Inner Products

An inner product is a scalar-valued pairing that equips a vector space with length, orthogonality, and geometry. In quantum mechanics, the same pairing produces probability amplitudes, expectation values, matrix elements, and projection coefficients.

The inner product is mathematical structure added to a vector space; it is not part of the vector-space axioms. Different inner products on the same vector space can produce different notions of normalization and orthogonality. The associated general language of length, distance, and convergence is developed in Norms and Metrics.

Why the Pairing Matters

An inner product turns algebraic questions into geometric ones:

• it measures the norm of a vector;
• it decides whether vectors are orthogonal;
• it selects orthogonal projections and best approximations;
• it identifies vectors with continuous dual functionals;
• it determines which linear maps are adjoints or unitary;
• it converts state vectors into amplitudes and matrix elements.

Bare linear independence does not require an inner product. Neither do span, dimension, or linearity. Orthogonality, adjoints, and unitarity do.

Definition and Convention

Let $V$ be a complex vector space. An inner product is a map

$$\langle\cdot\vert\cdot\rangle : V\times V\longrightarrow\mathbb C$$

with three defining properties.

Linearity in the second slot:

$$\langle u\vert av+bw\rangle = a\langle u\vert v\rangle +b\langle u\vert w\rangle.$$

Conjugate symmetry:

$$\langle u\vert v\rangle = \langle v\vert u\rangle^*.$$

Positive definiteness:

$$\langle v\vert v\rangle\geq0, \qquad \langle v\vert v\rangle=0 \Longleftrightarrow v=0.$$

Conjugate symmetry and second-slot linearity imply conjugate linearity in the first slot:

$$\langle au+bv\vert w\rangle = a^*\langle u\vert w\rangle +b^*\langle v\vert w\rangle.$$

This is the physics convention, used throughout these pages. Many mathematics texts choose the opposite convention, linear in the first slot and conjugate-linear in the second. Both are consistent. When combining sources, identify the convention before moving scalars through brackets or defining an adjoint.

On a real vector space, complex conjugation disappears. A real inner product is a symmetric positive-definite bilinear form.

Why Conjugation Is Necessary

The ordinary bilinear expression

$$u^{\mathsf T}v$$

is not a complex inner product. For example, if $u=(1,i)^{\mathsf T}$, then

$$u^{\mathsf T}u = 1+i^2 =0$$

even though $u\ne0$. The standard complex pairing uses the conjugate transpose:

$$u^\dagger u = \lvert1\rvert^2+\lvert i\rvert^2 =2.$$

Conjugation is what makes the self-pairing real and positive while preserving complex linearity in one slot.

Standard Finite-Dimensional Inner Product

For $u,v\in\mathbb C^n$, the standard inner product is

$$\langle u\vert v\rangle = u^\dagger v = \sum_{j=1}^{n}u_j^*v_j.$$

The corresponding norm is

$$\lVert v\rVert = \sqrt{\langle v\vert v\rangle} = \left( \sum_{j=1}^{n}\lvert v_j\rvert^2 \right)^{1/2}.$$

The component formula assumes an orthonormal coordinate basis. In a general basis, the coordinate dot product is replaced by a Gram matrix.

Gram Matrices and Weighted Pairings

Let $(e_1,\ldots,e_n)$ be any basis of a finite-dimensional inner-product space. Its Gram matrix is

$$G_{ij} = \langle e_i\vert e_j\rangle.$$

If $u$ and $v$ have coordinate columns $c$ and $d$ in this basis, then

$$\langle u\vert v\rangle = c^\dagger Gd.$$

The matrix $G$ is Hermitian and positive definite:

$$G^\dagger=G, \qquad c^\dagger Gc>0 \quad\text{for }c\ne0.$$

Conversely, every Hermitian positive-definite matrix $G$ defines an inner product on $\mathbb C^n$ by

$$\langle u\vert v\rangle_G = u^\dagger Gv.$$

This is not a cosmetic generalization. Nonorthogonal basis functions, curvilinear-coordinate discretizations, and generalized eigenvalue problems all produce overlap or mass matrices that play the role of $G$.

Function-Space Inner Products

For complex-valued functions on a measure space $X$, a common pairing is

$$\langle f\vert g\rangle = \int_X f(x)^*g(x)\,d\mu(x).$$

With a positive weight $w(x)$, one may instead use

$$\langle f\vert g\rangle_w = \int_X f(x)^*g(x)w(x)\,d\mu(x).$$

The weight and measure are part of the definition. Radial wavefunctions, spherical harmonics, orthogonal polynomials, and numerical quadrature all use nontrivial measures or weights. A function family orthogonal for one weight need not be orthogonal for another.

Square-integrable functions are identified when they differ only on a measure-zero set. After that quotient is taken, the pairing is positive definite on $L^2$ Spaces.

Another useful example is the Hilbert–Schmidt pairing of finite matrices:

$$\langle A\vert B\rangle_{\mathrm{HS}} = \operatorname{Tr}(A^\dagger B).$$

It treats matrices themselves as vectors and appears in operator expansions, density-operator calculations, and numerical error measures.

Norm and Distance

Every inner product induces a norm,

$$\lVert v\rVert = \sqrt{\langle v\vert v\rangle},$$

and hence a distance,

$$d(u,v)=\lVert u-v\rVert.$$

Not every norm arises from an inner product. Inner-product norms satisfy the parallelogram identity, while a general norm need not. The canonical treatment of norms, metrics, convergence, and completeness is Norms and Metrics.

Orthogonality

Vectors $u$ and $v$ are orthogonal when

$$\langle u\vert v\rangle=0.$$

Conjugate symmetry makes orthogonality symmetric. For orthogonal vectors,

$$\lVert u+v\rVert^2 = \lVert u\rVert^2+\lVert v\rVert^2,$$

the abstract Pythagorean theorem.

For a subset $S\subseteq V$, its orthogonal complement is

$$S^\perp = \lbrace v\in V: \langle s\vert v\rangle=0 \text{ for every }s\in S\rbrace.$$

An orthogonal set of nonzero vectors is automatically linearly independent. The converse is false. Orthonormal Bases develops expansions in orthogonal unit vectors.

Projection onto One Direction

Let $v\ne0$. The component of $u$ along $v$ is

$$\operatorname{proj}_v u = v\, \frac{\langle v\vert u\rangle} {\langle v\vert v\rangle}.$$

The residual

$$u_\perp = u-\operatorname{proj}_v u$$

satisfies

$$\langle v\vert u_\perp\rangle=0.$$

This coefficient formula depends on which slot is linear. Under the physics convention, the scalar multiplying $v$ is $\langle v\vert u\rangle/\langle v\vert v\rangle$. The general treatment of orthogonal projectors belongs to Projectors.

Cauchy–Schwarz Inequality

For any two vectors,

$$\lvert\langle u\vert v\rangle\rvert \leq \lVert u\rVert\,\lVert v\rVert.$$

To prove it, the zero-vector cases are immediate, so take $v\ne0$ and define

$$\alpha = \frac{\langle v\vert u\rangle} {\langle v\vert v\rangle}, \qquad w=u-\alpha v.$$

By construction, $\langle v\vert w\rangle=0$. Therefore

$$\begin{aligned} \lVert u\rVert^2 &= \lVert\alpha v+w\rVert^2 \\ &= \lvert\alpha\rvert^2\lVert v\rVert^2 +\lVert w\rVert^2 \\ &\geq \frac{ \lvert\langle v\vert u\rangle\rvert^2 }{ \lVert v\rVert^2 }. \end{aligned}$$

Multiplying by $\lVert v\rVert^2$ and taking square roots gives the inequality. Equality holds exactly when $w=0$, meaning that $u$ and $v$ are linearly dependent, including the cases where one vector is zero.

Cauchy–Schwarz controls the size of every overlap. It also proves the triangle inequality for the induced norm and bounds expectation values, correlations, and uncertainty relations.

Worked Example: A Nonorthogonal Basis

Use the standard inner product on $\mathbb C^2$, but choose the nonorthogonal basis

$$e_1= \begin{pmatrix} 1\\ 0 \end{pmatrix}, \qquad e_2= \begin{pmatrix} 1\\ 1 \end{pmatrix}.$$

Its Gram matrix is

$$G = \begin{pmatrix} 1&1\\ 1&2 \end{pmatrix}.$$

Consider the vector with basis coordinates

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

The correct squared norm is

$$\begin{aligned} \lVert v\rVert^2 &= c^\dagger Gc \\ &= \begin{pmatrix} 1&-i \end{pmatrix} \begin{pmatrix} 1&1\\ 1&2 \end{pmatrix} \begin{pmatrix} 1\\ i \end{pmatrix} \\ &=3. \end{aligned}$$

The naive sum of squared coordinate moduli gives $2$, which is wrong because the basis is not orthonormal. Indeed,

$$v=e_1+i e_2 = \begin{pmatrix} 1+i\\ i \end{pmatrix},$$

whose standard coordinate norm is

$$\lvert1+i\rvert^2+\lvert i\rvert^2=3.$$

The abstract inner product is unchanged; only its coordinate representation has acquired the Gram matrix.

Quantum Meaning

For normalized vectors $\lvert\psi\rangle$ and $\lvert\phi\rangle$, the scalar

$$\mathcal A(\phi\leftarrow\psi) = \langle\phi\vert\psi\rangle$$

is a transition amplitude for a specified rank-one outcome. Cauchy–Schwarz gives

$$0 \leq \lvert\langle\phi\vert\psi\rangle\rvert \leq 1.$$

The inner product supplies the amplitude. The Born rule supplies the physical probability

$$p(\phi\mid\psi) = \lvert\langle\phi\vert\psi\rangle\rvert^2.$$

That probability postulate does not follow from the inner-product axioms. See Probability Amplitudes for preparation, measurement context, phase, and higher-rank outcomes.

Under rephasing,

$$\begin{aligned} \lvert\psi\rangle &\longmapsto e^{i\alpha}\lvert\psi\rangle, \\ \lvert\phi\rangle &\longmapsto e^{i\beta}\lvert\phi\rangle, \end{aligned}$$

the overlap becomes

$$\langle\phi\vert\psi\rangle \longmapsto e^{i(\alpha-\beta)} \langle\phi\vert\psi\rangle.$$

Its phase changes, while its magnitude and the transition probability do not. Relative phases remain observable when several amplitudes interfere.

Inner-Product-Preserving Maps

A linear map $U$ is unitary when

$$\langle Uu\vert Uv\rangle = \langle u\vert v\rangle$$

for all $u,v$. It therefore preserves norms, orthogonality, angles, and transition probabilities. This geometric definition is equivalent in finite dimensions to

$$U^\dagger U=I.$$

Unitary Operators develops this structure and its role in changes of basis and quantum evolution.

Finite and Infinite Dimensions

In finite dimensions, every inner-product space is complete and all linear functionals are continuous. In infinite dimensions, these conclusions separate:

• an inner-product space need not contain the limits of all its Cauchy sequences;
• completing it produces a Hilbert space;
• the Riesz theorem identifies vectors with continuous dual functionals, not the full algebraic dual;
• an integral pairing is defined only for functions in its domain;
• generalized position and momentum eigenkets are not normalizable Hilbert-space vectors.

The notation $\langle x\vert x'\rangle=\delta(x-x')$ is distributional normalization, not an ordinary inner product between vectors of finite norm. See Hilbert Spaces for completeness and continuous duality.

Positive Semidefinite and Indefinite Forms

Some useful pairings resemble inner products but fail positive definiteness. A positive-semidefinite pairing permits nonzero null vectors with

$$\langle v\vert v\rangle=0.$$

One can sometimes quotient by the null subspace to obtain an inner-product space. An indefinite Hermitian form permits positive, zero, and negative self-pairings. Such forms occur in relativistic and constrained theories but are not Hilbert-space inner products.

Do not infer probabilistic meaning from angle-bracket notation alone. The positive-definite Hilbert-space pairing and the physical Born rule are both needed.

Common Mistakes

• Forgetting the active convention. Scalars leave the first slot conjugated under the physics convention.
• Using a transpose instead of an adjoint. Complex self-pairings require conjugation.
• Assuming coordinates are orthonormal. Insert the Gram matrix in a nonorthogonal basis.
• Ignoring the measure or weight. Function-space orthogonality depends on both.
• Confusing an inner product with componentwise multiplication. The pairing returns one scalar.
• Equating linear independence with orthogonality. Orthogonality is stronger.
• Treating delta-normalized kets as finite-norm vectors. Their pairing is distributional.
• Reading an overlap directly as a probability. The states must be normalized and a measurement context and Born-rule assignment supplied.

Exercises

1. On $\mathbb C^2$, define

$$\langle u\vert v\rangle_G = u^\dagger \begin{pmatrix} 2&1\\ 1&2 \end{pmatrix} v.$$

Show that this is an inner product.

Solution

The matrix $G$ is Hermitian, so the pairing has conjugate symmetry. Matrix multiplication gives linearity in the second slot and conjugate linearity in the first. For $v=(x,y)^{\mathsf T}$,

$$\begin{aligned} v^\dagger Gv &= 2\lvert x\rvert^2 +x^*y+y^*x +2\lvert y\rvert^2 \\ &= \lvert x+y\rvert^2 +\lvert x\rvert^2 +\lvert y\rvert^2. \end{aligned}$$

This is nonnegative and vanishes only when $x=y=0$, so the pairing is positive definite.

2. Let $u,v\ne0$. Prove that equality holds in Cauchy–Schwarz if and only if $u$ and $v$ are linearly dependent.

Solution

In the proof above,

$$u = \alpha v+w, \qquad \langle v\vert w\rangle=0,$$

and

$$\lVert u\rVert^2 = \lvert\alpha\rvert^2\lVert v\rVert^2 +\lVert w\rVert^2.$$

The inequality becomes an equality exactly when $\lVert w\rVert^2=0$. Positive definiteness then gives $w=0$, so $u=\alpha v$. Conversely, if $u=\alpha v$, direct substitution gives

$$\lvert\langle u\vert v\rangle\rvert = \lVert u\rVert\,\lVert v\rVert.$$

3. In the standard inner product on $\mathbb C^2$, let $e_1=(1,0)^{\mathsf T}$ and $e_2=(1,1)^{\mathsf T}$. Find the Gram matrix and the norm of the vector with basis coordinates $c=(2,-1)^{\mathsf T}$.

Solution

The Gram matrix is

$$G = \begin{pmatrix} 1&1\\ 1&2 \end{pmatrix}.$$

Therefore

$$\begin{aligned} \lVert v\rVert^2 &= c^\dagger Gc \\ &= \begin{pmatrix} 2&-1 \end{pmatrix} \begin{pmatrix} 1&1\\ 1&2 \end{pmatrix} \begin{pmatrix} 2\\ -1 \end{pmatrix} \\ &=2. \end{aligned}$$

Hence $\lVert v\rVert=\sqrt2$. As a check, $v=2e_1-e_2=(1,-1)^{\mathsf T}$ in the standard basis, which has the same norm.

4. Let normalized vectors satisfy $\lvert\langle\phi\vert\psi\rangle\rvert=1$. What does Cauchy–Schwarz imply, and what is the physical conclusion for pure states?

Solution

Equality in Cauchy–Schwarz implies linear dependence:

$$\lvert\psi\rangle = \lambda\lvert\phi\rangle.$$

Normalization gives $\lvert\lambda\rvert=1$, so $\lambda=e^{i\alpha}$ for some real $\alpha$. The vectors differ only by a global phase and therefore represent the same pure-state ray. This conclusion uses the physical identification of pure states with rays in addition to the mathematical equality condition.

References

• S. Axler, Linear Algebra Done Right, 3rd ed., Springer, 2015.
• R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed., Cambridge University Press, 2013.
• B. C. Hall, Quantum Theory for Mathematicians, Springer, 2013.
• M. Reed and B. Simon, Methods of Modern Mathematical Physics, Volume I: Functional Analysis, revised and enlarged ed., Academic Press, 1980.