Orthogonal matrices isolate the linear transformations of Euclidean space that preserve the metric structure: lengths, angles, and orthogonality. In linear algebra, a general invertible matrix may stretch space differently in different directions; an orthogonal matrix can rotate, reflect, or combine rotations and reflections without distortion. This makes orthogonal matrices the matrix-level meeting point of [inner product](/page/Inner%20Product), [linear map](/page/Linear%20Map), Euclidean norm, and change of orthonormal coordinates. They are also a bridge between algebra and analysis. In multivariable calculus, a Jacobian matrix that is orthogonal at a point says that the derivative preserves lengths of tangent vectors at that point; by itself, this pointwise condition is not a global isometry statement. In [numerical analysis](/page/Numerical%20Analysis), orthogonal changes of variables are stable because they do not magnify Euclidean errors. In geometry, the [orthogonal group](/page/Orthogonal%20Group) packages the linear distance-preserving transformations of Euclidean space that fix the origin. ## Definition The starting question is: which square matrices preserve the Euclidean inner product exactly? Since the inner product determines both length and angle, preserving it is the strongest finite-dimensional way to say that a linear transformation has no metric distortion. The transpose lets this condition be written as a single matrix identity. [definition: Orthogonal Matrix] Let $n \in \mathbb{N}$. A matrix $Q \in \mathbb{R}^{n \times n}$ is an orthogonal matrix if \begin{align*} Q^\top Q = I_n. \end{align*} [/definition] A single orthogonal matrix represents one rigid linear symmetry. Many arguments need to compose such symmetries, invert them, and compare them as a family. Naming the full collection records the fact that these matrices form a group of Euclidean symmetries. [definition: Orthogonal Group] Let $n \in \mathbb{N}$. The orthogonal group $O(n)$ is the set \begin{align*} O(n) = \{Q \in \mathbb{R}^{n \times n} : Q^\top Q = I_n\}. \end{align*} [/definition] Orientation is not recorded by the equation $Q^\top Q=I_n$. Rotations and reflections can both preserve distances, so the metric condition alone does not distinguish them. The determinant supplies the missing orientation information and leads to the orientation-preserving subgroup. [definition: Special Orthogonal Group] Let $n \in \mathbb{N}$. The special orthogonal group $SO(n)$ is the set \begin{align*} SO(n) = \{Q \in O(n) : \det Q = 1\}. \end{align*} [/definition] This subgroup isolates the determinant-$1$ part of $O(n)$. After the determinant constraint is established below, the complementary determinant-$-1$ part will account for transformations involving a reflection. ## Equivalent Characterisations The matrix equation $Q^\top Q = I_n$ is compact, but it can hide the elementary column-by-column meaning. Each entry of $Q^\top Q$ is a dot product of two columns of $Q$. The next characterisation explains why orthogonal matrices are exactly the matrices whose columns form an orthonormal coordinate frame. [quotetheorem:8283] This column criterion is often the fastest way to test a concrete matrix: compute column lengths and mutual dot products, rather than multiplying out the whole matrix product. Its square-matrix hypothesis matters. Having orthonormal columns fills all of $\mathbb{R}^n$, so the matrix is automatically invertible; for rectangular matrices the same column calculation describes an isometric embedding, not an orthogonal matrix in $O(n)$. For geometric and analytic applications, however, the column test is still too coordinate-bound. What matters next is the metric effect of multiplying by $Q$: lengths, distances, and inner products should be unchanged. A metric criterion is therefore needed because analytic estimates usually care about lengths and inner products rather than individual entries. The column criterion is computational, while applications to isometry require a coordinate-free interpretation. Orthogonality should mean that [Euclidean metric](/page/Euclidean%20Metric) data survives the transformation. The following theorem gives exactly that interpretation. [quotetheorem:439] Coordinate changes are where the metric characterisation becomes indispensable. An arbitrary [change of basis](/page/Change%20Of%20Basis) may turn a unit vector into a coordinate vector of a different Euclidean length, so matrix estimates can become contaminated by the chosen basis. When both bases are orthonormal, the change-of-basis matrix should preserve the coordinate inner product, and the next theorem records that orthonormal coordinate changes are exactly the safe ones. [quotetheorem:8598] This theorem is the coordinate bookkeeping behind many later uses of orthogonal matrices. To see the two basic geometric behaviours before moving to general properties, we start with the planar matrices that rotate or reflect the standard orthonormal frame. ## Standard Examples The lowest-dimensional nontrivial examples already show the two fundamental behaviours: rotation and reflection. These are not merely pictures; they encode the determinant split between $SO(2)$ and the rest of $O(2)$. [example: Planar Rotation Matrix] For $\theta \in \mathbb{R}$, define \begin{align*} R_\theta = \begin{pmatrix} \cos \theta & -\sin \theta \cr \sin \theta & \cos \theta \end{pmatrix}. \end{align*} Its transpose is \begin{align*} R_\theta^\top = \begin{pmatrix} \cos \theta & \sin \theta \cr -\sin \theta & \cos \theta \end{pmatrix}. \end{align*} Multiplying entry by entry gives \begin{align*} R_\theta^\top R_\theta = \begin{pmatrix} \cos \theta \cos \theta + \sin \theta \sin \theta & \cos \theta(-\sin \theta) + \sin \theta \cos \theta \cr (-\sin \theta)\cos \theta + \cos \theta \sin \theta & (-\sin \theta)(-\sin \theta) + \cos \theta \cos \theta \end{pmatrix}. \end{align*} The off-diagonal entries cancel because $-\cos\theta\sin\theta+\sin\theta\cos\theta=0$ and $-\sin\theta\cos\theta+\cos\theta\sin\theta=0$, while the diagonal entries are $1$ by $\cos^2\theta+\sin^2\theta=1$. Hence \begin{align*} R_\theta^\top R_\theta = \begin{pmatrix} 1 & 0 \cr 0 & 1 \end{pmatrix}=I_2. \end{align*} Thus $R_\theta \in O(2)$ by the definition of the orthogonal group. Its determinant is \begin{align*} \det R_\theta = \cos\theta\cos\theta - (-\sin\theta)\sin\theta = \cos^2\theta+\sin^2\theta = 1. \end{align*} Therefore $R_\theta \in SO(2)$, so the usual planar rotation matrices are orientation-preserving orthogonal matrices. [/example] Rotations preserve orientation, but reflections preserve distances while reversing orientation. The next example gives the simplest reflection and also shows why determinant information is not redundant. [example: Coordinate Reflection] Let \begin{align*} S = \begin{pmatrix} 1 & 0 \cr 0 & -1 \end{pmatrix}. \end{align*} Since $S$ is diagonal, its transpose is itself: \begin{align*} S^\top = \begin{pmatrix} 1 & 0 \cr 0 & -1 \end{pmatrix}. \end{align*} Multiplying the two matrices entry by entry gives \begin{align*} S^\top S = \begin{pmatrix} 1\cdot 1 + 0\cdot 0 & 1\cdot 0 + 0\cdot(-1) \cr 0\cdot 1 + (-1)\cdot 0 & 0\cdot 0 + (-1)(-1) \end{pmatrix}. \end{align*} Thus \begin{align*} S^\top S = \begin{pmatrix} 1 & 0 \cr 0 & 1 \end{pmatrix}=I_2. \end{align*} Therefore $S \in O(2)$ by the definition of the orthogonal group. Its determinant is \begin{align*} \det S = 1\cdot(-1)-0\cdot 0=-1. \end{align*} Since $SO(2)=\{Q\in O(2):\det Q=1\}$, this shows $S\notin SO(2)$. For a vector $(x_1,x_2)$, multiplication gives \begin{align*} S\begin{pmatrix} x_1 \cr x_2 \end{pmatrix}=\begin{pmatrix} x_1 \cr -x_2 \end{pmatrix}, \end{align*} so the transformation fixes every point on the first coordinate axis and reverses the second coordinate. [/example] Not every matrix whose entries look symmetric or normalized is orthogonal. The definition requires all column lengths and all mutual dot products to have the correct values. [example: A Non-Orthogonal Shear] Let \begin{align*} A = \begin{pmatrix} 1 & 1 \cr 0 & 1 \end{pmatrix}. \end{align*} Then \begin{align*} A^\top = \begin{pmatrix} 1 & 0 \cr 1 & 1 \end{pmatrix}. \end{align*} Multiplying entries gives \begin{align*} A^\top A = \begin{pmatrix} 1\cdot 1+0\cdot 0 & 1\cdot 1+0\cdot 1 \cr 1\cdot 1+1\cdot 0 & 1\cdot 1+1\cdot 1 \end{pmatrix}. \end{align*} Thus \begin{align*} A^\top A = \begin{pmatrix} 1 & 1 \cr 1 & 2 \end{pmatrix}. \end{align*} Since this matrix has off-diagonal entries equal to $1$ and lower-right entry equal to $2$, it is not \begin{align*} I_2=\begin{pmatrix} 1 & 0 \cr 0 & 1 \end{pmatrix}. \end{align*} Therefore $A^\top A \ne I_2$, so $A$ is not orthogonal by the definition of an orthogonal matrix. The same failure appears metrically: \begin{align*} A\begin{pmatrix} 0 \cr 1 \end{pmatrix}=\begin{pmatrix} 1\cdot 0+1\cdot 1 \cr 0\cdot 0+1\cdot 1 \end{pmatrix}=\begin{pmatrix} 1 \cr 1 \end{pmatrix}. \end{align*} The original vector has Euclidean length \begin{align*} \left|\begin{pmatrix} 0 \cr 1 \end{pmatrix}\right|=\sqrt{0^2+1^2}=1, \end{align*} while its image has Euclidean length \begin{align*} \left|\begin{pmatrix} 1 \cr 1 \end{pmatrix}\right|=\sqrt{1^2+1^2}=\sqrt{2}. \end{align*} So this shear changes the length of the vector $(0,1)$, which is exactly the kind of Euclidean distortion orthogonal matrices exclude. [/example] Coordinate relabelling needs its own matrix model because it appears in combinatorics, numerical linear algebra, and symmetry arguments. These transformations move standard basis vectors around without changing their lengths or mutual orthogonality. The relevant matrices have a single $1$ in each row and column, so they encode permutations of coordinates. [definition: Permutation Matrix] Let $n \in \mathbb{N}$. A matrix $P \in \mathbb{R}^{n \times n}$ is a permutation matrix if every row and every column contains exactly one entry equal to $1$, and all other entries are $0$. [/definition] The definition of a permutation matrix is combinatorial, but many applications need to use coordinate relabelling inside Euclidean estimates or symmetry arguments. To make that passage legitimate, we need the metric consequence: relabelling the standard coordinate directions is an orthogonal transformation. [quotetheorem:8599] This embeds a finite combinatorial object into $O(n)$ and reminds us that the orthogonal group contains both discrete and continuous families. ## Properties The first structural property is closure under the natural operations for invertible linear transformations. If two transformations preserve Euclidean lengths and angles, their composition should also preserve them. The inverse transformation should undo the motion without introducing distortion. [quotetheorem:3268] This group structure means that composing two Euclidean symmetries gives another Euclidean symmetry, and reversing such a symmetry again preserves all lengths and angles. The determinant measures signed volume scaling, so it is the right invariant to test whether a length-preserving linear map can still stretch volume. Orthogonality rules out any genuine volume expansion or contraction, but it leaves one possible ambiguity: the map may preserve orientation or reverse it. The following result isolates exactly this obstruction by forcing the determinant to have absolute value $1$. [quotetheorem:8600] The sign of the determinant separates $O(n)$ into orientation-preserving and orientation-reversing pieces. This is the algebraic source of the definition of $SO(n)$. Analytic estimates often ask how much a linear map can enlarge a vector. Orthogonal matrices are the best-conditioned invertible matrices because they preserve every Euclidean length. This makes their operator norm especially simple. [quotetheorem:8601] This result is often used implicitly in estimates. If $x$ is perturbed by an error $e$, then $Qx$ is perturbed by $Qe$, whose length is exactly $|e|$. Measure and integration add another way to express the absence of distortion. A linear change of variables scales [Lebesgue measure](/page/Lebesgue%20Measure) by the absolute value of the determinant. Orthogonal matrices therefore preserve Lebesgue measure exactly. [quotetheorem:3300] This connects orthogonal matrices to integration: rotating or reflecting a domain does not change its Lebesgue measure. For instance, integrals of radial functions over balls are naturally invariant under $O(n)$. The measure-preservation statement controls volume, but spectral arguments ask a sharper question: what can an orthogonal matrix do along an invariant complex line? Since an eigenvalue records the scaling factor on such a line, length preservation should force those scalars to sit on the unit circle. [quotetheorem:8602] The eigenvalue constraint is only the first spectral hint. In applications, the more powerful question is whether a matrix can be simplified by a change of coordinates that does not distort lengths or angles. Real symmetric matrices are the central case: their eigenvectors can be chosen as an orthonormal basis, so diagonalisation happens through an orthogonal matrix rather than an arbitrary invertible one. [quotetheorem:925] The matrix $Q$ records an orthonormal eigenbasis, so the diagonal form is obtained without changing the Euclidean metric. Complex vector spaces require an analogue of orthogonality that respects the complex inner product. Since complex inner products involve conjugation, ordinary transpose is no longer the right operation. This motivates the unitary matrix as the complex counterpart of a real orthogonal matrix. [definition: Unitary Matrix] Let $n \in \mathbb{N}$. A matrix $U \in \mathbb{C}^{n \times n}$ is a unitary matrix if \begin{align*} U^*U = I_n, \end{align*} where $U^*$ denotes the conjugate transpose of $U$. [/definition] Because many spectral arguments move between real and complex vector spaces, we need a precise comparison rather than a slogan. For real matrices, conjugate transpose reduces to transpose, so the real orthogonal condition should be exactly the real case of unitarity. [quotetheorem:8603] This places $O(n)$ as the real counterpart of the unitary group $U(n)$. ## Relationship to Other Concepts Orthogonal matrices are the finite-dimensional real model for isometries fixing the origin. Once translations are allowed, the same matrices become the linear parts of all distance-preserving self-maps of Euclidean space. [quotetheorem:8604] The affine classification separates every Euclidean isometry into a translation part and a linear part. To connect the metric language of isometries with the matrix language of $O(n)$, we first isolate the origin-fixing case, where the translation has disappeared and the remaining transformation can be tested for orthogonality. [definition: Euclidean Isometry Fixing the Origin] Let $n \in \mathbb{N}$. A map $T: \mathbb{R}^n \to \mathbb{R}^n$ is a Euclidean isometry fixing the origin if $T(0)=0$ and $|T(x)-T(y)| = |x-y|$ for all $x,y \in \mathbb{R}^n$. [/definition] This origin-fixing condition removes the translation part of a rigid motion. What remains should be visible in coordinates as a matrix condition, and the next theorem identifies that condition with membership in $O(n)$. The geometric definition above applies to maps, while the original definition applies to matrices. A bridge theorem is needed so that statements about isometry can be translated into statements about $O(n)$. The standard basis provides the translation. [quotetheorem:8282] Quadratic forms often become understandable only after a good coordinate change. Orthogonal coordinate changes are preferred because they simplify the expression without changing Euclidean lengths. The next identity is the algebra behind principal axes and orthogonal diagonalisation. [quotetheorem:8605] This is the computational form of rotating coordinates to simplify a quadratic expression. It is the algebra underneath principal axes, covariance diagonalisation, and many energy estimates. The concept also appears in differential calculus. If a differentiable map has an orthogonal Jacobian matrix at a point, then its derivative preserves lengths of tangent vectors at that point. [example: Orthogonal Jacobian of a Planar Rotation] Let $f: \mathbb{R}^2 \to \mathbb{R}^2$ be given by \begin{align*} f(x_1,x_2)=(x_1\cos\theta-x_2\sin\theta,\ x_1\sin\theta+x_2\cos\theta) \end{align*} for a fixed $\theta\in\mathbb{R}$. Writing $f=(f_1,f_2)$, the four partial derivatives are \begin{align*} \frac{\partial f_1}{\partial x_1}=\cos\theta,\quad \frac{\partial f_1}{\partial x_2}=-\sin\theta,\quad \frac{\partial f_2}{\partial x_1}=\sin\theta,\quad \frac{\partial f_2}{\partial x_2}=\cos\theta. \end{align*} Hence the Jacobian matrix at every point $x\in\mathbb{R}^2$ is \begin{align*} Jf_x=\begin{pmatrix}\cos\theta&-\sin\theta\cr \sin\theta&\cos\theta\end{pmatrix}. \end{align*} Its transpose is \begin{align*} (Jf_x)^\top=\begin{pmatrix}\cos\theta&\sin\theta\cr -\sin\theta&\cos\theta\end{pmatrix}. \end{align*} Multiplying entry by entry gives \begin{align*} (Jf_x)^\top Jf_x=\begin{pmatrix}\cos\theta\cos\theta+\sin\theta\sin\theta&\cos\theta(-\sin\theta)+\sin\theta\cos\theta\cr (-\sin\theta)\cos\theta+\cos\theta\sin\theta&(-\sin\theta)(-\sin\theta)+\cos\theta\cos\theta\end{pmatrix}. \end{align*} The diagonal entries are $\cos^2\theta+\sin^2\theta=1$, and the off-diagonal entries are $-\cos\theta\sin\theta+\sin\theta\cos\theta=0$ and $-\sin\theta\cos\theta+\cos\theta\sin\theta=0$. Therefore \begin{align*} (Jf_x)^\top Jf_x=\begin{pmatrix}1&0\cr 0&1\end{pmatrix}=I_2. \end{align*} So $Jf_x\in O(2)$ for every $x\in\mathbb{R}^2$ by the definition of the orthogonal group. Thus the total derivative $Df_x$ is represented at every point by a matrix that preserves Euclidean lengths and inner products. [/example] This example is the local analytic shadow of rigid motion. In nonlinear problems, orthogonal Jacobians are closely related to local isometries, though global conclusions require additional hypotheses. ## Common Pitfalls A frequent mistake is to confuse orthogonality of a matrix with pairwise orthogonality of its entries or with symmetry. Orthogonality concerns dot products of rows or columns, not entrywise products. [example: Symmetric Does Not Mean Orthogonal] Let \begin{align*} A = \begin{pmatrix} 2 & 0 \cr 0 & 1/2 \end{pmatrix}. \end{align*} Since $A$ is diagonal, transposing it leaves its entries unchanged: \begin{align*} A^\top = \begin{pmatrix} 2 & 0 \cr 0 & 1/2 \end{pmatrix}=A. \end{align*} Thus $A$ is symmetric. To test whether $A$ is orthogonal, compute $A^\top A$: \begin{align*} A^\top A=\begin{pmatrix} 2 & 0 \cr 0 & 1/2 \end{pmatrix}\begin{pmatrix} 2 & 0 \cr 0 & 1/2 \end{pmatrix}. \end{align*} Entry-by-entry multiplication gives \begin{align*} A^\top A=\begin{pmatrix} 2\cdot 2+0\cdot 0 & 2\cdot 0+0\cdot(1/2) \cr 0\cdot 2+(1/2)\cdot 0 & 0\cdot 0+(1/2)(1/2) \end{pmatrix}. \end{align*} Therefore \begin{align*} A^\top A=\begin{pmatrix} 4 & 0 \cr 0 & 1/4 \end{pmatrix}. \end{align*} This is not \begin{align*} I_2=\begin{pmatrix} 1 & 0 \cr 0 & 1 \end{pmatrix}, \end{align*} because the upper-left entry is $4\ne 1$ and the lower-right entry is $1/4\ne 1$. Hence $A^\top A\ne I_2$, so $A$ is not orthogonal by the definition of an orthogonal matrix. Geometrically, $A(1,0)=(2,0)$ stretches the first coordinate direction, while $A(0,1)=(0,1/2)$ compresses the second coordinate direction; symmetry alone does not imply length preservation. [/example] Another mistake is to think determinant $\pm 1$ is sufficient. It is necessary for orthogonality, but it does not force length preservation. [example: Determinant One Is Not Enough] Let \begin{align*} A=\begin{pmatrix}1&1\cr 0&1\end{pmatrix}. \end{align*} Its determinant is \begin{align*} \det A=1\cdot 1-1\cdot 0=1. \end{align*} So $A$ has determinant one. To test orthogonality, compute \begin{align*} A^\top=\begin{pmatrix}1&0\cr 1&1\end{pmatrix}. \end{align*} Then \begin{align*} A^\top A=\begin{pmatrix}1&0\cr 1&1\end{pmatrix}\begin{pmatrix}1&1\cr 0&1\end{pmatrix}. \end{align*} Entry-by-entry multiplication gives \begin{align*} A^\top A=\begin{pmatrix}1\cdot 1+0\cdot 0&1\cdot 1+0\cdot 1\cr 1\cdot 1+1\cdot 0&1\cdot 1+1\cdot 1\end{pmatrix}. \end{align*} Therefore \begin{align*} A^\top A=\begin{pmatrix}1&1\cr 1&2\end{pmatrix}. \end{align*} This is not \begin{align*} I_2=\begin{pmatrix}1&0\cr 0&1\end{pmatrix}, \end{align*} because the off-diagonal entries are $1$ instead of $0$ and the lower-right entry is $2$ instead of $1$. Hence $A^\top A\ne I_2$, so $A$ is not orthogonal. The determinant records signed area scaling, while orthogonality requires preservation of every Euclidean length and angle. [/example] These failures help locate the definition precisely. Orthogonality is stronger than invertibility, stronger than determinant having absolute value $1$, and independent from symmetry. ## Beyond and Connections Orthogonal matrices are the finite-dimensional model for isometries of inner product spaces. The characterisation by inner products connects this page directly to [Linear Map](/page/Linear%20Map), because an orthogonal matrix is exactly a linear map that preserves the Euclidean inner product in orthonormal coordinates. It also explains why [Orthonormal Basis](/page/Orthonormal%20Basis) is the natural language for change of coordinates: switching between orthonormal bases should not alter lengths, angles, or quadratic expressions. The determinant split points toward orientation. Matrices in $SO(n)$ preserve both metric data and orientation, while the determinant-$-1$ part preserves metric data but reverses orientation. This distinction becomes important in geometry, where rotations and reflections have different roles even though both are distance-preserving. The spectral and quadratic-form viewpoints lead in two complementary directions. Spectrally, orthogonal matrices have eigenvalues constrained by length preservation, so they are built from rotations, reflections, and fixed directions. For quadratic forms and multivariable calculus, orthogonal diagonalisation is useful because it simplifies formulas without distorting the Euclidean geometry of the problem. ## References [Linear Map](/page/Linear%20Map). [Inner Product](/page/Inner%20Product). [Orthonormal Basis](/page/Orthonormal%20Basis). Axler, *Linear Algebra Done Right* (2015). Strang, *Linear Algebra and Its Applications* (2006). Horn and Johnson, *Matrix Analysis* (2012).