Matrix Groups: GL, SL, O, and U

What symmetries preserve a shape, a length, or a volume? This simple question opens the door to some of the richest infinite groups in all of mathematics. Once we know the Binary Operations and the Group Axioms that define an abstract group, the natural next step is clear. We should build large families of concrete examples from linear algebra. Matrices under multiplication supply exactly the right raw material. Matrix groups matter because they are simultaneously algebraic and geometric. A matrix is a symbol we can multiply, invert, and manipulate by hand. But it is also a linear transformation acting on a vector space. It stretches, rotates, or reflects that space. This double life makes matrix groups the bridge between pure group theory and geometry, physics, and computation. Every rotation of a rigid body is secretly an element of one of these groups. So is every quantum-mechanical symmetry, every linear code, and every computer graphics transformation. The big picture is a tower of increasingly restrictive conditions on matrices. We start with the general linear group GLnGL_n, all invertible matrices. We impose a determinant condition to get the special linear group SLnSL_n. We impose a length- or inner-product-preserving condition to get the orthogonal group O(n)O(n). Its complex analogue is the unitary group U(n)U(n). Each restriction cuts down a large group to a smaller, more rigid subgroup. This foreshadows the general theory of Subgroups and the Subgroup Criterion. This article defines all four families precisely. It proves that each is indeed a group under matrix multiplication. It explains the geometric meaning of each condition. It also works out concrete numerical examples in dimension 2 and 3. We also examine what breaks when hypotheses are dropped. This is the surest way to understand why each hypothesis is there in the first place.

Key takeaways
  • GLn(F)GL_n(F) is the group of all n×nn\times n invertible matrices over a field FF, under matrix multiplication.
  • SLn(F)⊂GLn(F)SL_n(F)\subset GL_n(F) is the subgroup of matrices with determinant 11.
  • O(n)O(n) is the group of real matrices preserving the standard dot product; equivalently ATA=IA^TA = I.
  • U(n)U(n) is the complex analogue, preserving the standard Hermitian inner product, satisfying A∗A=IA^*A=I.
  • Determinants of orthogonal and unitary matrices have absolute value 11, giving natural subgroups SO(n)SO(n) and SU(n)SU(n).
  • These are all infinite (for n≥1n\ge 1, n≥2n\ge2 resp.) non-abelian (for n≥2n\ge 2) groups, foreshadowing Group Homomorphisms via the determinant map.
  • Key idea: invertibility gives a group under multiplication; extra geometric conditions (determinant, length-preservation) carve out subgroups that are still groups because those conditions are compatible with multiplication and inversion.

Where this lives today

The general linear group GLnGL_n appears as the group of change-of-basis matrices in every linear algebra course and as the "ambient" group whose subgroups organize representation theory. The special linear group SLn(Z)SL_n(\mathbb{Z}) acts on lattices and modular forms, and is central to number theory (e.g., SL2(Z)SL_2(\mathbb{Z}) acting on the upper half plane defines the modular group underlying elliptic curves and modular forms). The orthogonal group O(3)O(3) and its subgroup SO(3)SO(3) are literally the group of rotations (and rotoreflections) of three-dimensional space, used throughout robotics, computer graphics, computer vision, and crystallography to encode rigid body orientation. The unitary group U(n)U(n), especially SU(2)SU(2) and SU(3)SU(3), is the mathematical backbone of the Standard Model of particle physics, where SU(3)×SU(2)×U(1)SU(3)\times SU(2)\times U(1) encodes the gauge symmetries of the strong, weak, and electromagnetic forces; U(n)U(n) also governs the evolution of quantum states in quantum computing, since unitary matrices are exactly the operations allowed on qubit states to preserve total probability. In cryptography, matrix groups over finite fields (finite analogues of GLn,SLnGL_n, SL_n) underlie certain public-key schemes and error-correcting codes, and in statistics, orthogonal matrices arise as the change-of-basis matrices diagonalizing covariance matrices in principal component analysis.

Ask the tutor: Why does the determinant map from O(n)O(n) to {±1}\{\pm1\} being a homomorphism guarantee SO(n)SO(n) is a subgroup, and how does this generalize to other matrix groups?

DefinitionThe four matrix groups

Let FF denote a field (for concreteness, take F=RF=\mathbb{R} or F=CF=\mathbb{C} unless otherwise stated), and let n≥1n\ge 1 be an integer.

General linear group.

GLn(F)={ A∈Mn(F):det⁡(A)≠0 },GL_n(F) = \{\, A \in M_n(F) : \det(A) \ne 0 \,\},

the set of n×nn\times n matrices over FF with nonzero determinant, under matrix multiplication.

Special linear group.

SLn(F)={ A∈GLn(F):det⁡(A)=1 }.SL_n(F) = \{\, A \in GL_n(F) : \det(A) = 1 \,\}.

Orthogonal group (real case, F=RF=\mathbb{R}). Let ATA^T denote the transpose of AA and II the identity matrix.

O(n)={ A∈Mn(R):ATA=I }.O(n) = \{\, A \in M_n(\mathbb{R}) : A^T A = I \,\}.

Unitary group (complex case, F=CF=\mathbb{C}). Let A∗=A‾TA^* = \overline{A}^T denote the conjugate transpose.

U(n)={ A∈Mn(C):A∗A=I }.U(n) = \{\, A \in M_n(\mathbb{C}) : A^* A = I \,\}.

In each case the binary operation is ordinary matrix multiplication, and we will show below that each set is closed under this operation, contains the identity matrix, contains inverses of its elements, and inherits associativity from Mn(F)M_n(F).

Intuition

Think of an n×nn\times n invertible matrix AA as a machine that takes column vectors in FnF^n and outputs new column vectors, v↦Avv \mapsto Av. The condition det⁡(A)≠0\det(A)\ne 0 says exactly that this machine is reversible: no information about vv is destroyed, because AA has a matrix inverse A−1A^{-1} undoing the transformation. This is the entire content of GLnGL_n: it is the group of all reversible linear transformations of FnF^n, full stop, with no further restriction. The special linear group asks for more: not just reversibility, but volume preservation. The determinant of AA measures the signed factor by which AA scales volumes (areas in dimension 2, volumes in dimension 3). Demanding det⁡(A)=1\det(A) = 1 means AA preserves both volume and orientation exactly, with no stretching or shrinking at all — think of shear transformations and rotations, but not dilations. The orthogonal and unitary groups impose a different, stronger geometric condition: preservation of length and angle, measured via an inner product. In Rn\mathbb{R}^n, the standard inner product is ⟨u,v⟩=uTv\langle u,v\rangle = u^Tv, and A∈O(n)A\in O(n) exactly when ⟨Au,Av⟩=⟨u,v⟩\langle Au, Av\rangle = \langle u,v\rangle for all u,vu,v — rotations and reflections, the rigid motions fixing the origin. In Cn\mathbb{C}^n, the analogous Hermitian inner product ⟨u,v⟩=u∗v\langle u,v\rangle = u^*v is preserved exactly by unitary matrices. Since preserving all lengths and angles is a much stronger demand than preserving volume alone, we will see O(n)⊊GLn(R)O(n) \subsetneq GL_n(\mathbb{R}) strictly and, in fact, that every orthogonal matrix has determinant ±1\pm1, so O(n)O(n) sits inside the volume-computation picture too, but as a much smaller, "rigid" subgroup.

Figure
The tower of matrix groups inside GLnGL_n, ordered by containment
GLₙ(F): det(A) ≠ 0 SLₙ(F): det(A) = 1 O(n): AᵀA = I SO(n): det = 1 U(n) is the analogous complex group, A*A = I
TheoremEach of GLn(F)GL_n(F), SLn(F)SL_n(F), O(n)O(n), U(n)U(n) is a group under matrix multiplication

Let FF be a field and n≥1n \ge 1 an integer. Then: 1. GLn(F)GL_n(F) is a group under matrix multiplication, with identity II and inverses given by matrix inversion. 2. SLn(F)SL_n(F) is a subgroup of GLn(F)GL_n(F): it is closed under multiplication, contains II, and is closed under inversion. 3. O(n)O(n), defined for F=RF = \mathbb{R}, is a subgroup of GLn(R)GL_n(\mathbb{R}). 4. U(n)U(n), defined for F=CF = \mathbb{C}, is a subgroup of GLn(C)GL_n(\mathbb{C}). Moreover, every element of O(n)O(n) has determinant ±1\pm 1, and every element of U(n)U(n) has determinant of absolute value 11; consequently O(n)⊆GLn(R)O(n) \subseteq GL_n(\mathbb{R}) and U(n)⊆GLn(C)U(n) \subseteq GL_n(\mathbb{C}) genuinely (both matrices with the required determinant condition satisfied automatically), so O(n)O(n) and U(n)U(n) are indeed subsets of SLnSL_n-adjacent territory but not contained in SLnSL_n in general (only their determinant-11 subgroups SO(n),SU(n)SO(n), SU(n) are).

Proof

We use the subgroup criterion technique: to show a nonempty subset HH of a group GG is a subgroup, it suffices to show HH is closed under the group operation and under taking inverses (closure under the operation, together with inverses, forces the identity to lie in HH as well, since for any h∈Hh\in H, h−1∈Hh^{-1}\in H by assumption and then hh−1=e∈Hhh^{-1}=e\in H by closure). We apply this after first establishing GLn(F)GL_n(F) itself is a group, since all other claims are subgroup claims relative to it.

Step 1: Mn(F)M_n(F) under multiplication is a monoid, and GLn(F)GL_n(F) is its group of units. Matrix multiplication on Mn(F)M_n(F) is associative: for A,B,C∈Mn(F)A,B,C \in M_n(F), the (i,j)(i,j) entry of (AB)C(AB)C and of A(BC)A(BC) both equal

∑k,lAikBklClj,\sum_{k,l} A_{ik}B_{kl}C_{lj},

by direct expansion of the double sum defining matrix multiplication (this is a finite algebraic identity in the field FF, using only commutativity and associativity of +,×+,\times in FF and the distributive law, and holds entrywise). The identity matrix II satisfies IA=AI=AIA = AI = A for all AA, by direct computation of matrix entries (Iik=δikI_{ik} = \delta_{ik}, so (IA)ij=∑kδikAkj=Aij(IA)_{ij} = \sum_k \delta_{ik}A_{kj} = A_{ij}, and symmetrically on the other side). Thus (Mn(F),⋅)(M_n(F), \cdot) is a monoid with identity II. Now recall the basic linear-algebra fact (Cramer/adjugate theory over a field): a matrix A∈Mn(F)A \in M_n(F) has a two-sided multiplicative inverse in Mn(F)M_n(F) if and only if det⁡(A)≠0\det(A) \ne 0, and in that case the inverse is unique and is given by

A−1=1det⁡(A) adj(A),A^{-1} = \frac{1}{\det(A)}\,\mathrm{adj}(A),

where adj(A)\mathrm{adj}(A) is the adjugate (transpose of the cofactor matrix), an expression built from FF-polynomial combinations of the entries of AA, hence itself lying in Mn(F)M_n(F). We take this classical fact from linear algebra as known (it does not depend on group theory, so invoking it here is not circular). Consequently GLn(F)GL_n(F), defined as {A:det⁡A≠0}\{A : \det A \ne 0\}, is precisely the set of units of the monoid Mn(F)M_n(F). It is a general fact, verified directly, that the set of units of any monoid forms a group under the monoid operation: it is closed under multiplication (if A,BA, B are units with inverses A−1,B−1A^{-1}, B^{-1}, then B−1A−1B^{-1}A^{-1} satisfies (AB)(B−1A−1)=A(BB−1)A−1=AIA−1=I(AB)(B^{-1}A^{-1}) = A(BB^{-1})A^{-1} = AIA^{-1}=I and similarly on the other side, so ABAB is a unit with inverse B−1A−1B^{-1}A^{-1}), contains the identity (II is its own inverse), and is closed under inversion (if AA is a unit with inverse A−1A^{-1}, then A−1A^{-1} is a unit with inverse AA). Associativity is inherited from Mn(F)M_n(F). Hence GLn(F)GL_n(F) is a group. This proves claim (1).

Step 2: det⁡\det is multiplicative, so SLn(F)SL_n(F) is a subgroup. The determinant satisfies det⁡(AB)=det⁡(A)det⁡(B)\det(AB) = \det(A)\det(B) for all A,B∈Mn(F)A, B \in M_n(F); this is a standard identity of linear algebra (provable via the Leibniz expansion of the determinant as a signed sum over permutations, or via elementary row operations), independent of group theory, and we invoke it as known. Since det⁡(I)=1≠0\det(I) = 1 \ne 0, we have I∈GLn(F)I \in GL_n(F), and clearly det⁡(I)=1\det(I) = 1 so I∈SLn(F)I \in SL_n(F); thus SLn(F)SL_n(F) is nonempty. Closure under multiplication: if A,B∈SLn(F)A, B \in SL_n(F), then det⁡(A)=det⁡(B)=1\det(A) = \det(B) = 1, so det⁡(AB)=det⁡(A)det⁡(B)=1⋅1=1\det(AB) = \det(A)\det(B) = 1\cdot 1 = 1, and since 1≠01 \ne 0, also AB∈GLn(F)AB \in GL_n(F). Hence AB∈SLn(F)AB \in SL_n(F). Closure under inversion: if A∈SLn(F)A \in SL_n(F), then A∈GLn(F)A \in GL_n(F), so A−1A^{-1} exists in GLn(F)GL_n(F), and from AA−1=IAA^{-1} = I we get det⁡(A)det⁡(A−1)=det⁡(I)=1\det(A)\det(A^{-1}) = \det(I) = 1; since det⁡(A)=1\det(A) = 1, this forces det⁡(A−1)=1\det(A^{-1}) = 1. Hence A−1∈SLn(F)A^{-1} \in SL_n(F). By the subgroup criterion, SLn(F)SL_n(F) is a subgroup of GLn(F)GL_n(F), proving claim (2).

Step 3: O(n)O(n) is a subgroup of GLn(R)GL_n(\mathbb{R}). First we check O(n)⊆GLn(R)O(n) \subseteq GL_n(\mathbb{R}): if ATA=IA^TA = I, then taking determinants of both sides (using det⁡(AT)=det⁡(A)\det(A^T) = \det(A), another standard fact of linear algebra) gives det⁡(A)2=det⁡(AT)det⁡(A)=det⁡(ATA)=det⁡(I)=1\det(A)^2 = \det(A^T)\det(A) = \det(A^TA) = \det(I) = 1, so det⁡(A)=±1≠0\det(A) = \pm 1 \ne 0, hence A∈GLn(R)A \in GL_n(\mathbb{R}). This simultaneously proves the claimed fact that every element of O(n)O(n) has determinant ±1\pm1. Nonemptiness: ITI=I⋅I=II^T I = I \cdot I = I, so I∈O(n)I \in O(n). Closure under multiplication: suppose A,B∈O(n)A, B \in O(n), so ATA=IA^TA = I and BTB=IB^TB=I. Using (AB)T=BTAT(AB)^T = B^TA^T (a standard identity: transpose reverses order of a product, provable entrywise, ((AB)T)ij=(AB)ji=∑kAjkBki=∑k(BT)ik(AT)kj=(BTAT)ij((AB)^T)_{ij} = (AB)_{ji} = \sum_k A_{jk}B_{ki} = \sum_k (B^T)_{ik}(A^T)_{kj} = (B^TA^T)_{ij}), we compute

(AB)T(AB)=BTATAB=BT(ATA)B=BTIB=BTB=I.(AB)^T(AB) = B^TA^TAB = B^T(A^TA)B = B^TIB = B^TB = I.

Hence AB∈O(n)AB \in O(n). Closure under inversion: let A∈O(n)A \in O(n). Since A∈GLn(R)A \in GL_n(\mathbb{R}) by the above, A−1A^{-1} exists. From ATA=IA^TA = I, multiply both sides on the left by (AT)−1(A^T)^{-1} and on the right appropriately: actually more directly, ATA=IA^TA = I means ATA^T is a left inverse of AA; since AA is invertible with unique two-sided inverse A−1A^{-1}, a left inverse must equal the two-sided inverse, so AT=A−1A^T = A^{-1}. Then

(A−1)TA−1=(AT)TA−1=AA−1=I,(A^{-1})^T A^{-1} = (A^T)^T A^{-1} = A A^{-1} = I,

using (AT)T=A(A^T)^T = A. Hence A−1∈O(n)A^{-1} \in O(n). By the subgroup criterion, O(n)O(n) is a subgroup of GLn(R)GL_n(\mathbb{R}), proving claim (3).

Step 4: U(n)U(n) is a subgroup of GLn(C)GL_n(\mathbb{C}), by an identical argument with conjugate transpose. The conjugate transpose satisfies (AB)∗=B∗A∗(AB)^* = B^*A^* (since ∗^* is transpose composed with entrywise complex conjugation, and both transpose-reversal and conjugation of a product being the product of conjugates in either order are standard facts: AB‾=A‾ B‾\overline{AB} = \overline{A}\,\overline{B} entrywise since complex conjugation is a ring homomorphism of C\mathbb{C}, combined with the transpose-reversal identity above) and (A∗)∗=A(A^*)^* = A. If A∗A=IA^*A = I, then taking determinants: det⁡(A∗)det⁡(A)=det⁡(I)=1\det(A^*)\det(A) = \det(I) = 1. Since det⁡(A∗)=det⁡(A)‾\det(A^*) = \overline{\det(A)} (conjugating every entry of AA before taking the determinant, a polynomial in the entries with integer coefficients, conjugates the determinant, as complex conjugation is a field automorphism of C\mathbb{C} fixing Z\mathbb{Z}), we get det⁡(A)‾det⁡(A)=∣det⁡(A)∣2=1\overline{\det(A)}\det(A) = |\det(A)|^2 = 1, so ∣det⁡(A)∣=1≠0|\det(A)| = 1 \ne 0; in particular A∈GLn(C)A \in GL_n(\mathbb{C}). This proves the claimed fact that every unitary matrix has determinant of absolute value 11. Nonemptiness: I∗I=II^*I = I, so I∈U(n)I \in U(n). Closure under multiplication: for A,B∈U(n)A, B \in U(n),

(AB)∗(AB)=B∗A∗AB=B∗IB=B∗B=I,(AB)^*(AB) = B^*A^*AB = B^*IB = B^*B = I,

so AB∈U(n)AB \in U(n). Closure under inversion: as before, A∗A=IA^*A = I with AA invertible forces A∗=A−1A^* = A^{-1} (a left inverse of an invertible matrix is the two-sided inverse), and then

(A−1)∗A−1=(A∗)∗A−1=AA−1=I,(A^{-1})^*A^{-1} = (A^*)^*A^{-1} = AA^{-1} = I,

so A−1∈U(n)A^{-1} \in U(n). By the subgroup criterion, U(n)U(n) is a subgroup of GLn(C)GL_n(\mathbb{C}), proving claim (4). All four claims are established, and associativity throughout is inherited automatically from the ambient monoid Mn(F)M_n(F), so no separate verification of associativity is needed for the subgroups. ■\blacksquare

∎

Worked examples

Example 1: An explicit element of GL2(R)∖SL2(R)GL_2(\mathbb{R}) \setminus SL_2(\mathbb{R}), and its inverse, showing the group operation concretely. Let

A=(2153).A = \begin{pmatrix} 2 & 1 \\ 5 & 3 \end{pmatrix}.

Compute det⁡(A)=2⋅3−1⋅5=6−5=1\det(A) = 2\cdot 3 - 1\cdot 5 = 6 - 5 = 1. So in fact det⁡(A)=1\det(A)=1 and A∈SL2(R)A \in SL_2(\mathbb{R}); let's instead perturb it slightly to get a genuine GL2∖SL2GL_2 \setminus SL_2 example and also illustrate inversion. Take

A=(2154),det⁡(A)=2⋅4−1⋅5=8−5=3.A = \begin{pmatrix} 2 & 1 \\ 5 & 4 \end{pmatrix}, \qquad \det(A) = 2\cdot4 - 1\cdot 5 = 8-5=3.

Since det⁡(A)=3≠0\det(A) = 3 \ne 0, A∈GL2(R)A \in GL_2(\mathbb{R}), but since 3≠13 \ne 1, A∉SL2(R)A \notin SL_2(\mathbb{R}). By the adjugate formula for 2×22\times 2 matrices, (abcd)−1=1ad−bc(d−b−ca)\begin{pmatrix}a&b\\c&d\end{pmatrix}^{-1} = \frac{1}{ad-bc}\begin{pmatrix}d & -b\\ -c& a\end{pmatrix}, so

A−1=13(4−1−52)=(4/3−1/3−5/32/3).A^{-1} = \frac{1}{3}\begin{pmatrix} 4 & -1 \\ -5 & 2 \end{pmatrix} = \begin{pmatrix} 4/3 & -1/3 \\ -5/3 & 2/3 \end{pmatrix}.

Check: AA−1=(2154)(4/3−1/3−5/32/3)AA^{-1} = \begin{pmatrix}2&1\\5&4\end{pmatrix}\begin{pmatrix}4/3 & -1/3\\-5/3 & 2/3\end{pmatrix}. Row 1: 2⋅43+1⋅(−53)=83−53=12\cdot\tfrac{4}{3} + 1\cdot(-\tfrac{5}{3}) = \tfrac{8}{3}-\tfrac{5}{3} = 1, and 2⋅(−13)+1⋅23=−23+23=02\cdot(-\tfrac13) + 1\cdot\tfrac23 = -\tfrac23+\tfrac23=0. Row 2: 5⋅43+4⋅(−53)=203−203=05\cdot\tfrac43 + 4\cdot(-\tfrac53) = \tfrac{20}{3}-\tfrac{20}{3}=0, and 5⋅(−13)+4⋅23=−53+83=15\cdot(-\tfrac13)+4\cdot\tfrac23 = -\tfrac53+\tfrac83 = 1. So indeed AA−1=IAA^{-1} = I, confirming A∈GL2(R)A \in GL_2(\mathbb{R}) with the exhibited inverse, consistent with Theorem: Each of GLnGL_n, SLnSL_n, O(n)O(n), U(n)U(n) is a group. Note also det⁡(A−1)=1/det⁡(A)=1/3\det(A^{-1}) = 1/\det(A) = 1/3, matching the general fact det⁡(A)det⁡(A−1)=1\det(A)\det(A^{-1})=1 used in the proof.

Example 2: A genuinely asymmetric element of O(3)O(3) and verification that it is orthogonal, plus its determinant. We build a rotation by angle θ\theta with cos⁡θ=3/5,sin⁡θ=4/5\cos\theta = 3/5, \sin\theta = 4/5 (a 33–44–55 triangle angle, not a "nice" angle like 90°90°) about the zz-axis, combined with a reflection in the third coordinate to make the example asymmetric and to illustrate an orthogonal matrix with determinant −1-1:

A=(3/5−4/504/53/5000−1).A = \begin{pmatrix} 3/5 & -4/5 & 0 \\ 4/5 & 3/5 & 0 \\ 0 & 0 & -1 \end{pmatrix}.

We verify ATA=IA^TA = I directly. First,

AT=(3/54/50−4/53/5000−1).A^T = \begin{pmatrix} 3/5 & 4/5 & 0 \\ -4/5 & 3/5 & 0 \\ 0 & 0 & -1 \end{pmatrix}.

Compute ATAA^TA entry by entry. Entry (1,1)(1,1): (3/5)(3/5)+(4/5)(4/5)+0=9/25+16/25=25/25=1(3/5)(3/5) + (4/5)(4/5) + 0 = 9/25+16/25 = 25/25=1. Entry (1,2)(1,2): (3/5)(−4/5)+(4/5)(3/5)+0=−12/25+12/25=0(3/5)(-4/5)+(4/5)(3/5)+0 = -12/25+12/25=0. Entry (1,3)(1,3): 0+0+0⋅(−1)=00+0+0\cdot(-1) = 0. Entry (2,2)(2,2): (−4/5)(−4/5)+(3/5)(3/5)=16/25+9/25=1(-4/5)(-4/5)+(3/5)(3/5) = 16/25+9/25=1. Entry (2,3)(2,3): 00. Entry (3,3)(3,3): (−1)(−1)=1(-1)(-1)=1. By symmetry of the computation (ATAA^TA is automatically symmetric since (ATA)T=ATA(A^TA)^T = A^TA), the remaining entries (2,1),(3,1),(3,2)(2,1),(3,1),(3,2) equal their transposed counterparts, already computed as 00. So ATA=IA^TA = I, confirming A∈O(3)A \in O(3). Now compute det⁡(A)\det(A). Expanding along the third row (which has entries 0,0,−10,0,-1):

det⁡(A)=−1⋅det⁡(3/5−4/54/53/5)=−1⋅(925+1625)=−1⋅1=−1.\det(A) = -1 \cdot \det\begin{pmatrix} 3/5 & -4/5 \\ 4/5 & 3/5\end{pmatrix} = -1\cdot\left(\frac{9}{25} + \frac{16}{25}\right) = -1 \cdot 1 = -1.

This matches the theorem: every element of O(n)O(n) has det⁡=±1\det = \pm 1, and here we exhibit the −1-1 case explicitly, so A∈O(3)∖SO(3)A \in O(3) \setminus SO(3). This shows O(3)O(3) is strictly larger than SO(3)=O(3)∩SL3(R)SO(3) = O(3)\cap SL_3(\mathbb{R}), and that the "reflection part" genuinely changes the determinant sign while orthogonality is preserved — a good check that the two conditions (determinant 11, versus ATA=IA^TA=I) are logically independent.

CounterexampleWhy invertibility (the hypothesis det⁡A≠0\det A \ne 0) cannot be dropped from GLnGL_n

The theorem crucially uses that GLn(F)GL_n(F) consists only of matrices with det⁡A≠0\det A \ne 0. Suppose instead we tried to form a "group" Mn(F)M_n(F) of all n×nn\times n matrices (dropping the invertibility restriction) under multiplication, claiming it is a group. This fails, and the failure is instructive. Take F=RF = \mathbb{R}, n=2n=2, and

B=(1000).B = \begin{pmatrix} 1 & 0 \\ 0 & 0 \end{pmatrix}.

Then det⁡(B)=0\det(B) = 0, so B∉GL2(R)B \notin GL_2(\mathbb{R}). We claim BB has no multiplicative inverse in M2(R)M_2(\mathbb{R}): if C=(abcd)C = \begin{pmatrix}a&b\\c&d\end{pmatrix} satisfied BC=IBC = I, then

BC=(1000)(abcd)=(ab00),BC = \begin{pmatrix}1&0\\0&0\end{pmatrix}\begin{pmatrix}a&b\\c&d\end{pmatrix} = \begin{pmatrix}a & b\\ 0&0\end{pmatrix},

which can never equal I=(1001)I = \begin{pmatrix}1&0\\0&1\end{pmatrix} since the (2,2)(2,2) entry of BCBC is always 0≠10 \ne 1. So BB has no inverse in M2(R)M_2(\mathbb{R}), and (M2(R),⋅)(M_2(\mathbb{R}), \cdot) fails the inverse axiom of a group. Even worse, note M2(R)M_2(\mathbb{R}) under multiplication is not even a group's worth of "cancellative" structure: taking C=(0001)≠0C = \begin{pmatrix}0&0\\0&1\end{pmatrix} \ne 0, we get BC=(1000)(0001)=(0000)BC = \begin{pmatrix}1&0\\0&0\end{pmatrix}\begin{pmatrix}0&0\\0&1\end{pmatrix} = \begin{pmatrix}0&0\\0&0\end{pmatrix}, the zero matrix, illustrating that Mn(F)M_n(F) has zero divisors, a phenomenon fundamentally incompatible with group structure (in a group, xy=exy = e with y≠e−1x−1y\ne e^{-1}x^{-1} combinations behaving like zero divisors cannot occur, since every nonidentity element still has a well-defined inverse and cancellation always holds). This shows precisely why the determinant condition det⁡(A)≠0\det(A) \ne 0 is not an artificial restriction but the exact condition needed to guarantee invertibility, and hence group structure.

Scope, edge cases & common traps

SituationWhat happensWhy
n=1n=1GL1(F)=F×GL_1(F) = F^\times (nonzero scalars), SL1(F)={1}SL_1(F) = \{1\}, O(1)={±1}O(1) = \{\pm1\}, U(1)={z∈C:∣z∣=1}U(1) = \{z\in\mathbb{C} : |z|=1\}, the circle groupDeterminant of a 1×11\times1 matrix is just its entry
FF a finite fieldGLn(F)GL_n(F) is finite, e.g. ∣GLn(Fq)∣=∏i=0n−1(qn−qi)|GL_n(\mathbb{F}_q)| = \prod_{i=0}^{n-1}(q^n - q^i)Orthogonal/unitary analogues over finite fields exist but require care with the bilinear/Hermitian form definition
Non-abelian for n≥2n\ge2GLn,SLn,O(n),U(n)GL_n, SL_n, O(n), U(n) are all non-abelian once n≥2n \ge 2 (with the trivial exception SL1SL_1)Matrix multiplication generally does not commute in dimension ≥2\ge 2
O(n)O(n) vs SO(n)SO(n)O(n)O(n) has two connected pieces (det =1=1 and det =−1=-1); only SO(n)=O(n)∩SLnSO(n) = O(n)\cap SL_n is path-connectedDeterminant is continuous and takes only values ±1\pm1 on O(n)O(n), so it separates O(n)O(n) into two disjoint pieces
Assuming O(n)⊂SLnO(n) \subset SL_nFalse in generalO(n)O(n) allows det⁡=−1\det = -1; only the determinant-1 elements of O(n)O(n) lie in SLnSL_n

The mathematician behind it

Origins · 1 minCamille Jordan and how this was discovered
CJ
Camille Jordan
1838–1922 · France, matrix and group theory

Camille Jordan was a French mathematician whose 1870 treatise Traité des substitutions et des équations algébriques systematized the study of permutation groups and linear substitutions, laying groundwork that would later crystallize into the modern theory of matrix groups such as GLnGL_n and SLnSL_n. Jordan's work bridged Galois's theory of equations with the emerging language of linear transformations, and the Jordan canonical form — a cornerstone tool for understanding conjugacy inside GLnGL_n — bears his name to this day.

History
- 1870: Camille Jordan publishes his Traité, systematically studying linear substitution groups, an early ancestor of GLnGL_n and SLnSL_n as studied objects.
- 1888–1897: Sophus Lie and Wilhelm Killing develop the theory of continuous transformation groups, providing the structural (Lie group) framework in which GLn(R)GL_n(\mathbb{R}), O(n)O(n), and U(n)U(n) are now understood as smooth manifolds with group structure.
- 1897: Élie Cartan begins his systematic classification of simple Lie groups, placing SO(n)SO(n), SU(n)SU(n), and related groups into the classical families An,Bn,Cn,DnA_n, B_n, C_n, D_n.
- 1905–1920s: The unitary group U(n)U(n) becomes central to the emerging mathematics of quantum mechanics, particularly through Hermann Weyl's work connecting representation theory of U(n)U(n) to atomic spectra.
- 1939: Weyl's book The Classical Groups gives the first modern, unified treatment of GLnGL_n, O(n)O(n), SpnSp_n, and U(n)U(n) as a coherent family, the direct ancestor of the exposition given here.
Is GLn(F)GL_n(F) abelian?

No, except in trivial cases. For n=1n = 1, GL1(F)=F×GL_1(F) = F^\times is abelian because multiplication in the field FF is commutative. But for n≥2n \ge 2, GLn(F)GL_n(F) is non-abelian: for instance in GL2(R)GL_2(\mathbb{R}), take A=(1101)A = \begin{pmatrix}1&1\\0&1\end{pmatrix} and B=(1011)B=\begin{pmatrix}1&0\\1&1\end{pmatrix}; direct computation gives AB=(2111)AB = \begin{pmatrix}2&1\\1&1\end{pmatrix} while BA=(1112)BA = \begin{pmatrix}1&1\\1&2\end{pmatrix}, which are different matrices. This non-commutativity is one reason matrix groups are such a rich source of examples in group theory: they exhibit genuinely non-abelian behavior even in low dimensions, unlike simpler examples such as cyclic groups.

Why do we require FF to be a field in these definitions?

The definitions of GLn,SLnGL_n, SL_n rely on the determinant formula for the inverse, A−1=det⁡(A)−1adj(A)A^{-1} = \det(A)^{-1}\mathrm{adj}(A), which requires dividing by det⁡(A)\det(A); this division only makes sense when det⁡(A)\det(A) has a multiplicative inverse, guaranteed for nonzero elements precisely when the scalars form a field. Over a general commutative ring RR, one instead defines GLn(R)GL_n(R) as matrices whose determinant is a unit of RR (not merely nonzero), and the same group-theoretic proof goes through verbatim with "unit" replacing "nonzero." Over Z\mathbb{Z}, for instance, GLn(Z)GL_n(\mathbb{Z}) consists of integer matrices with determinant exactly ±1\pm1, since those are the units of Z\mathbb{Z}.

Are O(n)O(n) and U(n)U(n) finite or infinite groups?

Both are infinite for n≥1n \ge 1 (for U(n)U(n)) and effectively for n≥1n\ge1 (for O(n)O(n), e.g. O(1)={±1}O(1)=\{\pm1\} is the only finite case). For n≥2n \ge 2, O(n)O(n) contains a rotation by every possible angle θ∈[0,2π)\theta \in [0, 2\pi), giving uncountably many distinct elements, and similarly U(n)U(n) contains uncountably many diagonal matrices diag(eiθ1,…,eiθn)\mathrm{diag}(e^{i\theta_1}, \dots, e^{i\theta_n}). These groups are also naturally topological spaces (compact manifolds), which is part of why they are studied as Lie groups, though that structure is extra data layered on top of the pure group structure discussed here.

How exactly does SO(n)SO(n) relate to O(n)O(n), and why is it called "special"?

SO(n)SO(n) is defined as SO(n)=O(n)∩SLn(R)SO(n) = O(n) \cap SL_n(\mathbb{R}), i.e., the orthogonal matrices with determinant exactly 11 (as opposed to −1-1). Since the determinant map det⁡:O(n)→{1,−1}\det: O(n) \to \{1,-1\} is a group homomorphism (a fact foreshadowing Group Homomorphisms, since det⁡(AB)=det⁡(A)det⁡(B)\det(AB)=\det(A)\det(B)) onto the two-element group {±1}\{\pm1\}, and SO(n)SO(n) is exactly its kernel, SO(n)SO(n) is automatically a subgroup of O(n)O(n) of index 22 (assuming n≥1n\ge1 so that some orthogonal matrix has determinant −1-1, e.g. a reflection). Geometrically, SO(n)SO(n) consists precisely of the rotations, while O(n)∖SO(n)O(n)\setminus SO(n) consists of orientation-reversing maps like reflections.

Does every matrix in GLn(F)GL_n(F) preserve some geometric structure, the way O(n)O(n) preserves length?

Not necessarily in the strong "isometry" sense, but every A∈GLn(F)A \in GL_n(F) does preserve linear independence and hence dimension: since AA is invertible, Av1,…,AvkAv_1,\dots,Av_k are linearly independent in FnF^n whenever v1,…,vkv_1,\dots,v_k are. That is the entire content of invertibility — it is a purely algebraic/linear condition, not a metric one. SLnSL_n additionally preserves volume (up to sign in odd characteristic considerations, but exactly for det⁡=1\det=1), while O(n)O(n) and U(n)U(n) preserve the much finer metric structure of lengths and angles. So the four groups form a genuine hierarchy of "how much structure is preserved," from purely linear (GLnGL_n) to volume (SLnSL_n) to full metric rigidity (O(n)O(n), U(n)U(n)).

Can I check membership in O(n)O(n) using rows instead of columns?

Yes, and this gives a useful sanity check. The condition ATA=IA^TA=I is equivalent to AAT=IAA^T = I (since in a group, a two-sided inverse is unique, and ATA=IA^TA=I combined with AA invertible forces AAT=IAA^T=I as shown in the proof). Expanding ATA=IA^TA=I entrywise says the columns of AA form an orthonormal set with respect to the standard dot product; expanding AAT=IAA^T=I says the same about the rows. So you may check either the columns or the rows are pairwise orthogonal unit vectors — both give equivalent, correct verifications of orthogonality.

What is the smallest example showing SLnSL_n is a proper subgroup of GLnGL_n?

Take n=1n=1 over R\mathbb{R}: GL1(R)=R×GL_1(\mathbb{R}) = \mathbb{R}^\times, all nonzero reals under multiplication, while SL1(R)={1}SL_1(\mathbb{R}) = \{1\}, since the "determinant" of a 1×11\times1 matrix (a)(a) is just aa itself, and we need a=1a=1. So already at n=1n=1, SL1(R)SL_1(\mathbb{R}) is the trivial group while GL1(R)GL_1(\mathbb{R}) is infinite — an extreme but perfectly valid illustration that the determinant condition is a genuine, non-vacuous restriction, before even considering the more elaborate n≥2n\ge2 pictures used elsewhere in this article.

Before this

Where this leads