Normalization

Normalization fixes the total weight of a quantum state so that the Born rule defines a probability law. For an ordinary pure-state vector,

$$\langle\psi|\psi\rangle=1.$$

This condition is not cosmetic. If a complete measurement is applied to a unit vector, its outcome probabilities sum to one. The same abstract condition appears as a coefficient sum in a discrete basis, an integral for a wavefunction, or a trace condition for a density operator.

Several conventions that share the word “normalization” must be kept distinct:

Object	Normalization statement	Meaning
ordinary pure state	$\langle\psi\rvert\psi\rangle=1$	total probability one
density operator	$\operatorname{Tr}\rho=1$	total probability one
conditional branch	$\operatorname{Tr}\widetilde\rho=p\le1$	branch probability retained
continuum basis ket	$\langle a\rvert a'\rangle=\delta(a-a')$	distributional basis convention
box mode	$\langle n\rvert m\rangle=\delta_{nm}$	finite-volume basis convention
scattering mode	prescribed incident flux	comparison of rates or cross sections

Unit vectors and positive trace-one operators describe normalized physical states. Box modes are also ordinary unit vectors for the finite-volume system. Delta and flux normalization instead specify conventions for continuum modes. A subnormalized positive operator retains the probability that a selected branch occurred.

Helpful background. State Vectors explains norms and rays; Wavefunctions as Representations explains why integration measures matter.

Why Total Weight Must Be One

Let a projective measurement have mutually orthogonal projectors $\{P_a\}$ with

$$\sum_a P_a=I.$$

For a vector $|\psi\rangle$, the Born weights are

$$w(a)=\langle\psi|P_a|\psi\rangle.$$

Summing over all outcomes gives

$$\begin{aligned} \sum_a w(a) &= \langle\psi| \left(\sum_a P_a\right) |\psi\rangle \\ &= \langle\psi|\psi\rangle. \end{aligned}$$

Thus the weights form a normalized probability distribution precisely when $\|\psi\|=1$. For a normalized POVM $\{E_a\}$, the same argument uses $\sum_aE_a=I$.

If $|\chi\rangle$ is nonzero but not normalized, probabilities can be written as ratios:

$$p(a) = \frac{\langle\chi|P_a|\chi\rangle} {\langle\chi|\chi\rangle}.$$

This ratio is invariant under any nonzero complex rescaling $|\chi\rangle\mapsto\lambda|\chi\rangle$. Unit normalization packages the denominator once so subsequent probability formulas are simpler.

The Born Rule turns amplitudes into probabilities. Normalization ensures that a complete set of outcomes has total probability one.

Normalizing a Nonzero Vector

For any nonzero vector $|\chi\rangle$ with finite norm,

$$\|\chi\| = \sqrt{\langle\chi|\chi\rangle}>0.$$

A normalized representative is

$$|\psi\rangle = \frac{|\chi\rangle}{\|\chi\|}.$$

Indeed,

$$\langle\psi|\psi\rangle = \frac{\langle\chi|\chi\rangle} {\|\chi\|^2} =1.$$

The zero vector cannot be normalized because its norm vanishes. It represents no physical pure state: every Born weight computed from it is zero, and it does not define a ray.

Normalization also fails when $\|\chi\|=\infty$. Such an expression may be a generalized eigenfunction, a useful asymptotic solution, or simply an inadmissible candidate state. Dividing it by an informal “infinite norm” does not create a Hilbert-space vector.

Rays, Scale, and Global Phase

A physical pure state is a ray rather than one preferred vector. If

$$|\chi'\rangle = \lambda|\chi\rangle, \qquad \lambda\ne0,$$

then the two vectors define the same ray. Normalizing each gives

$$\frac{|\chi'\rangle}{\|\chi'\|} = \frac{\lambda}{|\lambda|} \frac{|\chi\rangle}{\|\chi\|}.$$

The magnitude $|\lambda|$ disappears, while its phase remains as a global phase. Therefore normalization selects the unit sphere inside the nonzero vectors but does not select one unique vector on each ray.

The sequence of quotients is conceptually useful:

$$\begin{aligned} \text{nonzero vectors} &\longrightarrow \text{unit vectors} \\ &\longrightarrow \text{physical rays}, \end{aligned}$$

where the first step removes positive scale and the second identifies constant phases. The geometry of the final quotient belongs to Projective Hilbert Space.

Finite-Dimensional States

In an orthonormal basis $\{|n\rangle\}_{n=1}^d$, write

$$|\chi\rangle = \sum_{n=1}^d a_n|n\rangle.$$

Orthonormality gives

$$\|\chi\|^2 = \sum_{n=1}^d|a_n|^2.$$

If the sum is nonzero, normalized coefficients are

$$c_n = \frac{a_n} {\sqrt{\sum_m|a_m|^2}}.$$

For example,

$$|\chi\rangle = 2|0\rangle-i|1\rangle$$

has squared norm $5$, so

$$|\psi\rangle = \frac{2}{\sqrt5}|0\rangle - \frac{i}{\sqrt5}|1\rangle.$$

The basis probabilities are $4/5$ and $1/5$. The relative phase $-i$ does not affect this particular measurement but can affect measurements in another basis.

Nonorthogonal Coordinates and the Gram Matrix

The coefficient-sum rule is not valid in an arbitrary basis. Let $\{|b_i\rangle\}$ be a nonorthogonal basis and

$$|\chi\rangle = \sum_i c_i|b_i\rangle.$$

Define the Gram matrix

$$G_{ij}=\langle b_i|b_j\rangle.$$

Then

$$\|\chi\|^2 = \sum_{i,j}c_i^*G_{ij}c_j = c^\dagger Gc.$$

The normalized coordinate column is therefore

$$c_{\mathrm{norm}} = \frac{c}{\sqrt{c^\dagger Gc}}.$$

For linearly independent basis vectors, $G$ is positive definite, so $c^\dagger Gc>0$ for every nonzero column. If the spanning set is linearly dependent, $G$ is singular and coordinate descriptions are not unique; one must first identify the actual vector or remove redundancy.

This example emphasizes the invariant rule: normalize using the Hilbert-space inner product, not by applying a remembered component formula outside its assumptions.

Infinite Discrete Expansions

For a countable orthonormal basis,

$$|\chi\rangle = \sum_{n=1}^{\infty}a_n|n\rangle,$$

the vector belongs to the Hilbert space only if

$$\sum_{n=1}^{\infty}|a_n|^2<\infty.$$

It can be normalized when this sum is positive and finite. The sequence $a_n=1/\sqrt n$, for example, does not define an element of $\ell^2$ because

$$\sum_{n=1}^{\infty}\frac1n$$

diverges. An expression may be a formally meaningful series without being a normalizable state.

For a normalized state, truncating after $N$ terms gives the subnormalized vector

$$|\psi_N\rangle = \sum_{n=1}^{N}c_n|n\rangle$$

with retained weight

$$w_N = \langle\psi_N|\psi_N\rangle = \sum_{n=1}^{N}|c_n|^2.$$

If $w_N>0$, the normalized truncation is

$$|\widehat\psi_N\rangle = \frac{|\psi_N\rangle}{\sqrt{w_N}}.$$

Renormalizing makes a valid state inside the truncated subspace, but it does not erase truncation error. The discarded probability weight is $1-w_N$, and the fidelity with the original pure state is $w_N$.

Position-Space Wavefunctions

For a particle on the line, normalization is

$$\int_{-\infty}^{\infty} dx\,|\psi(x)|^2=1.$$

If

$$\psi(x)=A f(x),$$

then

$$|A| = \left( \int dx\,|f(x)|^2 \right)^{-1/2},$$

provided the integral is finite and nonzero. Normalization fixes $|A|$ but not the constant phase of $A$.

For example, with $\alpha>0$,

$$\psi(x)=Ae^{-\alpha x^2/2}$$

has

$$\int dx\,e^{-\alpha x^2} = \sqrt{\frac\pi\alpha},$$

so

$$|A| = \left(\frac\alpha\pi\right)^{1/4}.$$

The function $e^{+\alpha x^2/2}$ cannot be normalized on the line because its squared modulus is not integrable. No finite constant $A\ne0$ repairs that failure.

Square Integrability and Almost-Everywhere Equality

Normalizable position wavefunctions live in an $L^2$ space. The condition

$$\int dx\,|\psi(x)|^2<\infty$$

states square integrability, while equality to one selects a unit-norm representative.

An $L^2$ vector is an equivalence class of functions that agree almost everywhere. Changing a wavefunction at finitely many points does not alter its norm or any interval probability. Conversely, pointwise decay is not by itself enough to guarantee square integrability; the rate and measure matter.

For a power-law tail

$$|\psi(x)|\sim |x|^{-s}$$

on the line, the tail contribution behaves as

$$\int_R^\infty dx\,x^{-2s},$$

which converges only for $s>1/2$. Local singularities require a separate integrability check. The mathematical structure is developed in L2 Spaces.

Measures, Coordinates, and Units

Normalization belongs to a function and a measure together. In coordinates $q=(q^1,\ldots,q^d)$, one may have

$$1 = \int dq\,J(q)|\psi(q)|^2,$$

where $J(q)$ is a Jacobian density. In spherical coordinates,

$$1 = \int_0^\infty dr \int_0^\pi d\theta \times \int_0^{2\pi}d\varphi\, r^2\sin\theta |\psi(r,\theta,\varphi)|^2.$$

The radial factor and angular Jacobian are part of the probability measure. Dropping them changes the norm.

If $\psi$ is normalized with respect to $d^dq$, then

$$[\psi]=[q]^{-d/2}$$

when all coordinates carry the same dimension. For a one-particle position wavefunction in $d$ spatial dimensions,

$$[\psi]=L^{-d/2}.$$

The abstract ket has no position-space units of this kind; the units belong to its coordinate representative and chosen continuum normalization.

One can absorb the Jacobian into a redefined function:

$$\widetilde\psi(q) = J(q)^{1/2}\psi(q),$$

so that

$$1=\int dq\,|\widetilde\psi(q)|^2.$$

The represented operators must be transformed at the same time.

Momentum-Space Normalization

With the symmetric Fourier convention,

$$\phi(p) = \frac{1}{\sqrt{2\pi\hbar}} \int dx\,e^{-ipx/\hbar}\psi(x),$$

and a normalized state satisfies

$$\int dp\,|\phi(p)|^2=1.$$

The equality

$$\int dx\,|\psi(x)|^2 = \int dp\,|\phi(p)|^2$$

is norm preservation under a unitary Fourier transform. It is not an independent normalization condition on a second state.

The momentum wavefunction has units

$$[\phi(p)]=[p]^{-1/2}$$

in one dimension. Relabelling momentum by $p=\hbar k$ requires

$$\widetilde\psi(k) = \sqrt\hbar\,\phi(\hbar k)$$

so that

$$|\widetilde\psi(k)|^2dk = |\phi(p)|^2dp.$$

The Jacobian is part of normalization. The full representation dictionary is in Momentum-Space Representation.

Multi-Component Wavefunctions

For a wavefunction with a discrete internal label $s$, normalization sums over that label and integrates the continuous variables:

$$\sum_s \int dx\,|\psi_s(x)|^2=1.$$

For a spin-$1/2$ particle,

$$\Psi(x) = \begin{pmatrix} \psi_\uparrow(x)\\ \psi_\downarrow(x) \end{pmatrix},$$

so

$$\begin{aligned} \int dx\, \Psi(x)^\dagger\Psi(x) &= \int dx\, |\psi_\uparrow(x)|^2 + \int dx\, |\psi_\downarrow(x)|^2 \\ &=1. \end{aligned}$$

The individual components need not each have norm one. Their squared norms are the probabilities of the corresponding internal outcomes when the continuous coordinate is ignored.

Many-Particle Wavefunctions

For $N$ distinguishable particles in three dimensions,

$$1 = \int \prod_{j=1}^{N}d^3r_j\, |\Psi(\mathbf r_1,\ldots,\mathbf r_N)|^2.$$

The wavefunction lives on $3N$-dimensional configuration space and has units

$$[\Psi]=L^{-3N/2}.$$

If there are spin labels, normalization also sums over every spin configuration. Exchange symmetry for identical particles restricts the allowed functions but does not change the unit-norm condition.

For a product state

$$|\Psi\rangle = |\psi\rangle\otimes|\phi\rangle,$$

the norm factors:

$$\|\Psi\|^2 = \|\psi\|^2\|\phi\|^2.$$

Normalized factors therefore give a normalized product. Entangled states are normalized by their full joint norm, not by trying to normalize nonexistent separate subsystem wavefunctions.

Density-Operator Normalization

A general quantum state is represented by a positive trace-one operator:

$$\rho\ge0, \qquad \operatorname{Tr}\rho=1.$$

For any complete POVM $\{E_a\}$,

$$\begin{aligned} \sum_a p(a) &= \sum_a\operatorname{Tr}(\rho E_a) \\ &= \operatorname{Tr} \left( \rho\sum_aE_a \right) \\ &= \operatorname{Tr}\rho =1. \end{aligned}$$

For a pure normalized vector,

$$\rho_\psi = |\psi\rangle\langle\psi|$$

has

$$\operatorname{Tr}\rho_\psi = \langle\psi|\psi\rangle =1.$$

For an ensemble

$$\rho = \sum_j q_j |\psi_j\rangle\langle\psi_j|,$$

with normalized $|\psi_j\rangle$, trace normalization requires

$$q_j\ge0, \qquad \sum_jq_j=1.$$

This decomposition is not unique, but $\operatorname{Tr}\rho=1$ is representation independent. The state concept and positivity requirements are canonical in Density Operators.

Subnormalized Conditional States

Intermediate states need not always carry total weight one. Suppose a measurement outcome $a$ is represented by an operation that produces

$$\widetilde\rho_a = M_a\rho M_a^\dagger.$$

Its trace is the probability of that branch:

$$p(a) = \operatorname{Tr}\widetilde\rho_a \le1.$$

The tilde signals that the operator is subnormalized. Conditional on the outcome occurring and $p(a)>0$, the normalized state is

$$\rho_a = \frac{\widetilde\rho_a}{p(a)}.$$

For a pure-state projection,

$$|\widetilde\psi_a\rangle = P_a|\psi\rangle$$

has squared norm

$$\langle\widetilde\psi_a |\widetilde\psi_a\rangle = \langle\psi|P_a|\psi\rangle = p(a),$$

and the conditional unit vector is

$$|\psi_a\rangle = \frac{P_a|\psi\rangle}{\sqrt{p(a)}}.$$

One must not renormalize a collection of branches before recording their traces, because those traces are the outcome probabilities. The dynamical and interpretive statement belongs to State Update Rule.

Normalization Under Unitary Evolution

Closed-system time evolution is unitary:

$$|\psi(t)\rangle = U(t,t_0)|\psi(t_0)\rangle, \qquad U^\dagger U=I.$$

Therefore

$$\begin{aligned} \langle\psi(t)|\psi(t)\rangle &= \langle\psi(t_0)| U^\dagger U |\psi(t_0)\rangle \\ &= \langle\psi(t_0)|\psi(t_0)\rangle. \end{aligned}$$

A normalized initial state remains normalized. In differential form, for a self-adjoint Hamiltonian and a state in the relevant domains,

$$\begin{aligned} \frac{d}{dt}\langle\psi|\psi\rangle &= \langle\dot\psi|\psi\rangle + \langle\psi|\dot\psi\rangle \\ &= \frac{i}{\hbar} \langle\psi|H|\psi\rangle - \frac{i}{\hbar} \langle\psi|H|\psi\rangle \\ &=0. \end{aligned}$$

Repeatedly renormalizing an exact closed-system solution should be unnecessary. In numerical work, norm drift is instead a diagnostic of discretization, solver, or implementation error unless the effective dynamics is intentionally nonunitary.

The structural reason for preservation is explained in Unitary Time Evolution.

Delta Normalization

Continuous spectral bases are commonly normalized by a Dirac delta:

$$\langle a|a'\rangle = \delta(a-a').$$

Position and momentum examples are

$$\langle x|x'\rangle = \delta(x-x'), \qquad \langle p|p'\rangle = \delta(p-p').$$

This is not unit normalization. The generalized ket $|a\rangle$ is not normally an element of the physical Hilbert space, and $\delta(0)$ is not a large finite norm. The delta is a distribution specifying how continuum basis objects pair under integrals.

The identity resolution

$$I = \int da\,|a\rangle\langle a|$$

allows a normalizable state to be written

$$|\psi\rangle = \int da\,\psi(a)|a\rangle,$$

with ordinary state normalization

$$\int da\,|\psi(a)|^2=1.$$

The generalized basis kets are delta normalized; the coefficient function of a physical state is square normalized. Confusing these two levels leads to illegal expressions such as $|\delta(a-a_0)|^2$ as an ordinary probability density.

Relabelling a Continuum Basis

Delta normalization depends on the spectral label. Let $b=f(a)$ be a smooth one-to-one relabelling with nonzero derivative on the branch under consideration. Since

$$\delta(a-a') = \left|\frac{db}{da}\right| \delta(b-b'),$$

normalized generalized bases are related by

$$|b\rangle = \left|\frac{da}{db}\right|^{1/2} |a(b)\rangle.$$

Their coefficient functions obey

$$\psi_b(b) = \left|\frac{da}{db}\right|^{1/2} \psi_a(a(b)).$$

Consequently,

$$|\psi_b(b)|^2\,|db| = |\psi_a(a)|^2\,|da|.$$

The absolute differentials emphasize that probability uses a positive measure, even for a decreasing coordinate map. This square-root Jacobian is the continuum analogue of a basis normalization factor. It appears when changing from momentum to wave number, from momentum to energy on a fixed branch, or between other spectral coordinates.

When $f$ is not one-to-one, each branch or degeneracy label must be retained. For a free particle, the energy $E=p^2/(2m)$ does not distinguish $+p$ from $-p$, so an energy representation requires an additional channel label.

Worked example: energy and two momentum directions

For a free particle on the full line, set $p_E=\sqrt{2mE}$ for $E>0$ and retain the direction $\sigma=\pm$. Starting with momentum kets normalized by $\langle p|p'\rangle=\delta(p-p')$, define

$$|E,\sigma\rangle=\sqrt{\frac{m}{p_E}}\,|\sigma p_E\rangle, \qquad \chi_\sigma(E)=\sqrt{\frac{m}{p_E}}\,\phi(\sigma p_E).$$

The square-root Jacobian gives $\langle E,\sigma|E',\sigma'\rangle=\delta_{\sigma\sigma'}\delta(E-E')$. Splitting the momentum integral into its positive and negative halves verifies

$$\sum_{\sigma=\pm}\int_0^\infty dE\,|\chi_\sigma(E)|^2 =\int_{-\infty}^\infty dp\,|\phi(p)|^2=1.$$

An energy measurement that does not resolve direction therefore has density

$$w_E(E)=\frac{m}{p_E} \left[|\phi(p_E)|^2+|\phi(-p_E)|^2\right].$$

The two probabilities add because the directions are orthogonal channels. The state remains the pair $(\chi_+,\chi_-)$: replacing that pair by $\chi_++\chi_-$ introduces an interference term belonging to a different measurement. Ignoring direction in the recorded energy does not, by itself, destroy coherence between directions.

For a concrete asymmetric packet, take the real momentum amplitude

$$\phi(p)=\frac{1}{(2\pi s^2)^{1/4}} \exp\!\left[-\frac{(p-s)^2}{4s^2}\right], \qquad s>0.$$

Its mean momentum and momentum standard deviation both equal $s$. With $E_*=s^2/(2m)$ and $\varepsilon=E/E_*$, the energy density becomes

$$E_*w_E(E)= \frac{e^{-(\sqrt\varepsilon-1)^2/2} +e^{-(\sqrt\varepsilon+1)^2/2}} {2\sqrt{2\pi\varepsilon}}.$$

This density diverges as $E^{-1/2}$ at the threshold, but the divergence is integrable: the state has no point probability at $E=0$. For the finite bin $0<E\le E_*$, the corresponding momentum interval is $-s\le p\le s$, so

$$\begin{aligned} \Pr(0<E\le E_*)&=\int_{-s}^{s}dp\,|\phi(p)|^2 =\frac12\operatorname{erf}(\sqrt2) \simeq0.477250,\\ \Pr(0<p\le s)&\simeq0.341345,\\ \Pr(-s\le p<0)&\simeq0.135905. \end{aligned}$$

Both channels contribute, with unequal weights. The first moment provides an independent check: $\langle E\rangle=\langle p^2\rangle/(2m)=s^2/m=2E_*$. These are amplitudes normalized with respect to $dE$; the separate convention for incident flux is explained in Normalization Conventions.

Box Normalization

Box normalization replaces a continuum by a finite region with specified boundary conditions. On a periodic interval of length $L$,

$$\psi_n(x) = \frac1{\sqrt L} e^{ik_nx}, \qquad k_n=\frac{2\pi n}{L}.$$

These modes satisfy

$$\int_0^Ldx\, \psi_m(x)^*\psi_n(x) = \delta_{mn}.$$

The finite-volume basis is genuinely orthonormal within the periodic Hilbert space. It is not the same object as a delta-normalized plane wave on the full line.

With

$$p_n=\hbar k_n, \qquad \Delta p=\frac{2\pi\hbar}{L},$$

the large-box relation between a continuum momentum amplitude and discrete coefficients is

$$c_n \simeq \sqrt{\Delta p}\,\phi(p_n).$$

Then

$$\sum_n|c_n|^2 \longrightarrow \int dp\,|\phi(p)|^2.$$

At the basis level,

$$|p_n\rangle_{\mathrm{box}} \simeq \sqrt{\Delta p}\,|p_n\rangle_{\delta}.$$

The factor relates the Kronecker-normalized discrete ket to the Dirac-delta-normalized continuum ket. Densities of states arise because sums carry one state per spacing $\Delta p$.

Boundary Conditions Are Part of the Convention

The phrase “normalize in a box” is incomplete without boundary conditions. Periodic boundary conditions yield plane-wave momentum modes. Hard-wall Dirichlet conditions yield standing waves instead. Other self-adjoint boundary conditions can shift or reorganize the spectrum.

The factor $1/\sqrt L$ normalizes a periodic plane wave because its modulus is constant on an interval of length $L$. It does not imply that the same plane wave is a physical eigenstate for every finite interval problem.

Box normalization is useful for:

• turning integrals into sums;
• defining finite-volume numerical bases;
• counting modes and deriving densities of states;
• regulating continuum expressions;
• connecting Kronecker and Dirac deltas.

The artificial volume should disappear from physical continuum predictions after sums, amplitudes, and state densities are converted consistently.

Flux Normalization

Scattering calculations often use stationary states that are not square normalizable. One may choose their amplitude so the incident probability current has a prescribed value, commonly unit flux. This is flux normalization, not total-probability normalization.

For a one-dimensional plane wave

$$\psi(x)=Ae^{ikx}, \qquad p=\hbar k,$$

the probability current is

$$j = \frac{\hbar k}{m}|A|^2 = \frac{p}{m}|A|^2.$$

Choosing unit incident flux would require

$$|A|^2=\frac{m}{|p|}$$

for the magnitude convention $|j|=1$, up to any additional continuum or channel normalization factors. The dimensions of such a mode differ from those of a unit-normalized bound-state wavefunction.

Flux-normalized states are designed so reflected, transmitted, or scattered flux ratios yield probabilities or cross sections. The detailed convention depends on dimension, channels, relativistic versus nonrelativistic kinematics, and the definition of the scattering matrix. Those applications belong to the scattering volume and Normalization Conventions.

Non-Normalizable Ideals and Limits

Standard non-normalizable objects include:

• exact position or momentum eigenkets;
• plane waves on all of space;
• energy-normalized scattering eigenfunctions;
• an unregularized Dirac delta used as a putative state;
• formal solutions with nonintegrable growth or tails.

They can still be useful as generalized basis vectors, asymptotic modes, Green function sources, or limits of normalized wave packets. Their usefulness does not make them ordinary physical states.

For a narrowing Gaussian of width $\sigma$, the probability measures $|\psi_\sigma(x)|^2dx$ converge weakly to a point mass: integrating a continuous bounded test function against the density approaches its value at the packet center. The wavefunctions themselves do not converge to a delta distribution. Their distributional limit is zero, even though every one has L2 norm one; no strong L2 limit exists. The exercise below computes the two limits explicitly.

Likewise, a unit-normalized packet spread over an increasingly large region has a decreasing local amplitude. It does not approach a nonzero plane wave pointwise. A plane-wave limit requires a compensating amplitude rescaling, which abandons unit normalization in favor of a continuum convention.

Conditional Normalization in Subregions

For an ideal Lüders measurement of whether position lies in a region $\Delta$, condition on the successful outcome. The unnormalized projected wavefunction is

$$\widetilde\psi_\Delta(x) = \mathbf 1_\Delta(x)\psi(x),$$

where $\mathbf 1_\Delta$ is the indicator function. Its squared norm is

$$p_\Delta = \int_\Delta dx\,|\psi(x)|^2.$$

If $p_\Delta>0$, the conditional normalized wavefunction is

$$\psi_\Delta(x) = \frac{\mathbf 1_\Delta(x)\psi(x)} {\sqrt{p_\Delta}}.$$

This renormalization changes the state because it conditions on new information or a measurement event. It should not be confused with merely choosing a unit representative of an unchanged ray.

A sharp interval cutoff can introduce jumps at the endpoints. The new state is still normalizable, but need not have finite kinetic energy or belong to the momentum domain. A finite-resolution detector requires a smoother measurement model.

If $p_\Delta=0$, no conditional state for that event is defined by this formula. Division by zero is not repaired by assigning an arbitrary vector to an impossible branch.

Numerical Normalization

On a grid $x_j$ with quadrature weights $w_j$, the continuum norm is approximated by

$$\|\psi\|^2 \approx \sum_j w_j|\psi_j|^2.$$

For a uniform grid with spacing $\Delta x$,

$$\|\psi\|^2 \approx \Delta x\sum_j|\psi_j|^2.$$

The raw Euclidean norm $\sum_j|\psi_j|^2=1$ is therefore not generally the continuum normalization unless the stored array has already absorbed $\sqrt{\Delta x}$.

For a nonorthogonal finite basis with overlap matrix $S$, a coefficient vector $c$ is normalized by

$$c^\dagger Sc=1.$$

In generalized eigenvalue problems,

$$Hc=ESc,$$

the same overlap matrix defines the physical inner product. Normalizing with $c^\dagger c=1$ instead can produce basis-dependent errors.

Useful numerical checks include:

1. include quadrature or overlap weights;
2. verify that the norm is real and nonnegative within tolerance;
3. reject zero or nearly zero vectors before division;
4. track discarded weight after truncation;
5. monitor norm drift during nominally unitary evolution;
6. distinguish deliberate subnormalization from numerical loss;
7. avoid hiding large errors by renormalizing at every step.

Normalization Does Not Guarantee Physical Admissibility

Unit norm is necessary for an ordinary state vector, but it is not sufficient for every physical question. A normalized wavefunction may fail to lie in the domain of an unbounded observable. For example, it may have finite norm but infinite kinetic-energy expectation or undefined boundary derivatives.

Likewise, a trace-one operator is not a state unless it is also positive. The matrix

$$\begin{pmatrix} 2&0\\ 0&-1 \end{pmatrix}$$

has trace one but is not positive and therefore is not a density operator.

Normalization controls total weight. Positivity, domains, symmetry constraints, boundary conditions, and finite expectation values impose additional requirements depending on the system and observable.

Common Mistakes

• Forgetting to normalize an ordinary state before applying the simplest Born formulas.
• Trying to normalize the zero vector or a vector of infinite norm.
• Thinking normalization removes global phase or chooses one unique vector on a ray.
• Using $\sum_i|c_i|^2$ in a nonorthogonal basis instead of $c^\dagger Gc$.
• Dropping coordinate Jacobians, spin sums, particle coordinates, quadrature weights, or overlap matrices.
• Calling $|\psi(x)|^2$ a probability instead of a density with respect to a measure.
• Treating delta normalization as ordinary unit normalization.
• Squaring a Dirac delta as though it were an ordinary function.
• Mixing box-normalized, delta-normalized, and flux-normalized amplitudes.
• Renormalizing a conditional branch before recording its probability.
• Assuming trace one is enough for a density operator without checking positivity.
• Renormalizing every numerical time step and thereby hiding nonunitary error.
• Assuming any unit-norm wavefunction lies in the domain of every observable.

Continue with applications

• Born Rule for Continuous Spectra connects densities to spectral projectors.
• Generalized Eigenvectors explains the distributional meaning of continuum kets.
• Normalization Conventions applies the conventions to bound, scattering, radial, and grid states.
• State Update Rule explains conditioning on measurement outcomes.

Exercises

1. A nonorthogonal two-vector basis

Let

$$|b_1\rangle=|0\rangle, \qquad |b_2\rangle = \frac{|0\rangle+|1\rangle}{\sqrt2},$$

where $|0\rangle,|1\rangle$ are orthonormal. Normalize $|\chi\rangle=|b_1\rangle+|b_2\rangle$ using the Gram matrix.

Solution

The Gram matrix is

$$G = \begin{pmatrix} 1&1/\sqrt2\\ 1/\sqrt2&1 \end{pmatrix},$$

and the coordinate column is

$$c = \begin{pmatrix} 1\\ 1 \end{pmatrix}.$$

Therefore

$$\|\chi\|^2 = c^\dagger Gc = 2+\sqrt2.$$

A normalized vector is

$$|\psi\rangle = \frac{|b_1\rangle+|b_2\rangle} {\sqrt{2+\sqrt2}}.$$

The naive coefficient sum would give $2$ and would miss the nonzero overlap.

2. A normalizable power-law state

Let

$$\psi(x) = \frac{A}{1+x^2/a^2}, \qquad a>0.$$

Find $|A|$ on the real line.

Solution

Set $u=x/a$. Then

$$\begin{aligned} 1 &= |A|^2a \int_{-\infty}^{\infty} \frac{du}{(1+u^2)^2} \\ &= |A|^2a\frac\pi2. \end{aligned}$$

Thus

$$|A| = \sqrt{\frac{2}{\pi a}}.$$

Its phase remains arbitrary.

3. A narrowing Gaussian: state versus density

Let $\psi_\sigma(x)=(2\pi\sigma^2)^{-1/4}e^{-x^2/(4\sigma^2)}$ with $\sigma>0$. For a smooth compactly supported test function $f$, compare $\int f(x)\psi_\sigma(x)\,dx$ with $\int f(x)|\psi_\sigma(x)|^2\,dx$ as $\sigma\to0$. Does the state have a strong L2 limit?

Solution

Using $x=\sigma u$ gives

$$\int f(x)\psi_\sigma(x)\,dx =\frac{\sqrt\sigma}{(2\pi)^{1/4}} \int f(\sigma u)e^{-u^2/4}\,du\longrightarrow0.$$

The boundedness of $f$ and integrability of the Gaussian bound the remaining integral uniformly. For the density,

$$\int f(x)|\psi_\sigma(x)|^2\,dx =\frac1{\sqrt{2\pi}}\int f(\sigma u)e^{-u^2/2}\,du \longrightarrow f(0).$$

Thus the density measures approach a point mass, whereas the state amplitudes approach zero as distributions. A strong L2 limit would also be the zero distribution, but norm convergence to zero contradicts $\|\psi_\sigma\|_2=1$. There is no strong L2 limit and no normalized position eigenvector.

4. Truncation and renormalization

Let a normalized state have orthonormal-basis coefficients $c_n$. Define

$$|\psi_N\rangle = \sum_{n=1}^{N}c_n|n\rangle$$

and $w_N=\sum_{n=1}^{N}|c_n|^2>0$. Normalize the truncation and find its fidelity with the original state.

Solution

The normalized truncation is

$$|\widehat\psi_N\rangle = \frac1{\sqrt{w_N}} \sum_{n=1}^{N}c_n|n\rangle.$$

Its overlap with the full state is

$$\begin{aligned} \langle\psi|\widehat\psi_N\rangle &= \frac1{\sqrt{w_N}} \sum_{n=1}^{N}|c_n|^2 \\ &= \sqrt{w_N}. \end{aligned}$$

Therefore the pure-state fidelity is

$$|\langle\psi|\widehat\psi_N\rangle|^2 = w_N.$$

Renormalization does not restore the discarded weight $1-w_N$.

5. A subnormalized measurement branch

A normalized qubit is

$$|\psi\rangle = \alpha|0\rangle+\beta|1\rangle.$$

Project onto $|0\rangle$. Find the unnormalized branch, its norm, and the conditional state when the outcome has nonzero probability.

Solution

With $P_0=|0\rangle\langle0|$,

$$|\widetilde\psi_0\rangle = P_0|\psi\rangle = \alpha|0\rangle.$$

Its squared norm is

$$\langle\widetilde\psi_0 |\widetilde\psi_0\rangle = |\alpha|^2,$$

which is the outcome probability. If $\alpha\ne0$, the conditional state is

$$|\psi_0\rangle = \frac{\alpha}{|\alpha|}|0\rangle,$$

which represents the same ray as $|0\rangle$. If $\alpha=0$, the branch is impossible and the division is undefined.

6. Continuum relabelling

Suppose $E=p^2/(2m)$ with $p>0$, and $\phi(p)$ is normalized on the positive momentum half-line. Find the energy amplitude $\chi(E)$ normalized with respect to $dE$.

Then use the full-line Gaussian in the worked example to test the incorrect scalar amplitude $\chi_++\chi_-$. What is its squared norm? Contrast it with the properly normalized change of channel basis $\chi_{\mathrm{even,odd}}=(\chi_+\pm\chi_-)/\sqrt2$, which resolves spatial parity as well as energy. Find the total probabilities of even and odd parity.

Solution

On the positive branch,

$$p(E)=\sqrt{2mE}, \qquad \frac{dp}{dE} = \frac{m}{p(E)}.$$

Probability invariance requires

$$|\chi(E)|^2dE = |\phi(p)|^2dp.$$

Therefore

$$\chi(E) = \sqrt{\frac{m}{p(E)}} \phi\bigl(p(E)\bigr)$$

up to a phase convention. If both signs of momentum are present, energy alone is not a complete label; one must retain a direction or channel index.

For the stated real Gaussian,

$$\begin{aligned} \int_0^\infty dE\,\chi_+(E)^*\chi_-(E) &=\int_0^\infty dp\,\phi(p)\phi(-p)\\ &=\frac12 e^{-1/2}. \end{aligned}$$

Consequently the incorrect scalar amplitude has squared norm $1+e^{-1/2}\simeq1.606531$, rather than one. By contrast, the even and odd channels form an orthonormal basis within each degenerate energy space. Their probabilities are

$$P_{\mathrm{even}}=\frac{1+e^{-1/2}}2\simeq0.803265, \qquad P_{\mathrm{odd}}=\frac{1-e^{-1/2}}2\simeq0.196735.$$

They sum to one. Parity tests a coherent combination of the directions; energy alone sums over the two channels in either basis.

7. Norm conservation from the Schrödinger equation

Assume $H=H^\dagger$ and that all domain conditions needed below hold. Starting from

$$i\hbar|\dot\psi\rangle=H|\psi\rangle,$$

show that $d\langle\psi|\psi\rangle/dt=0$.

Solution

The Schrödinger equation and its adjoint give

$$|\dot\psi\rangle = -\frac{i}{\hbar}H|\psi\rangle, \qquad \langle\dot\psi| = \frac{i}{\hbar}\langle\psi|H.$$

Hence

$$\begin{aligned} \frac{d}{dt}\langle\psi|\psi\rangle &= \langle\dot\psi|\psi\rangle + \langle\psi|\dot\psi\rangle \\ &= \frac{i}{\hbar}\langle\psi|H|\psi\rangle - \frac{i}{\hbar}\langle\psi|H|\psi\rangle \\ &=0. \end{aligned}$$

Self-adjointness makes the bra equation use the same $H$.

8. Trace one is not enough

Consider

$$A = \begin{pmatrix} 2&0\\ 0&-1 \end{pmatrix}.$$

Show that $A$ has unit trace but cannot be a density operator. Give an effect for which the Born expression is negative.

Solution

The trace is

$$\operatorname{Tr}A = 2-1 =1.$$

However, with the positive rank-one effect

$$E=|1\rangle\langle1|,$$

one obtains

$$\operatorname{Tr}(AE) = \langle1|A|1\rangle = -1.$$

A probability cannot be negative. The failure is that $A$ is not positive, so trace normalization alone does not make it a state.

References

• P. A. M. Dirac, The Principles of Quantum Mechanics, 4th ed., Oxford University Press, 1958, Chapters II–III. Continuum normalization conventions.
• B. C. Hall, Quantum Theory for Mathematicians, Springer, 2013. Hilbert-space states, distributions, and self-adjoint domains.
• J. J. Sakurai and J. Napolitano, Modern Quantum Mechanics, 3rd ed., Cambridge University Press, 2020, Chapters 1–2. State and scattering normalization.
• R. Shankar, Principles of Quantum Mechanics, 2nd ed., Springer, 1994, Chapters 1, 4–5. Normalization, representations, and wave packets.