Orthonormal Diagonalization | KEVOS® Mathematics

Overview

A general diagonalisation $S^{-1}AS = D$ places no constraint on $S$ beyond invertibility, and that freedom is expensive: the eigenvector matrix may be badly conditioned, the inverse must be computed, and the transformation distorts lengths and angles. Orthonormal diagonalization is the case where all three problems vanish at once. The transforming matrix is unitary, its inverse is its adjoint, and the change of basis is a rigid rotation of $\mathbb{C}^n$ .

The matrices that admit such a diagonalisation are not rare. Every Hermitian matrix — every matrix equal to its own conjugate transpose, and in the real case every symmetric matrix — has an orthonormal basis of eigenvectors. Since symmetric matrices arise wherever a physical quantity is defined by a quadratic form (stiffness, inertia, stress, covariance, energy), the guarantee covers a large fraction of the matrices an engineer actually meets.

Two structural facts do most of the work. A Hermitian matrix has only real eigenvalues, so quantities that must be physically meaningful — a natural frequency, a principal stress, a variance — never come out complex. And eigenvectors of a Hermitian matrix belonging to different eigenvalues are automatically orthogonal, so no work is needed to separate distinct modes. Only within a repeated eigenvalue's eigenspace is a choice required, and there the Gram-Schmidt procedure supplies an orthonormal basis.

The result is a decomposition $A = UDU^{*}$ in which $A$ is written as a weighted sum of orthogonal projectors, one per eigenvalue. This is the spectral decomposition, and it is the algebraic content of principal axes in mechanics, principal components in statistics, normal modes in vibration and observables in quantum mechanics. Because $U$ is unitary the decomposition is also numerically ideal: errors are neither amplified nor rotated into unexpected directions.

Definition

Hermitian Matrix

A square matrix $A$ is Hermitian, or self-adjoint, when it equals its own adjoint: $A^{*} = A$ , where $A^{*} = \overline{A}^{\,t}$ is the conjugate transpose. Entrywise this says $[A]_{ij} = \overline{[A]_{ji}}$ , which forces every diagonal entry to be real. A matrix with only real entries is Hermitian exactly when it is symmetric, $A^{t} = A$ .

Every theorem on this page stated for a Hermitian matrix may be read with "real symmetric" substituted throughout, with the adjoint replaced by the transpose and unitary replaced by real orthogonal.

Unitary Matrix

A square matrix $U$ is unitary when $U^{*}U = I_n$ . Equivalently, the columns of $U$ form an orthonormal set, and equivalently again $U^{-1} = U^{*}$ , so the inverse is available at no computational cost. A real unitary matrix is called orthogonal and satisfies $Q^{t}Q = I_n$ .

Unitary matrices preserve the inner product, $\langle U\vec{x}, U\vec{y} \rangle = \langle \vec{x}, \vec{y} \rangle$ , hence also lengths and angles. Geometrically they are the rotations and reflections of $\mathbb{C}^n$ .

Orthonormally Diagonalisable Matrix

ODM

A square matrix $A$ is orthonormally diagonalisable (equivalently unitarily diagonalisable) when there is a unitary $U$ and a diagonal $D$ with $U^{*}AU = D$ . Because $U^{*} = U^{-1}$ this is a similarity transformation, so $D$ carries the eigenvalues of $A$ ; the additional content is that the eigenvector basis can be chosen orthonormal.

This is strictly stronger than being diagonalisable. The matrix $\begin{bmatrix} 1 & 1 \\ 0 & 2 \end{bmatrix}$ is diagonalisable but not orthonormally so, because its two eigenvectors are not perpendicular.

Normal Matrix

NRM

A square matrix $A$ is normal when it commutes with its own adjoint, $A^{*}A = AA^{*}$ . Hermitian, skew-Hermitian and unitary matrices are all normal. Normality is the exact condition for orthonormal diagonalisability: a matrix is unitarily diagonalisable if and only if it is normal, a statement obtained by applying the Schur decomposition and then forcing the triangular factor to be diagonal.

Concepts

Hermitian matrices have real eigenvalues

Let $\vec{x} \ne \vec{0}$ satisfy $A\vec{x} = \lambda\vec{x}$ with $A$ Hermitian. Evaluating the inner product $\langle A\vec{x}, \vec{x} \rangle$ in two ways gives $\lambda \langle \vec{x}, \vec{x} \rangle$ on one side and, moving $A$ across the inner product using $A^{*} = A$ , $\overline{\lambda} \langle \vec{x}, \vec{x} \rangle$ on the other. Since $\vec{x}$ is non-zero, $\langle \vec{x}, \vec{x} \rangle$ is a positive real number and may be cancelled, leaving $\lambda = \overline{\lambda}$ . A number equal to its own conjugate has zero imaginary part, so $\lambda \in \mathbb{R}$ . The engineering consequence is direct: for a symmetric stiffness or covariance matrix, no eigenvalue can be complex, and quantities read off the diagonal are always physically interpretable.

Eigenvectors for distinct eigenvalues are already orthogonal

Take $A\vec{x} = \lambda\vec{x}$ and $A\vec{y} = \rho\vec{y}$ with $\lambda \ne \rho$ and $A$ Hermitian. The same two-sided evaluation of $\langle A\vec{x}, \vec{y} \rangle$ produces $\lambda \langle \vec{x}, \vec{y} \rangle$ on one side and $\rho \langle \vec{x}, \vec{y} \rangle$ on the other, using the fact that $\rho$ is real. Subtracting gives $(\lambda - \rho)\langle \vec{x}, \vec{y} \rangle = 0$ , and since $\lambda \ne \rho$ the inner product must vanish. Eigenvectors from different eigenspaces are therefore perpendicular with no orthogonalisation work at all — distinct vibration modes of a symmetric structure are automatically decoupled.

Orthogonality within a repeated eigenspace is a choice, not a gift

When an eigenvalue is repeated, its eigenspace has dimension greater than one and any basis of it consists of eigenvectors. Such a basis need not be orthogonal — row reduction of $A - \lambda I_n$ produces a perfectly valid basis with no orthogonality property whatsoever. The Gram-Schmidt procedure applied within that eigenspace fixes this: every vector it produces is still a linear combination of eigenvectors for $\lambda$ , hence still an eigenvector for $\lambda$ , and the output is orthonormal. Doing this eigenspace by eigenspace, then concatenating, yields a full orthonormal eigenvector basis because different eigenspaces are already mutually orthogonal.

The spectral theorem and its decomposition

Combining the previous three results gives the spectral theorem: for every Hermitian $A$ of size $n$ there is a unitary $U$ and a real diagonal $D$ with $U^{*}AU = D$ , equivalently $A = UDU^{*}$ . Expanding the product columnwise turns this into $A = \sum_{i=1}^{n} \lambda_i \vec{u}_i \vec{u}_i^{\,*}$ , a sum of rank-one orthogonal projectors scaled by the eigenvalues. Grouping the terms belonging to one eigenvalue gives the projector onto its eigenspace, so $A$ is a weighted sum of orthogonal projectors onto mutually perpendicular subspaces that together fill $\mathbb{C}^n$ .

Why unitarity is worth more than diagonality alone

Three properties follow from $U^{-1} = U^{*}$ and none of them holds for a general diagonalisation. First, the inverse is free: no linear solve is needed to apply $U^{-1}$ . Second, the condition number $\kappa_2(U) = 1$ , so the transformation neither amplifies input error nor loses accuracy when the eigendecomposition is used to compute $A^{k}$ , $e^{At}$ or $A^{-1}$ . Third, the change of basis preserves the inner product, so orthogonality, lengths, energy norms and least-squares residuals all mean the same thing before and after. A general $S$ preserves none of these.

Quadratic forms and principal axes

For Hermitian $A$ the scalar $\vec{x}^{*}A\vec{x}$ is always real, and substituting $\vec{x} = U\vec{z}$ turns it into $\sum_i \lambda_i |z_i|^2$ — all cross terms disappear. This is the principal axis theorem: every quadratic form can be written as a weighted sum of squares in a rotated coordinate frame, the weights being the eigenvalues. It classifies the form (positive definite when every $\lambda_i > 0$ , indefinite when signs are mixed), it bounds the Rayleigh quotient between the smallest and largest eigenvalues, and it identifies the principal stress directions of a stress tensor and the principal axes of an inertia tensor as the same computation.

Beyond the Hermitian case

Orthonormal diagonalisability extends exactly as far as normality, $A^{*}A = AA^{*}$ . Unitary matrices are normal, and so are skew-Hermitian matrices; their eigenvalues lie on the unit circle and on the imaginary axis respectively, rather than on the real line. Every square matrix, normal or not, admits a Schur decomposition $A = QTQ^{*}$ with $Q$ unitary and $T$ upper triangular, and normality is precisely the condition that forces $T$ to be diagonal. When a matrix is non-normal, the Schur form is the right numerical target and the eigenvectors should not be trusted as a basis.

Procedure: construct an orthonormal diagonalisation

Confirm the structureVerify

A^{*} = A

entrywise (or

A^{t} = A

for real data). If the matrix is only nearly symmetric because of round-off, symmetrise it explicitly as

\tfrac{1}{2}(A + A^{*})

before proceeding.

Find the eigenvaluesCompute the roots of

p_A(x) = \det(A - xI_n)

. They are guaranteed real, so any imaginary part appearing in a numerical result is round-off and should be discarded.

Compute each eigenspaceRow-reduce

A - \lambda_i I_n

and read a basis for its null space. The dimensions are guaranteed to sum to

n

, so a shortfall indicates an arithmetic error rather than a defective matrix.

Orthonormalise within each eigenspaceApply Gram-Schmidt separately inside each eigenspace, then normalise. Vectors from different eigenspaces need no treatment: they are orthogonal already.

Assemble $U$ and $D$ Place the orthonormal eigenvectors as the columns of

U

and the matching eigenvalues on the diagonal of

D

, in the same order.

Verify unitarity and the diagonalisationCheck

U^{*}U = I_n

and

AU = UD

. Both are multiplication-only tests; neither requires an inverse.

Equations

Hermitian condition

EQ-OD-01

A^{*} = {\bar{A}}^{t} = A, [A]_{i j} = \bar{[A]_{j i}}

The defining symmetry. For a real matrix the conjugation is inert and the condition reduces to $A^{t} = A$ .

Unitary condition

EQ-OD-02

U^{*} U = U U^{*} = I_{n} \Leftrightarrow U^{- 1} = U^{*}

Equivalently, the columns of $U$ form an orthonormal basis of $\mathbb{C}^n$ . The real case is $Q^{t}Q = I_n$ , an orthogonal matrix.

Spectral theorem for Hermitian matrices

EQ-OD-03

U^{*} A U = D = diag (λ_{1}, λ_{2}, \dots, λ_{n}), λ_{i} \in ℝ

For every Hermitian $A$ such a unitary $U$ exists. The eigenvalues are real and the columns of $U$ are an orthonormal basis of eigenvectors.

Spectral decomposition as a sum of projectors

EQ-OD-04

A = U D U^{*} = \sum_{i}^{=} λ_{i} {\vec{u}}_{i} {\vec{u}}_{i}^{}

Each term $\vec{u}_i \vec{u}_i^{\,*}$ is the orthogonal projector onto the line spanned by $\vec{u}_i$ . Grouping equal eigenvalues gives the projector onto the full eigenspace.

Reality of the eigenvalues

EQ-OD-05

λ ⟨ \vec{x}, \vec{x} ⟩ = ⟨ A \vec{x}, \vec{x} ⟩ = ⟨ \vec{x}, A \vec{x} ⟩ = \bar{λ} ⟨ \vec{x}, \vec{x} ⟩ \Rightarrow λ = \bar{λ}

The one-line proof: since $\langle \vec{x}, \vec{x} \rangle \ne 0$ for an eigenvector, the scalar equals its own conjugate and is therefore real.

Orthogonality across distinct eigenvalues

EQ-OD-06

(λ - ρ) ⟨ \vec{x}, \vec{y} ⟩ = 0, λ \neq ρ \Rightarrow \vec{x} ⊥ \vec{y}

Eigenvectors of a Hermitian matrix for different eigenvalues are perpendicular automatically, which is why distinct modes decouple without any orthogonalisation step.

Diagonalisation of a quadratic form

EQ-OD-07

{\vec{x}}^{*} A \vec{x} = {\vec{z}}^{*} D \vec{z} = \sum_{i}^{=} λ_{i} | z_{i} |^{2}, \vec{x} = U \vec{z}

The principal axis theorem. In the rotated frame every cross term vanishes and the form is a weighted sum of squares, with the eigenvalues as weights.

Normality is exactly the condition for unitary diagonalisability

EQ-OD-08

A^{*} A = A A^{*} \Leftrightarrow \exists U unitary with U^{*} A U diagonal

Hermitian, skew-Hermitian and unitary matrices are all normal. A non-normal matrix may still be diagonalisable, but never by a unitary matrix.

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$A$	Subject matrix	The Hermitian (or normal) matrix being diagonalised	n x n complex matrix
$A^{*}$	Adjoint	Conjugate transpose $\overline{A}^{\,t}$ ; equals $A^{t}$ for real matrices	n x n complex matrix
$U$	Unitary diagonalising matrix	Columns are an orthonormal basis of eigenvectors of $A$	n x n unitary matrix
$D$	Diagonal form	Real diagonal matrix of eigenvalues, in the order of the columns of $U$	n x n real diagonal matrix
$λ_{i}$	Eigenvalue	Real eigenvalue of a Hermitian matrix; a diagonal entry of $D$	real scalar
${\vec{u}}_{i}$	Orthonormal eigenvector	Column $i$ of $U$ , with $\langle \vec{u}_i, \vec{u}_j \rangle = 0$ for $i \ne j$ and unit norm	unit vector in C^n
$⟨ \vec{x}, \vec{y} ⟩$	Inner product	Conjugate-linear inner product used to define orthogonality and norm	complex scalar
$κ_{2} (U)$	Spectral condition number	Ratio of largest to smallest singular value; exactly $1$ for a unitary matrix	real, at least 1
$Q$	Real orthogonal matrix	The real special case of $U$ , satisfying $Q^{t}Q = I_n$	n x n real orthogonal matrix

Worked Numerical Example

Problem statement

A three-channel vibration monitoring array returns the sample covariance matrix below, in units of $\mathrm{mg}^{2}$ . Diagonalise it orthonormally to obtain the principal components: the uncorrelated directions in sensor space and the variance carried by each.

Confirm the structure and state the matrix
A sample covariance matrix is real and symmetric by construction, so the spectral theorem applies before any computation is done. Every eigenvalue will be real and an orthonormal eigenvector basis is guaranteed to exist.
$A = [\begin{matrix} 5 & 1 & 1 \\ 1 & 5 & 1 \\ 1 & 1 & 5 \end{matrix}] = A^{t}$
Find the eigenvalues
The trace is $15$ and the determinant is $5(24) - 1(4) + 1(-4) = 112$ . Writing $A = 4I_3 + J$ where $J$ is the all-ones matrix makes the spectrum immediate: $J$ has rank one with eigenvalues $3, 0, 0$ , so $A$ has eigenvalues $7, 4, 4$ . These check against the trace $7 + 4 + 4 = 15$ and the determinant $7 \cdot 16 = 112$ .
$p_{A} (x) = (x - 7) (x - 4)^{2}, α_{A} (7) = 1, α_{A} (4) = 2$
Compute the eigenspace for $\lambda = 7$
Row-reducing $A - 7I_3$ leaves two pivots and one free variable, so the eigenspace is a line. Its direction is the equal-weight combination of all three channels: the common-mode signal.
$Compute the eigenspace for $\lambda = 7$$
Compute the eigenspace for $\lambda = 4$
Here $A - 4I_3$ is the all-ones matrix, whose reduced form is a single row. The eigenspace is the plane $x_1 + x_2 + x_3 = 0$ , of dimension $2$ , matching $\alpha_A(4) = 2$ exactly as the spectral theorem requires. The basis produced by row reduction is not orthogonal.
$Compute the eigenspace for $\lambda = 4$$
Orthogonalise inside the repeated eigenspace
The two basis vectors have inner product $1$ , so Gram-Schmidt is needed. Keep $\vec{v}_1 = (1, -1, 0)^{t}$ and subtract its component from $(1, 0, -1)^{t}$ : the coefficient is $\langle (1,0,-1)^{t}, \vec{v}_1 \rangle / \langle \vec{v}_1, \vec{v}_1 \rangle = \tfrac{1}{2}$ . Clearing the fraction by a factor of $2$ leaves an integer vector, still an eigenvector for $\lambda = 4$ because the eigenspace is closed under linear combination.
${\vec{v}}_{2} = [\begin{matrix} 1 \\ 0 \\ - 1 \end{matrix}] - \frac{1}{2} [\begin{matrix} 1 \\ - 1 \\ 0 \end{matrix}] = \frac{1}{2} [\begin{matrix} 1 \\ 1 \\ - 2 \end{matrix}] \sim [\begin{matrix} 1 \\ 1 \\ - 2 \end{matrix}]$
Check the cross-eigenspace orthogonality
No work is required across eigenspaces, but the check is instructive: $\langle (1,1,1)^{t}, (1,-1,0)^{t} \rangle = 0$ and $\langle (1,1,1)^{t}, (1,1,-2)^{t} \rangle = 1 + 1 - 2 = 0$ . Both vanish automatically, exactly as the theorem on distinct eigenvalues predicts.
Normalise and assemble $U$ and $D$
Divide each vector by its norm: $\sqrt{3}$ , $\sqrt{2}$ and $\sqrt{6}$ respectively. The resulting matrix is orthogonal, so its inverse is its transpose and no inversion is needed.
$U = [\begin{matrix} \frac{1}{\sqrt{3}} & \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{6}} \\ \frac{1}{\sqrt{3}} & - \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{6}} \\ \frac{1}{\sqrt{3}} & 0 & - \frac{2}{\sqrt{6}} \end{matrix}], D = [\begin{matrix} 7 & 0 & 0 \\ 0 & 4 & 0 \\ 0 & 0 & 4 \end{matrix}]$
Verify and write the spectral decomposition
Check $U^{t}U = I_3$ column by column, then $AU = UD$ . Grouping the two $\lambda = 4$ terms gives the projector onto the plane, so the decomposition collapses to two terms: the rank-one common-mode projector and its orthogonal complement.
$A = 7 \cdot \frac{1}{3} [\begin{matrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{matrix}] + 4 (I_{3} - \frac{1}{3} [\begin{matrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{matrix}])$

Result

The array has one distinguished direction, the common mode $(1,1,1)^{t}/\sqrt{3}$ , carrying variance $7\ \mathrm{mg}^{2}$ , and a two-dimensional differential subspace carrying $4\ \mathrm{mg}^{2}$ in every direction. Total variance is $15\ \mathrm{mg}^{2}$ , of which the first principal component explains $7/15 \approx 47\%$ . Because the repeated eigenvalue makes the differential subspace isotropic, the individual directions within it are not identifiable from the data — any orthonormal pair spanning that plane is an equally valid answer, a fact worth knowing before reporting principal component loadings as physically meaningful.

Applications & Industry Use

Structural & mechanical engineering

Mass-normalised vibration modes

Stiffness and mass matrices are real symmetric, so the generalised eigenproblem yields real natural frequencies and mode shapes that are orthogonal with respect to both matrices. Mass-normalising the modes makes the modal transformation orthonormal, which diagonalises the equations of motion and lets each mode be integrated, damped or truncated independently.

Solid mechanics

Principal stresses and principal axes

The Cauchy stress tensor is symmetric, so it is orthogonally diagonalisable with real eigenvalues. The eigenvalues are the principal stresses and the orthonormal eigenvectors give the planes on which shear vanishes — the input to every maximum-shear and von Mises failure assessment. The inertia tensor is handled by exactly the same computation.

Data science & statistics

Principal component analysis

A covariance matrix is symmetric positive semi-definite. Its orthonormal diagonalisation produces uncorrelated components ranked by variance, giving the optimal low-rank approximation of a data set in the least-squares sense. Orthonormality is what makes the components genuinely uncorrelated rather than merely independent-looking.

Quantum & computational physics

Observables and measurement outcomes

Physical observables are represented by Hermitian operators precisely because measured values must be real and states associated with distinct outcomes must be distinguishable. The spectral theorem supplies both: real eigenvalues as the possible measurements and an orthonormal eigenbasis as the states, with the projectors giving the outcome probabilities.

Signal processing

Unitary transforms and energy preservation

The discrete Fourier transform matrix, scaled by $1/\sqrt{n}$ , is unitary, and it orthonormally diagonalises every circulant matrix. Because the transform preserves the inner product, signal energy is conserved between the time and frequency domains, and quantisation noise introduced in one domain is not amplified when returning to the other.

Network analysis & optimisation

Graph Laplacians and spectral partitioning

A graph Laplacian is symmetric positive semi-definite, so it diagonalises orthonormally with non-negative eigenvalues. The second smallest eigenvalue measures algebraic connectivity and its eigenvector supplies the partition used in spectral clustering, mesh decomposition and load balancing for parallel solvers.

Design Considerations

Symmetrise before you decompose

A matrix assembled from measurements or from a finite element code is often symmetric in theory but not to the last bit in practice. Feeding it to a general eigensolver invites complex eigenvalues with tiny imaginary parts and non-orthogonal eigenvectors. Replace $A$ by $\tfrac{1}{2}(A + A^{*})$ explicitly and call the Hermitian solver, which is both faster and better conditioned.

Eigenvalues are well conditioned; individual eigenvectors need not be

For a Hermitian matrix each eigenvalue is perturbed by at most the norm of the perturbation — an unconditional guarantee with no dependence on eigenvector geometry. Eigenvectors carry no such guarantee: when two eigenvalues are close, their individual eigenvectors are highly sensitive even though the two-dimensional invariant subspace they span is stable. Report subspaces, not individual directions, in the presence of near-degeneracy.

Repeated eigenvalues make individual axes arbitrary

Within an eigenspace of dimension greater than one, every orthonormal basis is equally correct. Two runs of the same solver, or two different libraries, may legitimately return different axes. Any downstream interpretation that assigns physical meaning to a single direction inside a degenerate eigenspace is unsound; interpret the subspace instead.

Never invert the transforming matrix

The whole point of unitarity is that $U^{-1} = U^{*}$ . Code that forms an explicit inverse of an orthogonal eigenvector matrix is doing $O(n^3)$ of unnecessary work and introducing error into a quantity that was exact. The same applies to solving with $U$ : use a transpose-multiply.

Exploit banded and sparse structure before diagonalising

Dense Hermitian eigensolvers cost $O(n^3)$ and $O(n^2)$ memory, which is prohibitive for large finite element or graph problems. When only the extreme eigenvalues are needed, Lanczos-type iterative methods work with matrix-vector products alone and preserve sparsity, at the cost of returning only part of the spectrum and requiring reorthogonalisation to remain reliable.

Check normality before assuming an orthonormal basis exists

The guarantee applies to Hermitian and, more broadly, normal matrices. A general system matrix in control or fluid dynamics is usually non-normal, and its eigenvectors can be far from orthogonal even when they form a basis. In that regime eigenvalues alone do not predict transient behaviour, and a Schur decomposition or a pseudospectral analysis is the appropriate instrument.

Standards & Codes

Notation, interchange and numerical standards that govern how this material is written down, stored and computed in production systems.

Applicable standards, conventions and reference implementations
Reference	Title	Relevance to this topic
`ISO 80000-2`	Quantities and units — Part 2: Mathematics	Fixes the notation for conjugate, transpose and adjoint used throughout, including the overline for complex conjugation and the superscript for the adjoint $A^{*}$ .
`IEEE 754-2019`	IEEE Standard for Floating-Point Arithmetic	Defines the arithmetic in which a theoretically symmetric matrix acquires an asymmetric last bit, which is why explicit symmetrisation before an eigendecomposition is a practical necessity rather than a formality.
`LAPACK / BLAS reference`	Linear Algebra PACKage reference implementation	Provides the dedicated Hermitian and symmetric drivers `xHEEV`, `xSYEV` and the divide-and-conquer variants, which exploit the guaranteed real spectrum and orthonormal basis for roughly half the cost of the general solver.
`ISO 2041`	Mechanical vibration, shock and condition monitoring — Vocabulary	Defines mode shape, modal mass and normal mode, the mechanical vocabulary for the orthonormal eigenvectors produced by this decomposition.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the adjoint, inner product and matrix notation on this page as structured content, so the distinction between $A^{t}$ and $A^{*}$ survives for assistive technology.

Material Selection

For a mathematical topic, "material" is the numeric representation: the scalar field, storage format and precision the computation is built from.

Representation and precision selection
Representation	Select when	Trade-off
IEEE 754 binary64, real symmetric storage	The default for stiffness, covariance, correlation and Laplacian matrices from real measurement or discretisation.	Uses a dedicated symmetric solver at roughly half the cost of the general one, with a guaranteed real spectrum, but still requires the input to be symmetrised exactly.
Packed or banded symmetric storage	Large structural or finite difference matrices where only a narrow band around the diagonal is non-zero.	Halves or better the memory footprint and speeds the tridiagonal reduction substantially, at the cost of index arithmetic and reduced cache locality relative to a dense layout.
Complex Hermitian (binary64 pairs)	Electromagnetics, quantum mechanics and signal processing where the operator is genuinely complex but self-adjoint.	Doubles storage and roughly quadruples arithmetic per operation, but preserves the real spectrum and the unitary basis exactly as in the real case.
Exact rational arithmetic	Small integer or rational symmetric matrices where the eigenvalues happen to be rational and an exact projector decomposition is required.	Exact and free of tolerance decisions, but the eigenvalues of a symmetric integer matrix are usually irrational, so exactness typically cannot be carried through to the eigenvectors.
Sparse storage with an iterative eigensolver	Very large graph Laplacians or finite element operators where only a handful of extreme eigenpairs is needed.	Costs only matrix-vector products and preserves sparsity, but returns a partial spectrum and needs explicit reorthogonalisation to keep the computed basis orthonormal.
IEEE 754 binary32 or reduced precision	Streaming principal component analysis and embedded condition monitoring where throughput dominates.	Halves memory traffic, but a covariance matrix formed in single precision loses roughly half the significant digits of the smallest eigenvalues, which are exactly the ones used to detect degeneracy.

Manufacturing Notes

Implementation notes — how the result is actually produced by hand, by algorithm and by library, including cost and numerical behaviour.

Algorithm and cost

The standard dense method reduces the Hermitian matrix to real tridiagonal form by Householder reflections at about $\tfrac{4}{3}n^3$ operations, then extracts the spectrum with the implicit QL/QR iteration or a divide-and-conquer scheme. Eigenvalues alone cost roughly $\tfrac{4}{3}n^3$ ; accumulating the orthogonal transformations to obtain eigenvectors brings the total to around $9n^3$ . Both figures are well below the general non-symmetric eigenproblem.

Gram-Schmidt by hand, and its numerical replacement

Classical Gram-Schmidt is the right procedure on paper: subtract from each new vector its projection onto the ones already accepted, then normalise. In floating point it loses orthogonality progressively. Use the modified variant, which subtracts projections one at a time against the updated vector, or better a Householder QR factorisation, which produces an orthonormal basis that is orthogonal to working precision regardless of the input conditioning.

Verification without an inverse

Two multiplication-only residuals suffice. Orthonormality: $\|U^{*}U - I_n\|$ should be at the level of the unit round-off times $n$ . Diagonalisation: $\|AU - UD\|$ relative to $\|A\|$ should be of the same order. A large first residual indicates a Gram-Schmidt failure; a large second indicates a wrong eigenvalue or a mismatched column order.

Library behaviour

numpy.linalg.eigh and scipy.linalg.eigh assume symmetry and read only one triangle of the input, silently ignoring the other — an asymmetric argument produces a wrong answer with no warning. They return eigenvalues in ascending order with matching orthonormal columns. The general routine eig should not be used on symmetric input: it is slower and returns a non-orthogonal, sometimes complex, basis.

Reusing the decomposition

With $A = UDU^{*}$ in hand, matrix functions cost two matrix-vector products plus $n$ scalar evaluations: $A^{-1} = UD^{-1}U^{*}$ , $e^{At} = U e^{Dt} U^{*}$ , and the matrix square root of a positive semi-definite $A$ is $U D^{1/2} U^{*}$ . Whitening a data set, computing a Mahalanobis distance and evaluating a Gaussian log-likelihood all reduce to this single decomposition.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Applying the spectral theorem to a non-symmetric matrix	high	Assuming any square matrix has an orthonormal eigenvector basis, when the guarantee requires normality.	Test $A^{}A = AA^{}$ , or in practice confirm $A^{*} = A$ . For non-normal matrices use a Schur decomposition and do not treat the eigenvectors as a coordinate frame.
Orthogonalising across eigenspaces	medium	Running Gram-Schmidt over the whole concatenated list of eigenvectors instead of within each eigenspace.	Vectors from distinct eigenvalues are already orthogonal, so the extra step is unnecessary; worse, in the presence of round-off it can mix eigenspaces and destroy the eigenvector property. Orthogonalise inside each eigenspace only.
Forgetting to normalise	medium	Assembling orthogonal but non-unit columns, producing $S$ with $S^{t}S$ diagonal rather than the identity.	Divide each column by its norm before assembling $U$ , and verify $U^{*}U = I_n$ rather than merely checking that off-diagonal entries vanish.
Interpreting axes inside a degenerate eigenspace	medium	Assigning physical meaning to individual principal components when the corresponding eigenvalues are equal or nearly equal.	Report the eigenvalue gap alongside the eigenvectors; when the gap is small relative to the perturbation level, interpret only the spanned subspace.
Using the general eigensolver on symmetric input	medium	Calling a generic routine out of habit, which returns complex eigenvalues with round-off-sized imaginary parts and a non-orthogonal basis.	Call the dedicated Hermitian or symmetric driver, which is faster, returns real ordered eigenvalues and guarantees orthonormal columns.
Confusing $A^{t}$ with $A^{*}$ for complex data	high	Transposing without conjugating, so a complex Hermitian matrix is treated as complex symmetric.	Complex symmetric matrices are not normal in general and have no orthonormal eigenbasis. Always conjugate when transposing complex quantities.
Trusting classical Gram-Schmidt in floating point	medium	Loss of orthogonality accumulates rapidly when the input vectors are nearly dependent.	Use modified Gram-Schmidt or a Householder QR factorisation, and check the orthogonality residual $\\|U^{*}U - I_n\\|$ explicitly.
Assuming positive definiteness follows from symmetry	low	Conflating the guaranteed reality of the eigenvalues with a guarantee about their sign.	Symmetry gives real eigenvalues only. Positive definiteness is the separate condition that every eigenvalue is strictly positive, best tested by an attempted Cholesky factorisation.

FAQs

What is the difference between diagonalisable and orthonormally diagonalisable?

Diagonalisable requires only that some invertible $S$ makes $S^{-1}AS$ diagonal, so the eigenvectors need only be independent. Orthonormally diagonalisable requires them to be orthonormal, so that $S$ can be taken unitary. The second is strictly stronger: $\begin{bmatrix} 1 & 1 \\ 0 & 2 \end{bmatrix}$ satisfies the first but not the second.

Why must a Hermitian matrix have real eigenvalues?

Because the scalar $\langle A\vec{x}, \vec{x} \rangle$ can be evaluated in two ways for an eigenvector: once as $\lambda \langle \vec{x}, \vec{x} \rangle$ , once, by moving $A$ across the inner product using $A^{*} = A$ , as $\overline{\lambda} \langle \vec{x}, \vec{x} \rangle$ . Since $\langle \vec{x}, \vec{x} \rangle$ is a non-zero real number it cancels, forcing $\lambda = \overline{\lambda}$ .

Do I always need Gram-Schmidt to build the orthonormal basis?

Only inside eigenspaces of dimension greater than one. Eigenvectors belonging to different eigenvalues of a Hermitian matrix are orthogonal automatically, so if all $n$ eigenvalues are distinct the row-reduction output is already orthogonal and needs nothing more than normalisation.

Which matrices other than Hermitian ones diagonalise orthonormally?

Exactly the normal matrices, those satisfying $A^{*}A = AA^{*}$ . This class includes unitary matrices, whose eigenvalues lie on the unit circle, and skew-Hermitian matrices, whose eigenvalues are purely imaginary. Normality is necessary as well as sufficient, which follows from the Schur decomposition.

How does the spectral decomposition relate to the singular value decomposition?

For a Hermitian positive semi-definite matrix they coincide: the singular values are the eigenvalues and the left and right singular vectors are both the orthonormal eigenvectors. For a Hermitian matrix with negative eigenvalues the singular values are the absolute values, with a sign absorbed into one of the singular vector sets. For a general matrix the SVD exists while a spectral decomposition may not.

Is a real symmetric matrix guaranteed to have positive eigenvalues?

No. Symmetry guarantees only that the eigenvalues are real; they may be positive, negative or zero. Positive definiteness is an additional condition, equivalent to $\vec{x}^{t}A\vec{x} > 0$ for all non-zero $\vec{x}$ , and is usually tested by attempting a Cholesky factorisation rather than by computing eigenvalues.

Why is a unitary transformation preferred numerically?

Because its spectral condition number is exactly one, so it neither amplifies nor attenuates errors. Any computation routed through $A = UDU^{*}$ inherits that stability, whereas a general diagonalisation $SDS^{-1}$ amplifies error by up to $\kappa(S)$ , which is unbounded as the matrix approaches a defective one.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section SD, subsection OD, with subsection PEE.EHM. Licensed under the GNU Free Documentation License v1.2.
Horn, R. A. and Johnson, C. R. Matrix Analysis, 2nd edition. Cambridge University Press, 2013. Chapter 2 on unitary similarity and normal matrices.
Golub, G. H. and Van Loan, C. F. Matrix Computations, 4th edition. Johns Hopkins University Press, 2013. Chapter 8, the symmetric eigenvalue problem.
Parlett, B. N. The Symmetric Eigenvalue Problem. Society for Industrial and Applied Mathematics, 1998.
Anderson, E. et al. LAPACK Users' Guide, 3rd edition. Society for Industrial and Applied Mathematics, 1999.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.

AI Suggested Questions

Show me a matrix that is diagonalisable but not orthonormally diagonalisable, and explain exactly which hypothesis fails.
Work through the Schur decomposition argument that normality is equivalent to unitary diagonalisability.
How do the eigenvalues and eigenvectors of a symmetric matrix respond to a small perturbation, and why is the eigenvalue bound unconditional while the eigenvector bound is not?
Derive the modal decoupling of a damped multi-degree-of-freedom system and identify where the orthonormality of the mode shapes is used.
Compare the spectral decomposition and the singular value decomposition for an indefinite symmetric matrix, term by term.
Explain why the normalised discrete Fourier transform matrix is unitary and what that implies for circulant matrices.

Related Calculators

Orthonormal Diagonalisation Calculator

Produce $U$ and $D$ for a Hermitian or symmetric matrix, with Gram-Schmidt applied inside repeated eigenspaces.

Spectral Decomposition Calculator

Express $A$ as a weighted sum of orthogonal projectors, one per distinct eigenvalue, with the eigenvalue gaps reported.

Gram-Schmidt Calculator

Orthonormalise a set of vectors using the modified procedure and report the resulting orthogonality residual.

Quadratic Form Classifier

Diagonalise $\vec{x}^{*}A\vec{x}$ into principal axes and classify the form as definite, semi-definite or indefinite.