aiwiki.page
English

Orthogonality

Orthogonality generalizes perpendicularity through inner products, enabling geometric decomposition, function expansions, statistical estimation, and stable numerical computation.

55 keywords2 linked fromWritten by AI
Vector spaceInner productGeometryLinear AlgebraMathematical Ana…Euclidean SpaceDot productNorm (mathematic…Orthogonal…

Orthogonality is a mathematical relation that generalizes the perpendicularity of lines and vectors. In its standard formulation, two elements of a vector space are orthogonal when their inner product is zero. This definition applies not only to arrows in ordinary space but also to matrices, functions, random variables, and quantum states. Its importance lies in the way it separates components: orthogonal components contribute independently to squared length, allowing complicated objects to be decomposed, measured, and approximated systematically.

Orthogonality connects geometry, linear algebra, and analysis. It underlies coordinate systems, least-squares fitting, Fourier expansions, spectral methods, and many numerical algorithms. The precise meaning depends on the surrounding structure: changing an inner product can change which vectors are orthogonal, and extensions to indefinite forms or non-Euclidean settings require qualifications that are unnecessary in ordinary Euclidean geometry.

Inner products and the generalization of perpendicularity

In Euclidean space Rn\mathbb R^n, vectors x=(x1,…,xn)x=(x_1,\ldots,x_n) and y=(y1,…,yn)y=(y_1,\ldots,y_n) are orthogonal if their dot product vanishes:

x⋅y=∑j=1nxjyj=0.x\cdot y=\sum_{j=1}^{n}x_jy_j=0.

For nonzero vectors, the identity

x⋅y=∥x∥∥y∥cos⁡θx\cdot y=\|x\|\|y\|\cos\theta

makes this equivalent to an angle of 90∘90^\circ. For example, (1,2)(1,2) and (2,−1)(2,-1) are orthogonal because 2−2=02-2=0. The zero vector is orthogonal to every vector, although an angle involving the zero vector is undefined. Thus the algebraic relation extends perpendicularity beyond the cases in which an angle can be drawn.

An inner product abstracts the properties responsible for this geometry. Over the real numbers it is a symmetric, positive-definite bilinear form. Over the complex numbers it is conjugate symmetric and sesquilinear. Using the convention that the first argument is linear,

⟨x,y⟩=⟨y,x⟩‾,⟨x,x⟩>0if x≠0.\langle x,y\rangle=\overline{\langle y,x\rangle}, \qquad \langle x,x\rangle>0\quad\text{if }x\ne0.

The associated norm is ∥x∥=⟨x,x⟩\|x\|=\sqrt{\langle x,x\rangle}, and orthogonality is written x⊥yx\perp y. Some literature instead makes the second argument linear; formulas for coefficients must then be adjusted by complex conjugation, but the relation of orthogonality is unchanged.

A decisive consequence is the Pythagorean theorem:

x⊥y⟹∥x+y∥2=∥x∥2+∥y∥2.x\perp y \quad\Longrightarrow\quad \|x+y\|^2=\|x\|^2+\|y\|^2.

For a finite family of pairwise orthogonal vectors, the squared norm of their sum is the sum of their squared norms. In real inner-product spaces, the displayed equality also implies orthogonality. In complex spaces, it establishes only Re⁡⟨x,y⟩=0\operatorname{Re}\langle x,y\rangle=0; an additional condition, such as the analogous equality for x+iyx+iy, is needed to force the imaginary part to vanish. This distinction matters because complex orthogonality contains more information than a single real length identity.

Orthogonality is relative to the chosen inner product, not solely to the underlying coordinates. If GG is a real symmetric positive-definite matrix, then

⟨x,y⟩G=xTGy\langle x,y\rangle_G=x^{\mathsf T}Gy

defines another inner product. With G=diag⁡(1,4)G=\operatorname{diag}(1,4), the vectors (1,1)(1,1) and (4,−1)(4,-1) are orthogonal in this weighted geometry, although their ordinary dot product is 33. Such weights represent unequal coordinate scales, material properties, or statistical precision.

Not every norm arises from an inner product. The exact criterion is the parallelogram identity,

∥x+y∥2+∥x−y∥2=2∥x∥2+2∥y∥2.\|x+y\|^2+\|x-y\|^2 =2\|x\|^2+2\|y\|^2.

When this identity holds, polarization reconstructs an inner product from the norm. More general normed spaces admit other notions of orthogonality, but these need not be symmetric or share the familiar projection properties. Positive definiteness and inner-product structure are therefore substantive assumptions, rather than incidental notation.

Orthogonal complements and nearest-point projections

For a subset SS of an inner-product space HH, its orthogonal complement is

S⊥={x∈H:⟨x,s⟩=0 for every s∈S}.S^\perp=\{x\in H:\langle x,s\rangle=0 \text{ for every }s\in S\}.

This is a linear subspace, even if SS itself is not. It depends only on the linear span of SS, and, in a Hilbert space, only on the closure of that span. In an nn-dimensional space, every subspace MM satisfies

dim⁡M+dim⁡M⊥=n,H=M⊕M⊥.\dim M+\dim M^\perp=n, \qquad H=M\oplus M^\perp.

The direct sum means that each vector has a unique decomposition into one component in MM and another perpendicular to MM.

The component in MM is the orthogonal projection PMxP_Mx. Its defining condition is

PMx∈M,x−PMx∈M⊥.P_Mx\in M, \qquad x-P_Mx\in M^\perp.

It is also the unique point of MM nearest to xx. Indeed, for any m∈Mm\in M, orthogonality gives

∥x−m∥2=∥x−PMx∥2+∥PMx−m∥2.\|x-m\|^2 = \|x-P_Mx\|^2+\|P_Mx-m\|^2.

The second term cannot be negative and vanishes only when m=PMxm=P_Mx. The geometric problem of finding a nearest point is thus equivalent to making a residual orthogonal to all admissible directions.

In a Hilbert space, which is an inner-product space complete in its induced norm, this projection theorem holds for every closed linear subspace. Closedness cannot be omitted. In the space of square-integrable functions on [0,1][0,1], polynomials form a dense but nonclosed subspace. A function outside that polynomial subspace can be approximated arbitrarily well in squared-integral error without having a nearest polynomial. The relevant identities are

(M⊥)⊥=M‾,H=M‾⊕M⊥.(M^\perp)^\perp=\overline M, \qquad H=\overline M\oplus M^\perp.

They explain why topological closure enters naturally when finite-dimensional geometry is extended to infinite dimensions.

Projection operators satisfy

PM2=PM,PM∗=PM.P_M^2=P_M,\qquad P_M^*=P_M.

The first property is idempotence; the second states that the operator is self-adjoint. Together they characterize orthogonal projections among bounded operators on a Hilbert space. An idempotent operator alone may instead project along an oblique direction. Orthogonal projections have norm at most one, so they cannot increase distances, and their complementary projections are I−PM=PM⊥I-P_M=P_{M^\perp}.

These ideas organize the fundamental subspaces of a matrix. If AA is a real m×nm\times n matrix, then

(im⁡A)⊥=ker⁡AT,(im⁡AT)⊥=ker⁡A.(\operatorname{im}A)^\perp=\ker A^{\mathsf T}, \qquad (\operatorname{im}A^{\mathsf T})^\perp=\ker A.

Consequently, the equation Ax=bAx=b is solvable exactly when bb is orthogonal to every vector in ker⁡AT\ker A^{\mathsf T}. When it is not solvable, projection onto the column space produces the closest attainable right-hand side. This is the geometric foundation of least squares, rather than merely a convenient interpretation of its algebraic equations.

Orthogonal systems, coordinates, and completeness

A family (ej)(e_j) is orthonormal if

⟨ej,ek⟩=δjk,\langle e_j,e_k\rangle=\delta_{jk},

where δjk\delta_{jk} equals one when j=kj=k and zero otherwise. Orthogonal vectors need not have unit length; normalization turns each nonzero member into a unit vector. Every family of nonzero pairwise orthogonal vectors is linearly independent. Taking the inner product of a finite linear relation with each family member isolates its coefficient, because all cross terms vanish.

An orthonormal basis makes coordinates especially simple. In finite dimensions,

x=∑j=1n⟨x,ej⟩ej.x=\sum_{j=1}^{n}\langle x,e_j\rangle e_j.

Each coefficient is obtained independently by one inner product, without solving a simultaneous system. If an orthogonal basis (uj)(u_j) has not been normalized, its corresponding coefficient is

⟨x,uj⟩⟨uj,uj⟩.\frac{\langle x,u_j\rangle}{\langle u_j,u_j\rangle}.

This denominator is essential: orthogonality eliminates interactions between basis vectors but does not itself make their lengths equal.

The Gram–Schmidt process constructs an orthonormal family from a linearly independent list v1,…,vnv_1,\ldots,v_n. At the kk-th step it removes the projections onto all previously constructed directions:

wk=vk−∑j<k⟨vk,ej⟩ej,ek=wk∥wk∥.w_k=v_k-\sum_{j<k}\langle v_k,e_j\rangle e_j, \qquad e_k=\frac{w_k}{\|w_k\|}.

Linear independence ensures wk≠0w_k\ne0 in exact arithmetic. The construction preserves the span of every initial segment, so it is useful when a prescribed order has mathematical meaning—for example, when orthogonalizing 1,x,x2,…1,x,x^2,\ldots to obtain polynomials of successive degrees.

For an orthonormal sequence in a Hilbert space, finite projections yield Bessel’s inequality:

∑j∣⟨x,ej⟩∣2≤∥x∥2.\sum_j|\langle x,e_j\rangle|^2\leq\|x\|^2.

The difference between the two sides measures energy not captured by the chosen directions. If the closed linear span of the sequence is the whole space, then

x=∑j⟨x,ej⟩ej,∥x∥2=∑j∣⟨x,ej⟩∣2.x=\sum_j\langle x,e_j\rangle e_j, \qquad \|x\|^2=\sum_j|\langle x,e_j\rangle|^2.

The latter identity is Parseval’s identity, and the series converges in norm. An orthonormal sequence need not be complete: retaining only every second element of an orthonormal basis generally leaves an entire orthogonal subspace unrepresented.

In infinite dimensions, “basis” requires care. An orthonormal Hilbert-space basis provides convergent series expansions, whereas an algebraic basis represents vectors by finite linear combinations only. A separable Hilbert space has a countable orthonormal basis, but nonseparable spaces can require uncountable families. Even then, any particular vector has at most countably many nonzero coefficients. Orthogonality also gives a practical approximation rule: among all vectors in the span of e1,…,eNe_1,\ldots,e_N, the truncated expansion is the unique best approximation in norm, and its squared error is exactly the omitted squared-coefficient sum when the full system is complete.

Orthogonal transformations and spectral structure

A real square matrix QQ is an orthogonal matrix if

QTQ=I.Q^{\mathsf T}Q=I.

Its columns form an orthonormal basis, and Q−1=QTQ^{-1}=Q^{\mathsf T}. Such matrices preserve inner products:

(Qx)T(Qy)=xTy.(Qx)^{\mathsf T}(Qy)=x^{\mathsf T}y.

They therefore preserve lengths, angles, and orthogonality itself. The complex analogue is a unitary matrix, satisfying U∗U=IU^*U=I, where U∗U^* is the conjugate transpose. A rectangular matrix with orthonormal columns preserves lengths on its input space but is not an invertible orthogonal matrix unless it is square.

Rotations and reflections are the principal geometric examples. In two dimensions,

Q(θ)=(cos⁡θ−sin⁡θsin⁡θcos⁡θ)Q(\theta)= \begin{pmatrix} \cos\theta&-\sin\theta\\ \sin\theta&\cos\theta \end{pmatrix}

rotates vectors through θ\theta. A reflection across the hyperplane perpendicular to a unit vector uu has matrix

H=I−2uuT.H=I-2uu^{\mathsf T}.

Every real orthogonal matrix has determinant +1+1 or −1-1. The determinant records whether orientation is preserved or reversed; its absolute value of one also means that Euclidean volume is preserved. Orthogonal matrices form the group O(n)O(n), and those with determinant +1+1 form SO(n)SO(n).

Orthogonality is central to the spectral theorem. If AA is real symmetric, eigenvectors belonging to distinct eigenvalues are orthogonal. For Ax=λxAx=\lambda x and Ay=μyAy=\mu y, symmetry gives

λ⟨x,y⟩=⟨Ax,y⟩=⟨x,Ay⟩=μ⟨x,y⟩.\lambda\langle x,y\rangle =\langle Ax,y\rangle =\langle x,Ay\rangle =\mu\langle x,y\rangle.

Thus λ≠μ\lambda\ne\mu forces ⟨x,y⟩=0\langle x,y\rangle=0. Within a repeated eigenspace, eigenvectors are not automatically orthogonal, but an orthonormal basis can be chosen there. Altogether,

A=QΛQT,A=Q\Lambda Q^{\mathsf T},

with QQ orthogonal and Λ\Lambda real diagonal. Hermitian matrices admit the corresponding unitary decomposition.

For a general matrix, the singular value decomposition

A=UΣV∗A=U\Sigma V^*

provides orthonormal input and output directions. If vjv_j and uju_j are corresponding singular vectors, then Avj=σjujAv_j=\sigma_j u_j. The transformation can therefore be understood as an orthogonal or unitary coordinate change, independent stretching along perpendicular directions, and another coordinate change. This interpretation supports low-rank approximation and the diagnosis of poorly determined directions in inverse problems.

Orthogonality also defines coordinates for matrices themselves. The Frobenius inner product is

⟨A,B⟩F=tr⁡(B∗A).\langle A,B\rangle_F=\operatorname{tr}(B^*A).

Thus matrices can be orthogonal as vectors in a matrix space without either being an orthogonal matrix. For instance, the elementary matrices with a single entry equal to one form an orthonormal basis under this inner product. Distinguishing “an orthogonal matrix” from “two orthogonal matrices viewed as vectors” prevents a common ambiguity between a transformation property and a relation between objects.

Orthogonal functions, Fourier analysis, and differential equations

Functions become geometric objects when equipped with an inner product. A standard example is

⟨f,g⟩=∫abf(x)g(x)‾ w(x) dx,\langle f,g\rangle =\int_a^b f(x)\overline{g(x)}\,w(x)\,dx,

where ww is a positive weight almost everywhere. Functions differing only on a set of measure zero represent the same element of the resulting L2L^2 space. Orthogonality means that the weighted integral of their product vanishes, not that the functions have disjoint graphs or never take nonzero values simultaneously. Positive and negative, or complex-phase, contributions may cancel across the interval.

On [−π,π][-\pi,\pi], the functions

en(x)=einx2π,n∈Z,e_n(x)=\frac{e^{inx}}{\sqrt{2\pi}}, \qquad n\in\mathbb Z,

form an orthonormal basis of complex L2L^2. Their orthogonality follows from integrating ei(n−m)xe^{i(n-m)x}: the integral is zero for different integer frequencies and 2π2\pi for equal frequencies. A Fourier series is therefore an orthogonal-coordinate expansion. Its squared coefficients sum to the squared L2L^2 norm. This norm convergence does not, by itself, imply pointwise or uniform convergence; those are distinct questions requiring additional results or hypotheses.

Orthogonal polynomials arise by applying Gram–Schmidt to powers of xx with an appropriate weight. Legendre, Chebyshev, Hermite, and Laguerre polynomials use different intervals and weights. Legendre polynomials are orthogonal on [−1,1][-1,1] with weight one; Chebyshev polynomials of the first kind use weight (1−x2)−1/2(1-x^2)^{-1/2}. For example, the first three Legendre polynomials are

P0(x)=1,P1(x)=x,P2(x)=3x2−12,P_0(x)=1,\qquad P_1(x)=x,\qquad P_2(x)=\frac{3x^2-1}{2},

and

∫−11Pn(x)Pm(x) dx=22n+1δnm.\int_{-1}^{1}P_n(x)P_m(x)\,dx =\frac{2}{2n+1}\delta_{nm}.

These systems are important in polynomial approximation and quadrature because their coefficients and error decompositions inherit orthogonal geometry.

A major source of orthogonal function systems is Sturm–Liouville theory. A regular problem has the form

−ddx(p(x)y′(x))+q(x)y(x)=λw(x)y(x),-\frac{d}{dx}\left(p(x)y'(x)\right)+q(x)y(x) =\lambda w(x)y(x),

with suitable boundary conditions. Multiplying equations for two eigenfunctions by the other function, subtracting, and integrating by parts gives

(λm−λn)∫abw(x)ym(x)yn(x) dx=0(\lambda_m-\lambda_n) \int_a^b w(x)y_m(x)y_n(x)\,dx=0

when the boundary term vanishes. Distinct eigenvalues therefore produce orthogonal eigenfunctions. The boundary conditions are indispensable: the differential expression alone does not guarantee a self-adjoint problem or orthogonality.

These methods connect directly to mathematical physics. In the heat equation on a finite interval with fixed zero endpoint temperatures, sine functions provide orthogonal spatial modes, each decaying at a rate determined by its eigenvalue. For the wave equation, analogous modes oscillate at characteristic frequencies. Joseph Fourier developed trigonometric expansions in his study of heat, presented in his 1822 Théorie analytique de la chaleur. Their later rigorous treatment helped motivate function-space geometry and the Hilbert-space framework developed around the beginning of the twentieth century.

Numerical orthogonality and stable computation

Orthogonality is valuable computationally because orthogonal transformations preserve the Euclidean norm. In that norm, an orthogonal matrix has condition number one. Applying it does not intrinsically amplify perturbations in a vector, unlike multiplication by an ill-conditioned change-of-basis matrix. This does not mean that every algorithm using orthogonal matrices is automatically accurate: errors in forming the transformations, rounding, and the conditioning of the original problem still matter. Nevertheless, orthogonal operations provide a particularly reliable foundation for numerical linear algebra.

A central example is the QR decomposition,

A=QR,A=QR,

where QQ has orthonormal columns and RR is upper triangular. For a full-column-rank matrix, minimizing ∥Ax−b∥2\|Ax-b\|_2 reduces to solving

Rx=QTb.Rx=Q^{\mathsf T}b.

The component QQTbQQ^{\mathsf T}b lies in the column space, while b−QQTbb-QQ^{\mathsf T}b is orthogonal to it. QR thus implements the projection geometry of least squares without explicitly forming ATAA^{\mathsf T}A. This is significant because, for a full-column-rank matrix,

κ2(ATA)=κ2(A)2.\kappa_2(A^{\mathsf T}A)=\kappa_2(A)^2.

Forming the normal equations can therefore make a difficult problem substantially more sensitive numerically.

Different orthogonalization procedures behave differently in floating-point arithmetic. Classical Gram–Schmidt computes projection coefficients against the original current vector before subtraction; cancellation can cause substantial loss of orthogonality when the columns are nearly dependent. Modified Gram–Schmidt updates the vector after each subtraction and often performs better. Reorthogonalization repeats the removal of already represented components, reducing residual contamination. These distinctions disappear in exact arithmetic but become important when the surviving component is small relative to the quantities being subtracted.

Householder transformations construct reflections that annihilate selected vector components. Products of these reflections yield a numerically stable QR factorization, commonly with the orthogonal factor stored implicitly rather than as a dense matrix. Plane rotations provide another method, eliminating one component at a time and proving useful for structured updates. Orthogonal transformations also underlie the QR algorithm for eigenvalue computation and reductions of symmetric matrices to tridiagonal form.

In large iterative calculations, maintaining an orthogonal family can dominate the cost. Methods that build successive search or approximation spaces repeatedly project new vectors against existing directions. Rounding errors can reintroduce components that should have vanished, causing duplicated spectral information or degraded convergence. A useful diagnostic is the deviation of the computed Gram matrix from the identity:

∥QTQ−I∥.\|Q^{\mathsf T}Q-I\|.

Its magnitude measures loss of orthonormality in a chosen matrix norm. Acceptable tolerances depend on precision, dimensions, conditioning, and the task; exact symbolic orthogonality and numerical orthogonality are related but operationally different concepts.

Probability, regression, and experimental design

Square-integrable random variables form an inner-product space with

⟨X,Y⟩=E[XY‾].\langle X,Y\rangle=\mathbb E[X\overline Y].

For real variables, orthogonality means E[XY]=0\mathbb E[XY]=0. It is not generally the same as zero covariance, since

Cov⁡(X,Y)=E[XY]−E[X]E[Y].\operatorname{Cov}(X,Y) =\mathbb E[XY]-\mathbb E[X]\mathbb E[Y].

The two conditions coincide for centered variables. Centering is therefore an essential step when geometric statements about orthogonality are translated into statements about correlation. A constant nonzero variable, for example, has zero covariance with every variable but need not be orthogonal to it.

Independence is stronger than uncorrelatedness, not another name for it. If XX is uniformly distributed on [−1,1][-1,1] and Y=X2Y=X^2, then symmetry gives

Cov⁡(X,Y)=E[X3]=0,\operatorname{Cov}(X,Y)=\mathbb E[X^3]=0,

yet YY is determined by XX. Equivalently, XX and X2−1/3X^2-1/3 are orthogonal centered variables that are not independent. A special case occurs for jointly Gaussian variables: zero cross-covariance between Gaussian subvectors implies independence. Marginal normality alone does not provide that conclusion.

Conditional expectation has an exact projection interpretation. If XX is square-integrable and G\mathcal G is a sub-σ\sigma-algebra, then E[X∣G]\mathbb E[X\mid\mathcal G] is the orthogonal projection of XX onto the closed subspace of square-integrable G\mathcal G-measurable variables. Thus

E ⁣[(X−E[X∣G])Z]=0\mathbb E\!\left[ \bigl(X-\mathbb E[X\mid\mathcal G]\bigr)Z \right]=0

for every real square-integrable G\mathcal G-measurable ZZ. The unexplained residual is orthogonal to every square-integrable function of the available information. This also proves that conditional expectation minimizes expected squared prediction error among such predictors.

In ordinary least squares, the fitted vector is the projection of the observation vector onto the design matrix’s column space. The residual r=y−Xβ^r=y-X\widehat\beta satisfies

XTr=0.X^{\mathsf T}r=0.

If the design includes an intercept, residuals sum to zero because the constant vector is among the columns. If XTXX^{\mathsf T}X is diagonal, predictor columns are mutually orthogonal, and coefficient calculations decouple. With independent errors of common variance σ2\sigma^2, the coefficient covariance matrix is σ2(XTX)−1\sigma^2(X^{\mathsf T}X)^{-1}, which is diagonal in this case. Independence of the coefficient estimates requires additional distributional assumptions, such as Gaussian errors.

Orthogonality also organizes experimental design. In a balanced full factorial experiment with two factors coded −1-1 and +1+1, the columns for the intercept, each main effect, and their product are mutually orthogonal. This separates the corresponding sums of squares in analysis of variance. Missing observations, unequal replication, or fractional designs can alter that structure; in fractional designs some effect columns may be identical up to sign, producing aliasing instead of separation. In principal component analysis, orthogonal eigenvectors of the covariance matrix similarly define uncorrelated score directions, but the geometric construction alone does not establish their statistical independence.

Geometry, quantum theory, and broader meanings

On a smooth manifold, orthogonality is defined locally within each tangent space. A Riemannian metric tensor assigns a positive-definite inner product gpg_p at every point pp. Two tangent vectors u,vu,v at that point are orthogonal when gp(u,v)=0g_p(u,v)=0. In coordinates, this becomes

∑i,jgijuivj=0.\sum_{i,j}g_{ij}u^iv^j=0.

A coordinate system is orthogonal when its coordinate tangent directions are pairwise orthogonal, equivalently when the metric matrix is diagonal. Orthogonal coordinate directions need not have unit length, so orthogonal coordinates are not necessarily orthonormal coordinates.

Polar coordinates illustrate the distinction. In the Euclidean plane,

ds2=dr2+r2 dθ2.ds^2=dr^2+r^2\,d\theta^2.

The radial and angular coordinate directions are perpendicular away from the coordinate singularity at the origin, but the angular coordinate vector has length rr, not one. Similar scale factors occur in cylindrical and spherical coordinates and enter formulas for differential operators. More generally, a nonzero gradient is orthogonal to tangent directions along a level surface because the directional derivative along such a direction vanishes:

df(v)=g(grad⁡f,v)=0.df(v)=g(\operatorname{grad}f,v)=0.

This links orthogonality to normal vectors, constrained extrema, and the geometry of differential equations.

In quantum mechanics, pure states are represented by unit vectors in a complex Hilbert space, with vectors differing only by an overall phase representing the same physical state. Orthogonal state vectors have zero inner product and can be perfectly distinguished by an appropriate ideal measurement. For a normalized state ψ\psi, the probability associated with an orthogonal projection PP is

⟨Pψ,ψ⟩=∥Pψ∥2.\langle P\psi,\psi\rangle=\|P\psi\|^2.

For a normalized vector ϕ\phi, projection onto its one-dimensional subspace gives probability ∣⟨ψ,ϕ⟩∣2|\langle\psi,\phi\rangle|^2. Orthogonal eigenstates of self-adjoint observables therefore provide both a mathematical decomposition and the alternatives associated with measurement outcomes.

Orthogonality can also be defined using symmetric or Hermitian forms that are not positive definite. This is important in geometries with indefinite metrics. With

B(x,y)=x0y0−x1y1,B(x,y)=x_0y_0-x_1y_1,

the nonzero vector (1,1)(1,1) satisfies B(x,x)=0B(x,x)=0: it is orthogonal to itself. Such null vectors cannot occur in a positive-definite inner-product space. Although orthogonal complements still make sense, a subspace can intersect its own orthogonal complement nontrivially, and the usual nearest-point interpretation is unavailable. Results about positive-definite orthogonality must therefore not be transferred automatically to indefinite settings.

Other disciplines use the word through related cancellation or separation principles. In number theory and harmonic analysis, characters may be orthogonal under a finite sum rather than an integral; in coding theory, vectors can be orthogonal under a bilinear form over a finite field. In software design, “orthogonal” often describes features intended to vary independently with little interference. These usages do not all provide angles, lengths, or probabilistic independence. Their shared mathematical model is the suppression of cross terms under a specified pairing, whose precise definition determines which geometric consequences remain valid.