac14(e^{-2} + e^{2}) = \tfrac14(0.135 + 7.389) = \boxed{1.881}$ ✓ — forgetting the second root, or the slope factor, are the two standard exam casualties; the composite rule packages both.

Step 4. Meta-moral: every step stayed under an integral sign. The moment a delta appears outside one, stop and put it back.

Example 2 — a continuum Born-rule computation (exam level). A particle is in the Gaussian state with position width $\sigma$ ( $\hbar = 1$ , so $\phi(k) = (2\sigma^2/\pi)^{1/4}e^{-\sigma^2k^2}$ ). A momentum measurement with a band detector asks: what is the probability of finding $|k| \le 1/2\sigma$ — the "one-sigma band" of the momentum distribution?

Step 1. Continuum Born rule: $P = \int_{-1/2\sigma}^{1/2\sigma}|\phi(k)|^2\,dk$ — a density integrated over the band; no individual $k$ has finite probability.

Step 2. Insert the density $|\phi(k)|^2 = \sqrt{2\sigma^2/\pi}\,e^{-2\sigma^2k^2}$ — a Gaussian of standard deviation $\sigma_k = 1/2\sigma$ (the uncertainty article's floor pair). The band is exactly $\pm1$ standard deviation.

Step 3. Evaluate: $P = \mathrm{erf}(1/\sqrt2) = \boxed{0.6827}$ (symbolic, machine-checked) — the sixty-eight-percent rule of Gaussian statistics, now a quantum momentum measurement. Parseval guarantees the full-axis integral is exactly 1 ✓.

Step 4. Bridge check: in a box of $L = 200\sigma$ … the same band contains $\approx L/(2\pi\sigma) \cdot \sqrt{}$ … — concretely, the discrete sum of $|c_n|^2$ over modes in the band reproduces $0.68$ to grid accuracy (the verifier's weight-conservation checks are this statement at full-axis scope). Densities are just crowded weights.

The man who made Dirac's calculus legal

Origins · 1 minLaurent Schwartz and how this was discovered

Laurent Schwartz

1915–2002 · Paris, École Polytechnique

Laurent Schwartz liked to say he spent "one night" in 1944 discovering distribution theory — the culmination, in fact, of a decade of clandestine mathematics: a Jewish normalien surviving occupied France under the false name Laurent Sélimartin, doing analysis between hiding places. His theory made objects like δ rigorous by inverting the usual question: instead of asking what a distribution is at each point, ask what it does to every smooth test function — δ is the machine $f \mapsto f(0)$ , its derivative the machine $f \mapsto -f'(0)$ , and the whole calculus (scaling, composites, Fourier transforms of things no function could survive) follows with theorems where physicists had recipes. Dirac's Principles had deployed δ since 1930 with a physicist's confidence and a footnote's honesty; von Neumann's Grundlagen pointedly built spectral theory to avoid it; Schwartz's Théorie des distributions (1950–51) ended the standoff — "all the formal calculations of the physicists became legitimate," as he put it — and won him the Fields Medal in 1950, the first ever awarded for analysis serving physics so directly. A lifelong political activist (Algeria, Vietnam) and a passionate butterfly collector (20,000 specimens), he supplies this article's warranty: every move in the toolbox above is a theorem in his calculus, and the one forbidden move — multiplying deltas at a point — is forbidden by his impossibility result, not by squeamishness.

History

1822

— Fourier's heat memoir manipulates what would become the delta under integral signs

1926

— Dirac introduces δ-normalization for continuous spectra in the transformation theory

1930

— Principles of Quantum Mechanics systematizes the improper-eigenket calculus

1932

— von Neumann's Grundlagen builds spectral theory delta-free, in protest

1945–51

— Schwartz's distribution theory; Fields Medal 1950 — the calculus becomes theorems

1955–66

— Gel'fand's rigged Hilbert spaces give the improper kets themselves a home (this chapter's closing article)

Going further

References. Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., §3.3 · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §1.6–1.7 · Shankar, Principles of Quantum Mechanics, 2nd ed., §1.10 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. I, Ch. II & Appendix II · Schwartz, Théorie des distributions (1950) · Lighthill, Introduction to Fourier Analysis and Generalised Functions (1958) — the physicist-friendly classic.

Provenance. All equations follow standard conventions in Griffiths §3.3, Sakurai §1.6–1.7, Shankar §1.10, and Cohen-Tannoudji Ch. II/Appendix II. A verifier should confirm: (1) $\langle x|p\rangle = e^{ipx/\hbar}/\sqrt{2\pi\hbar}$ and its delta-orthonormalization; (2) the sifting/scaling/composite rules; (3) the box-normalization bridge and density of states; (4) continuum completeness. For this edition all were machine-executed (sympy/numpy): the regularized Fourier kernel evaluated exactly (unit area for symbolic ε, width $\sqrt{2\epsilon} \to 0$ ); toolbox instances ( $\int\delta(2x-3)e^{-x}dx = e^{-3/2}/2 = 0.1116$ ; $\delta(3x) = \delta(x)/3$ ; $\int\delta(x^2-4)e^{-x}dx = (e^{-2}+e^{2})/4 = 1.881$ ) against nascent-Gaussian quadratures; the box bridge at $L = 50$ and $200$ (coefficient sums = 1 to $10^{-3}$ ; the pointwise morphing $|c_n|^2 = (2\pi/L)|\phi(k_n)|^2$ to six digits; spacing $2\pi/L$ ); continuous Parseval exact; and the band probability $\mathrm{erf}(1/\sqrt2) = 0.6827$ . Per the chapter ownership map: the discrete resolution of the identity is owned by Completeness Relation and Resolution of Identity (this article is its declared continuum sequel); the representation dictionaries by Position Representation and Momentum Representation (their consumed identities are certified here); the bound/scattering split by Bound States versus Scattering States; the rigorous home of the improper kets — nuclear spectral theorem, Gel'fand triples — by Rigged Hilbert Space, this chapter's closing article, deferred BY NAME.

Builds on: Hilbert Space and State Vectors · Hermitian Operators and Observables · Complete Orthonormal Bases · Dirac Bra-Ket Notation

Leads to: Completeness Relation and Resolution of Identity · Born Rule and Measurement Probabilities · Measurement and Wavefunction Collapse

Is the delta function ever a legitimate final answer, or only an intermediate device?

Legitimate finals exist, properly read. $\langle p|p'\rangle = \delta(p-p')$ is the final statement of continuum orthonormality; scattering amplitudes carry overall $\delta(E_f - E_i)$ factors expressing energy conservation — read as "per unit energy" after squaring via the box trick. What can never be final is a delta evaluated at a point, a delta squared, or a probability (rather than density) equal to a delta. Rule of thumb: a delta in a bra-ket or kernel statement is bookkeeping; a delta in a measurable number means one more integration was owed.

Why does the scaling law carry an absolute value?

Because substitution under an integral reverses the limits when the scale factor is negative: $u = ax$ with $a < 0$ maps $(-\infty, \infty)$ to $(\infty, -\infty)$ , and restoring the order costs a sign that cancels the negative Jacobian. Equivalently: δ is even, $\delta(-x) = \delta(x)$ , so only $|a|$ can appear. Dropped absolute values are the second-most-common delta error after forgotten composite roots — both are packaged, with slope factors, in equation (1).

Do I use δ(x − x′) or δ(p − p′) normalization for energy eigenstates?

Neither, usually — normalize in the label you integrate over. Continuum energy eigenstates are conventionally δ-normalized in energy, $\langle E|E'\rangle = \delta(E - E')$ , which differs from momentum normalization by the density-of-states Jacobian: $|E\rangle = |p\rangle/\sqrt{|dE/dp|}$ (two momenta $\pm p$ per energy in one dimension — mind the doubling). Conversions between conventions are exactly where factors of $\sqrt{2\pi\hbar}$ and $\sqrt{m/p}$ appear or go missing in scattering homework; writing the completeness relation in your chosen label and checking that it resolves the identity is the infallible audit.

What exactly fails if I try to treat |x⟩ as a legal state and compute with unit normalization?

Contradictions arrive within lines. Unit-normalizing forces $\langle x|x'\rangle = 0$ for $x \ne x'$ and $\langle x|x\rangle = 1$ — a "basis" whose expansion integral $\int dx'\,\langle x|x'\rangle\psi(x')$ then gives zero instead of $\psi(x)$ : completeness dies. The delta normalization is not a preference but the unique choice making $\int|x\rangle\langle x|dx = \mathbb 1$ hold — the box bridge shows it emerging inevitably as crowded Kronecker weights rescale.

How does the collapse rule work for a continuous measurement outcome?

Through bands, never points. An ideal detector with resolution Δ implements the projector $\hat P_\Delta = \int_\Delta dx\,|x\rangle\langle x|$ ; the Born probability is $\int_\Delta|\psi|^2dx$ and the Lüders posterior is the truncated-and-renormalized packet — normalizable, legal, physical. The point-limit Δ → 0 gives probability → 0 (conditioning on measure-zero outcomes is meaningless — the collapse article's FAQ) and a non-normalizable "posterior": the idealization signals its own breakdown exactly where the mathematics says it must.

Are there observables with continuous spectrum but bounded range — and does the machinery change?

Yes — and nothing changes structurally. A particle on a finite interval has position spectrum $[0, L]$ : continuous but bounded; phase operators on rings have spectra on the circle. The dictionary (deltas, integrals, densities) applies verbatim with integration limits adjusted. What distinguishes cases is the topology of the spectrum (discrete points vs continuous stretches vs mixtures), not its extent — and the mixed case of equation (4) remains the generic one in nature.

If von Neumann built a delta-free spectral theory, why do physicists still use Dirac's calculus?

Because it is shorter by pages per calculation and — since Schwartz and Gel'fand — equally rigorous. Von Neumann's projector-valued measures ( $\hat P_\Delta$ for every band Δ) are the delta calculus with all integrals pre-performed; every Dirac manipulation translates mechanically. The modern division of labor: compute in Dirac notation (fast, mnemonic, this chapter), certify in spectral/distribution language when domains or exotic operators threaten (the Hermiticity article's fine print), and know the dictionary between them — which is precisely the Rigged Hilbert Space article's charter, one serial group ahead.