01.01.10 · foundations / linear-algebra

Adjoint operator and isometry on a finite-dimensional inner-product space

shipped3 tiersLean: none

Anchor (Master): Shilov *Linear Algebra* Ch. 9 §9.1–§9.2; Halmos *Finite-Dimensional Vector Spaces* §44–§74; Axler *Linear Algebra Done Right* §7.A + §7.D; Lang *Algebra* Ch. XV; Reed-Simon *Methods of Modern Mathematical Physics* Vol. I §VI for the Hilbert-space adjoint

Intuition Beginner

The dot product lets you measure lengths and angles. Once a linear map $T$ is acting on vectors, a natural question is: if you push a vector through $T$ and then dot it against a second vector, can you instead leave the first vector alone and push the second one through some companion map? The answer is yes, and that companion map is called the adjoint of $T$ , written $T^{*}$ . It is the map that lets you slide $T$ from one side of the dot product to the other.

When you write $T$ as a matrix using a perpendicular, unit-length coordinate system, the adjoint has a beautifully simple description: you flip the matrix across its diagonal and (for complex numbers) flip the sign of every imaginary part. For real matrices this is just the transpose. So the abstract idea of "the map on the other side of the dot product" turns into a one-line recipe on the entries.

A second character in this unit is the isometry: a linear map that does not change any length or any angle. Rotations and reflections are the basic examples. An isometry takes a perpendicular unit-length frame to another perpendicular unit-length frame, so it moves the whole space rigidly while pinning the origin. The two ideas connect: a map is an isometry exactly when applying its adjoint undoes it.

Visual Beginner

The figure shows the adjoint as a mirror across the dot product, and an isometry as a rigid rotation of a square grid. On the left, two vectors and two ways to compute the same number: dot $T v$ against $w$ , or dot $v$ against $T^{*} w$ . Both routes give the same value, which is what defines $T^{*}$ . On the right, a rotation carries the standard unit square to a tilted unit square of the same size, with every length and every right angle preserved.

The two panels capture the whole unit. The adjoint is the bookkeeping device that moves a map across the dot product. The isometry is the kind of map that the dot product cannot tell apart from the identity, because it leaves every measurement unchanged.

Worked example Beginner

Take the real matrix $A = (1324)$ acting on the plane with the ordinary dot product. Its adjoint is the transpose $A^{*} = (1234)$ .

Check the defining property on two specific vectors. Let $v = (1, 0)$ and $w = (0, 1)$ .

Step 1. Compute $A v$ . Multiplying, $A v = (1, 3)$ .

Step 2. Dot $A v$ with $w$ . We get $(1, 3) \cdot (0, 1) = 0 + 3 = 3$ .

Step 3. Compute $A^{*} w$ . Multiplying, $A^{*} w = (3, 4)$ .

Step 4. Dot $v$ with $A^{*} w$ . We get $(1, 0) \cdot (3, 4) = 3 + 0 = 3$ .

Both routes give $3$ . The number you get by dotting $A v$ against $w$ equals the number you get by dotting $v$ against $A^{*} w$ . That equality, holding for every choice of $v$ and $w$ , is exactly what it means for $A^{*}$ to be the adjoint of $A$ .

What this tells us: the transpose is not just a symbol-shuffling operation. It is the matrix that answers "what map sits on the other side of the dot product?" For a real matrix in standard coordinates, that map is the transpose, and the worked equality $3 = 3$ is one instance of the rule.

Check your understanding Beginner

Exercise (easy, multiple choice).

For a real matrix $A$ in standard coordinates, the adjoint $A^{*}$ is:

A. The inverse of $A$ B. The transpose of $A$ C. The matrix $A$ with its diagonal deleted D. The negative of $A$

Hint

The adjoint flips the matrix across its main diagonal. Over the real numbers there is no imaginary part to conjugate.

Answer

B. The transpose of $A$ . For a real matrix in an orthonormal (perpendicular, unit-length) coordinate system, the adjoint is the transpose. Over the complex numbers it is the conjugate transpose, which also flips the sign of imaginary parts. Feedback-correct: the adjoint is the map on the other side of the dot product, and in standard real coordinates that is the transpose. Feedback-wrong: the inverse and negative are different operations and need not equal the transpose.

Formal definition Intermediate+

Let $F \in {R, C}$ and let $(V, ⟨ \cdot, \cdot ⟩)$ be a finite-dimensional inner-product space over $F$ in the sense of 01.01.09, with the inner product linear in the first slot and conjugate-linear in the second (the mathematician convention; the conjugate is the identity over $R$ ). Write $∥ v ∥ = ⟨ v, v ⟩$ for the induced norm.

Definition (adjoint). Let $T : V \to V$ be linear. The adjoint of $T$ is the linear operator $T^{*} : V \to V$ satisfying $$ \langle T v, w \rangle = \langle v, T^* w \rangle \qquad \text{for all } v, w \in V. $$ Such an operator exists and is unique; existence is the content of the key theorem below. An operator is self-adjoint (Hermitian over $C$ , symmetric over $R$ ) when $T = T^{*}$ , and anti-self-adjoint (skew-Hermitian / skew-symmetric) when $T = - T^{*}$ .

Definition (isometry). A linear operator $U : V \to V$ is an isometry when $$ \langle U v, U w \rangle = \langle v, w \rangle \qquad \text{for all } v, w \in V. $$ Over $R$ an isometry is called an orthogonal operator; over $C$ , a unitary operator. The isometries of $V$ form a group under composition: the orthogonal group $O (V) ≅ O (n)$ when $F = R$ and $dim V = n$ , and the unitary group $U (V) ≅ U (n)$ when $F = C$ .

Three notational conventions used below. The conjugate transpose (Hermitian transpose) of a matrix $A = (a_{ij})$ is $A^{*} = \overline{A}^{⊤}$ , with entries $(A^{*})_{ij} = \overline{a_{j i}}$ ; over $R$ it is the transpose $A^{⊤}$ . The Kronecker delta $δ_{ij}$ is $1$ if $i = j$ and $0$ otherwise. The orthogonal complement of a subset $S \subseteq V$ is $S^{⊥} = {v \in V : ⟨ v, s ⟩ = 0 for all s \in S}$ .

Counterexamples to common slips

The adjoint depends on the inner product, not on $T$ alone. The same linear operator has different adjoints under different inner products: if $G$ is the positive-definite Gram matrix of the inner product in a fixed basis and $A$ is the matrix of $T$ , the matrix of $T^{*}$ in that basis is $G^{- 1} A^{*} G$ , which reduces to $A^{*}$ only when $G = I$ , i.e., when the basis is orthonormal.
Norm preservation $∥ U v ∥ = ∥ v ∥$ for all $v$ is equivalent to inner-product preservation, but only because the polarisation identity recovers the inner product from the norm. Over a general (non-inner-product) normed space the two conditions diverge; the equivalence is special to inner-product spaces.
An operator with $U^{*} U = I$ on a finite-dimensional space automatically satisfies $U U^{*} = I$ as well, so it is invertible with $U^{- 1} = U^{*}$ . This collapse of one-sided into two-sided invertibility uses finite-dimensionality; on an infinite-dimensional Hilbert space the unilateral shift satisfies $U^{*} U = I$ but $U U^{*} \neq = I$ , and is an isometry that is not unitary.

Key theorem with proof Intermediate+

Theorem (existence, uniqueness, and matrix of the adjoint; Shilov §9.1 ^{[source pending]}, Halmos §44 ^{[source pending]}). Let $(V, ⟨ \cdot, \cdot ⟩)$ be a finite-dimensional inner-product space over $F \in {R, C}$ and $T : V \to V$ linear. There is a unique linear operator $T^ : V \to V $w i t h$ \langle T v, w \rangle = \langle v, T^* w \rangle $f or a l l$ v, w \in V $. I f$ {e_1, \ldots, e_n} $i s an or t h o n or ma l ba s i s an d$ A = (a_{ij}) $i s t h e ma t r i x o f$ T $in t hi s ba s i s, t h e n t h e ma t r i x o f$ T^ $in t h es am e ba s i s i s t h eco nj ug a t e t r an s p ose$ A^ = \overline{A}^\top$.*

Proof. Existence and uniqueness. Fix $w \in V$ and define $φ_{w} : V \to F$ by $φ_{w} (v) = ⟨ T v, w ⟩$ . Linearity of $T$ and linearity of the inner product in its first slot make $φ_{w}$ a linear functional on $V$ . By the Riesz representation theorem on a finite-dimensional inner-product space ^{[source pending]}, every linear functional $φ$ on $V$ is $φ (v) = ⟨ v, u ⟩$ for a unique $u \in V$ . Apply this to $φ_{w}$ : there is a unique vector, which we name $T^{*} w$ , with $$ \langle T v, w \rangle = \varphi_w(v) = \langle v, T^* w \rangle \qquad \text{for all } v \in V. $$ This defines a map $w \mapsto T^{*} w$ .

To see it is linear, fix $w_{1}, w_{2} \in V$ and $α, β \in F$ . For every $v$ , $$ \langle v, T^(\alpha w_1 + \beta w_2) \rangle = \langle T v, \alpha w_1 + \beta w_2 \rangle = \overline{\alpha} \langle T v, w_1 \rangle + \overline{\beta} \langle T v, w_2 \rangle = \langle v, \alpha T^ w_1 + \beta T^* w_2 \rangle, $$ using conjugate-linearity of the inner product in the second slot twice. Since this holds for all $v$ and the inner product is non-degenerate, $T^{*} (α w_{1} + β w_{2}) = α T^{*} w_{1} + β T^{*} w_{2}$ , so $T^{*}$ is linear. Uniqueness of $T^{*}$ follows from uniqueness in the Riesz step applied at each $w$ : if $S$ and $T^{*}$ both satisfy the defining identity, then $⟨ v, (S - T^{*}) w ⟩ = 0$ for all $v$ , forcing $(S - T^{*}) w = 0$ for all $w$ .

Matrix form. Let ${e_{1}, \dots, e_{n}}$ be orthonormal, so $⟨ e_{i}, e_{j} ⟩ = δ_{ij}$ . The matrix entries of $T$ are recovered by $a_{ij} = ⟨ T e_{j}, e_{i} ⟩$ , since $T e_{j} = \sum_{i} a_{ij} e_{i}$ and orthonormality isolates the $i$ -th coefficient. Let $b_{ij} = ⟨ T^{*} e_{j}, e_{i} ⟩$ be the matrix entries of $T^{*}$ . Using the defining identity with $v = e_{i}$ and $w = e_{j}$ , $$ b_{ij} = \langle T^* e_j, e_i \rangle = \overline{\langle e_i, T^* e_j \rangle} = \overline{\langle T e_i, e_j \rangle} = \overline{a_{ji}}. $$ The first equality is the entry formula for $T^{*}$ ; the second is conjugate symmetry of the inner product; the third is the defining identity $⟨ e_{i}, T^{*} e_{j} ⟩ = ⟨ T e_{i}, e_{j} ⟩$ ; the fourth is the entry formula for $T$ with the roles of $i, j$ swapped. Hence $b_{ij} = \overline{a_{j i}}$ , which is precisely the $(i, j)$ entry of the conjugate transpose $A^{*} = \overline{A}^{⊤}$ . Over $R$ the conjugation is the identity and $A^{*} = A^{⊤}$ . $□$

Bridge. The adjoint builds toward the entire operator-theoretic structure of an inner-product space, and it appears again in 01.01.13 (the spectral theorem), where the normality condition $T T^{*} = T^{*} T$ is the exact hypothesis that produces an orthonormal eigenbasis. The foundational mechanism is the Riesz representation theorem: it is the device that converts the linear functional $w \mapsto ⟨ T v, w ⟩$ into an honest vector, and without it the adjoint would be only formally defined. The central computation is that in an orthonormal basis the adjoint is conjugate transposition, so the abstract identity $⟨ T v, w ⟩ = ⟨ v, T^{*} w ⟩$ and the concrete operation $A \mapsto \overline{A}^{⊤}$ are two faces of one object. Putting these together, the adjoint controls the four fundamental subspaces through $ker T^{*} = (im T)^{⊥}$ , supplies the self-adjoint operators whose real spectrum and orthogonal eigenspaces drive the spectral theorem, and supplies the isometries via the single relation $U^{*} U = I$ . This bridge continues into 01.01.12 (singular value decomposition), where the adjoint appears through $T^{*} T$ , and into the polar decomposition $T = U ∣ T ∣$ that factors any operator into an isometry times a positive self-adjoint part.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Verify that the rotation matrix $U = (cos θ sin θ - sin θ cos θ)$ satisfies $U^{*} U = I$ , and conclude $U$ is an isometry of $R^{2}$ .

Hint

Over $R$ , $U^{*} = U^{⊤}$ . Multiply $U^{⊤} U$ and use $cos^{2} θ + sin^{2} θ = 1$ .

Answer

The transpose is $U^{⊤} = (cos θ - sin θ sin θ cos θ)$ . Then $$ U^\top U = \begin{pmatrix} \cos\theta & \sin\theta \ -\sin\theta & \cos\theta \end{pmatrix} \begin{pmatrix} \cos\theta & -\sin\theta \ \sin\theta & \cos\theta \end{pmatrix} = \begin{pmatrix} \cos^2\theta + \sin^2\theta & 0 \ 0 & \sin^2\theta + \cos^2\theta \end{pmatrix} = \begin{pmatrix} 1 & 0 \ 0 & 1 \end{pmatrix}. $$ The off-diagonal entries are $- cos θ sin θ + sin θ cos θ = 0$ and its mirror. So $U^{⊤} U = I$ , and by the equivalence theorem $U$ preserves the inner product, hence is an isometry. Rubric: full credit for the matrix product with the Pythagorean simplification.

Exercise 4 (medium, short-answer).

Prove the order-reversal law $(S T)^{*} = T^{*} S^{*}$ for operators $S, T$ on a finite-dimensional inner-product space.

Hint

Apply the defining identity $⟨(S T) v, w ⟩ = ⟨ v, (S T)^{*} w ⟩$ and also move $S$ then $T$ across the inner product one at a time.

Answer

For all $v, w \in V$ , slide $S$ then $T$ across the inner product: $$ \langle S T v, w \rangle = \langle T v, S^* w \rangle = \langle v, T^* S^* w \rangle. $$ The first step uses the defining identity for $S^{*}$ (with $T v$ in the first slot); the second uses the defining identity for $T^{*}$ . On the other hand, by definition of $(S T)^{*}$ , $⟨ S T v, w ⟩ = ⟨ v, (S T)^{*} w ⟩$ . Subtracting, $⟨ v, [(S T)^{*} - T^{*} S^{*}] w ⟩ = 0$ for all $v$ , so non-degeneracy forces $(S T)^{*} w = T^{*} S^{*} w$ for all $w$ , i.e., $(S T)^{*} = T^{*} S^{*}$ . Rubric: full credit for the two-step slide plus the non-degeneracy cancellation.

Exercise 7 (hard, short-answer).

Prove that for a self-adjoint operator $T$ , eigenvectors belonging to distinct eigenvalues are orthogonal.

Hint

For $T v_{1} = λ_{1} v_{1}$ and $T v_{2} = λ_{2} v_{2}$ , compute $⟨ T v_{1}, v_{2} ⟩$ two ways and use that $λ_{1}, λ_{2}$ are real.

Answer

Let $T v_{1} = λ_{1} v_{1}$ and $T v_{2} = λ_{2} v_{2}$ with $λ_{1} \neq = λ_{2}$ . The eigenvalues of a self-adjoint operator are real (proved in the Advanced results), so $\overline{λ_{2}} = λ_{2}$ . Compute $$ \lambda_1 \langle v_1, v_2 \rangle = \langle \lambda_1 v_1, v_2 \rangle = \langle T v_1, v_2 \rangle = \langle v_1, T^* v_2 \rangle = \langle v_1, T v_2 \rangle = \langle v_1, \lambda_2 v_2 \rangle = \overline{\lambda_2} \langle v_1, v_2 \rangle = \lambda_2 \langle v_1, v_2 \rangle. $$ The third equality uses $T = T^{*}$ ; the conjugate-linearity in the second slot turns $λ_{2}$ into $\overline{λ_{2}}$ , which equals $λ_{2}$ by reality. Subtracting, $(λ_{1} - λ_{2}) ⟨ v_{1}, v_{2} ⟩ = 0$ , and since $λ_{1} \neq = λ_{2}$ we conclude $⟨ v_{1}, v_{2} ⟩ = 0$ . Rubric: full credit for the equation chain plus explicit use of reality of the eigenvalues.

Exercise 8 (hard, short-answer).

Prove that the columns of a unitary matrix $U \in M_{n} (C)$ form an orthonormal basis of $C^{n}$ , and conversely.

Hint

The $(i, j)$ entry of $U^{*} U$ is the Hermitian inner product of column $i$ with column $j$ .

Answer

Write $U = [c_{1} ∣ \dots ∣ c_{n}]$ with columns $c_{k} \in C^{n}$ . The $(i, j)$ entry of $U^{*} U$ is $\sum_{k} \overline{(U)_{k i}} (U)_{k j} = ⟨ c_{j}, c_{i} ⟩$ in the standard Hermitian inner product (entry-wise, $(U^{*} U)_{ij} = \sum_{k} \overline{u_{k i}} u_{k j}$ ). The condition $U^{*} U = I$ reads $(U^{*} U)_{ij} = δ_{ij}$ , i.e., $⟨ c_{j}, c_{i} ⟩ = δ_{ij}$ : the columns are orthonormal. Since there are $n$ of them and they are orthonormal, they are linearly independent (an orthonormal set is independent, 01.01.09), hence a basis of the $n$ -dimensional space $C^{n}$ . Conversely, if the columns form an orthonormal basis then $(U^{*} U)_{ij} = ⟨ c_{j}, c_{i} ⟩ = δ_{ij}$ , so $U^{*} U = I$ and $U$ is unitary (in finite dimensions $U^{*} U = I$ gives $U U^{*} = I$ as well). The real-orthogonal case is identical with $U^{*} = U^{⊤}$ . Rubric: full credit for the column-inner-product identification of $U^{*} U$ plus the independence-to-basis step in both directions.

Lean formalization Intermediate+

Mathlib carries the adjoint and the isometry as named objects, but no single theorem packages the orthonormal-basis matrix correspondence or the three-way isometry equivalence used in this unit. The readable view of the statements:

import Mathlib.Analysis.InnerProductSpace.Adjoint
import Mathlib.Analysis.InnerProductSpace.Basic

variable {𝕜 V : Type*} [RCLike 𝕜]
  [NormedAddCommGroup V] [InnerProductSpace 𝕜 V] [FiniteDimensional 𝕜 V]

/-- Defining property of the adjoint: ⟪T x, y⟫ = ⟪x, T† y⟫.
Mathlib: `LinearMap.adjoint_inner_left`. -/
example (T : V →ₗ[𝕜] V) (x y : V) :
    inner (T x) y = inner x (LinearMap.adjoint T y) :=
  (LinearMap.adjoint_inner_left T x y).symm

/-- The matrix of the adjoint in an orthonormal basis is the conjugate
transpose. This wiring is NOT a single named Mathlib lemma. -/
theorem adjoint_toMatrix_conjTranspose
    {n : ℕ} (b : OrthonormalBasis (Fin n) 𝕜 V) (T : V →ₗ[𝕜] V) :
    (LinearMap.toMatrix b.toBasis b.toBasis (LinearMap.adjoint T))
      = (LinearMap.toMatrix b.toBasis b.toBasis T).conjTranspose :=
  sorry  -- expand both sides via b.repr and adjoint_inner_left

/-- Three-way isometry equivalence via polarisation:
U†U = 1 ↔ inner-product preserving ↔ norm preserving. The full
biconditional chain is not a single named Mathlib theorem. -/
theorem isometry_tfae (U : V →ₗ[𝕜] V) :
    (LinearMap.adjoint U * U = 1) ↔ (∀ x, ‖U x‖ = ‖x‖) :=
  sorry  -- (→) expand ‖U x‖² = ⟪U†U x, x⟫; (←) polarisation identity

The proof gap to a clean contribution: (i) prove adjoint_toMatrix_conjTranspose from LinearMap.adjoint_inner_left and the orthonormal b.repr coordinate formula; (ii) prove the polarisation step recovering the inner product from the norm in both $R$ and $C$ to close isometry_tfae; (iii) wire ker_adjoint = (range _)ᗮ to the rank-nullity refinement. Each step is reachable from existing Mathlib infrastructure, but no current named theorem presents the matrix correspondence or the polarisation equivalence in this packaged form.

Advanced results Master

Theorem (polarisation identity and the three characterisations of an isometry; Halmos §72–§73 ^{[source pending]}). Let $U : V \to V$ be linear on a finite-dimensional inner-product space. The following are equivalent: (i) $U^ U = I $; (ii)$ \langle U v, U w \rangle = \langle v, w \rangle $f or a l l$ v, w $; (iii)$ |U v| = |v| $f or a l l$ v$. On a complex space the inner product is recovered from the norm by* $$ \langle v, w \rangle = \tfrac{1}{4} \sum_{k=0}^{3} i^k , |v + i^k w|^2, $$ and on a real space by $⟨ v, w ⟩ = \frac{1}{4} (∥ v + w ∥^{2} - ∥ v - w ∥^{2})$ .

The equivalence (i) $\Leftrightarrow$ (ii) is the adjoint definition: $⟨ U v, U w ⟩ = ⟨ v, U^{*} U w ⟩$ equals $⟨ v, w ⟩$ for all $v, w$ exactly when $U^{*} U = I$ . The implication (ii) $\Rightarrow$ (iii) is the case $v = w$ . The reverse (iii) $\Rightarrow$ (ii) is the polarisation identity: the inner product is a fixed linear combination of squared norms, so any norm-preserving map preserves the inner product. The full chain is recorded in the proof set below.

Theorem (spectral consequences of self-adjointness; Hilbert 1906 ^{[source pending]}). A self-adjoint operator $T = T^ $o na f ini t e - d im e n s i o na l inn er - p r o d u c t s p a ce ha sr e a l s p ec t r u m, an d e i g e n v ec t or s a t d i s t in c t e i g e n v a l u es a r eor t h o g o na l . A u ni t a r y o p er a t or$ U $ha ss p ec t r u m o n t h e u ni t c i r c l e$ {|\lambda| = 1} $, an d$ |\det U| = 1$.*

Self-adjointness forces $⟨ T v, v ⟩ = ⟨ v, T v ⟩ = \overline{⟨ T v, v ⟩}$ , so $⟨ T v, v ⟩ \in R$ for all $v$ , and applying this to an eigenvector gives $λ ∥ v ∥^{2} \in R$ , hence $λ \in R$ . The eigenspace orthogonality is the computation in Exercise 7. For a unitary $U$ , an eigenvalue $λ$ satisfies $∥ v ∥ = ∥ U v ∥ = ∣ λ ∣ ∥ v ∥$ , so $∣ λ ∣ = 1$ ; the determinant is the product of eigenvalues over $C$ , each of modulus $1$ , so $∣ det U ∣ = 1$ . Over $R$ , $det U = \pm 1$ separates rotations ( $SO (n)$ , determinant $+ 1$ ) from orientation-reversing isometries (determinant $- 1$ ). These facts are the entry point to the spectral theorem 01.01.13, which upgrades them to a full orthonormal eigenbasis under the normality hypothesis $T T^{*} = T^{*} T$ .

Theorem (polar decomposition; Axler §7.D ^{[source pending]}). Every operator $T$ on a finite-dimensional inner-product space factors as $T = U ∣ T ∣$ , where $|T| = (T^ T)^{1/2} $i s t h e u ni q u e p os i t i v ese mi d e f ini t es q u a r er oo t o f t h e p os i t i v ese mi d e f ini t eo p er a t or$ T^* T $, an d$ U $i s ani so m e t r y o n$ \operatorname{im}|T| $e x t e n d e d t o ani so m e t r y o f$ V $. W h e n$ T $i s in v er t ib l e,$ U $i s u ni q u e l y d e t er min e d an d u ni t a r y; in g e n er a l$ U $i s u ni q u eo n$ (\ker T)^\perp$.*

The polar decomposition is the operator analogue of the polar form $z = r e^{i θ}$ of a complex number, with $∣ T ∣$ the radial (positive self-adjoint) part and $U$ the angular (isometric) part. It is the structural bridge from this unit to the singular value decomposition 01.01.12: diagonalising the positive self-adjoint $∣ T ∣ = V Σ V^{*}$ by the spectral theorem and absorbing the isometries yields $T = U V Σ V^{*} = (U V) Σ V^{*}$ , which is the SVD $T = W Σ V^{*}$ with $W = U V$ unitary.

Theorem (Cartan-Dieudonné; Cartan 1937, Dieudonné 1948 ^{[source pending]}). Every isometry of an $n$ -dimensional non-degenerate inner-product space over a field of characteristic $\neq = 2$ is a composition of at most $n$ reflections in hyperplanes. The identity uses zero reflections; a single hyperplane reflection $s_{u} (v) = v - 2 \frac{⟨ v , u ⟩}{⟨ u , u ⟩} u$ generates the whole orthogonal group together with its iterates.

A reflection $s_{u}$ is self-adjoint ( $s_{u} = s_{u}^{*}$ ) and isometric ( $s_{u}^{*} s_{u} = I$ ), and has determinant $- 1$ . The Cartan-Dieudonné bound makes the reflections the generating set of the orthogonal group, and the parity of the reflection count equals the sign of the determinant: an even number of reflections gives a rotation in $SO (n)$ , an odd number an orientation-reversing isometry. This generation result is the finite-dimensional root of Coxeter-group theory and of the description of Weyl groups as reflection groups.

Synthesis. The adjoint is the single operation from which the entire isometry-and-spectral edifice grows, and the matrix-level statement that the adjoint is conjugate transposition in an orthonormal basis is what ties the abstract identity $⟨ T v, w ⟩ = ⟨ v, T^{*} w ⟩$ to concrete computation. The four-subspace orthogonality $ker T^{*} = (im T)^{⊥}$ refines rank-nullity 01.01.05 from a dimension count to an orthogonal-complement statement, and the self-adjoint reality theorem feeds directly into the spectral theorem 01.01.13. Putting these together, the three characterisations of an isometry — operator relation $U^{*} U = I$ , inner-product preservation, and norm preservation — are linked by the polarisation identity into one notion, and that notion organises the orthogonal and unitary groups $O (n)$ and $U (n)$ as the symmetry groups of the inner product. The polar decomposition $T = U ∣ T ∣$ and the Cartan-Dieudonné reflection factorisation are the two structural theorems that close the unit: the first writes any operator as isometry-times-positive and bridges to the singular value decomposition 01.01.12, and the second exhibits every isometry as a product of at most $n$ reflections, identifying the orthogonal group as a reflection group. The adjoint, the isometry, the spectral theorem, and the singular value decomposition are four readings of the same inner-product geometry, and the conjugate transpose is the matrix shadow of all of them.

Full proof set Master

Proposition (algebraic laws of the adjoint). For operators $S, T$ on a finite-dimensional inner-product space over $F$ and $λ \in F$ : (a) $(T^)^* = T $; (b)$ (S + T)^* = S^* + T^ $; (c)$ (ST)^ = T^* S^ $; (d)$ (\lambda T)^ = \overline{\lambda} T^ $; (e)$ I^ = I$.*

Proof. Each law is read off the defining identity and non-degeneracy. For (a), $⟨ T^{*} v, w ⟩ = \overline{⟨ w, T^{*} v ⟩} = \overline{⟨ T w, v ⟩} = ⟨ v, T w ⟩$ for all $v, w$ , so the operator $(T^{*})^{*}$ characterised by $⟨ T^{*} v, w ⟩ = ⟨ v, (T^{*})^{*} w ⟩$ equals $T$ . For (b), $⟨(S + T) v, w ⟩ = ⟨ S v, w ⟩ + ⟨ T v, w ⟩ = ⟨ v, S^{*} w ⟩ + ⟨ v, T^{*} w ⟩ = ⟨ v, (S^{*} + T^{*}) w ⟩$ , and non-degeneracy gives $(S + T)^{*} = S^{*} + T^{*}$ . Law (c) is Exercise 4: $⟨ S T v, w ⟩ = ⟨ T v, S^{*} w ⟩ = ⟨ v, T^{*} S^{*} w ⟩$ . For (d), $⟨ λ T v, w ⟩ = λ ⟨ T v, w ⟩ = λ ⟨ v, T^{*} w ⟩ = ⟨ v, \overline{λ} T^{*} w ⟩$ , the last step by conjugate-linearity in the second slot, so $(λ T)^{*} = \overline{λ} T^{*}$ . For (e), $⟨ I v, w ⟩ = ⟨ v, w ⟩ = ⟨ v, I w ⟩$ , so $I^{*} = I$ . $□$

Proposition (reality of self-adjoint eigenvalues and orthogonality of eigenspaces). Let $T = T^ $o na f ini t e - d im e n s i o na l inn er - p r o d u c t s p a ce . E v er y e i g e n v a l u eo f$ T$ is real, and eigenvectors at distinct eigenvalues are orthogonal.*

Proof. Let $T v = λ v$ with $v \neq = 0$ . Then $$ \lambda \langle v, v \rangle = \langle \lambda v, v \rangle = \langle T v, v \rangle = \langle v, T^* v \rangle = \langle v, T v \rangle = \langle v, \lambda v \rangle = \overline{\lambda} \langle v, v \rangle, $$ using $T = T^{*}$ at the middle and conjugate-linearity in the second slot at the end. Since $⟨ v, v ⟩ = ∥ v ∥^{2} > 0$ , cancel to get $λ = \overline{λ}$ , so $λ \in R$ . For orthogonality, take $T v_{1} = λ_{1} v_{1}$ , $T v_{2} = λ_{2} v_{2}$ with $λ_{1} \neq = λ_{2}$ , both real. Then $⟨ T v_{1}, v_{2} ⟩ = ⟨ v_{1}, T v_{2} ⟩$ gives $λ_{1} ⟨ v_{1}, v_{2} ⟩ = \overline{λ_{2}} ⟨ v_{1}, v_{2} ⟩ = λ_{2} ⟨ v_{1}, v_{2} ⟩$ , so $(λ_{1} - λ_{2}) ⟨ v_{1}, v_{2} ⟩ = 0$ and $⟨ v_{1}, v_{2} ⟩ = 0$ . $□$

Proposition (three characterisations of an isometry, with polarisation). For linear $U$ on a finite-dimensional inner-product space, the conditions $U^ U = I$, inner-product preservation, and norm preservation are equivalent.*

Proof. ( $U^{*} U = I \Rightarrow$ inner-product preservation.) $⟨ U v, U w ⟩ = ⟨ v, U^{*} U w ⟩ = ⟨ v, w ⟩$ . (Inner-product $\Rightarrow$ norm.) Set $w = v$ : $∥ U v ∥^{2} = ⟨ U v, U v ⟩ = ⟨ v, v ⟩ = ∥ v ∥^{2}$ . (Norm $\Rightarrow$ inner-product.) Over $C$ , the polarisation identity expresses $$ \langle U v, U w \rangle = \tfrac{1}{4} \sum_{k=0}^{3} i^k |U v + i^k U w|^2 = \tfrac{1}{4} \sum_{k=0}^{3} i^k |U(v + i^k w)|^2 = \tfrac{1}{4} \sum_{k=0}^{3} i^k |v + i^k w|^2 = \langle v, w \rangle, $$ using linearity of $U$ in the middle and norm preservation at the third equality; over $R$ the same argument runs with the two-term identity $⟨ v, w ⟩ = \frac{1}{4} (∥ v + w ∥^{2} - ∥ v - w ∥^{2})$ . (Inner-product $\Rightarrow U^{*} U = I$ .) $⟨ v, U^{*} U w ⟩ = ⟨ U v, U w ⟩ = ⟨ v, w ⟩ = ⟨ v, I w ⟩$ for all $v, w$ , so non-degeneracy gives $U^{*} U = I$ . In finite dimensions $U^{*} U = I$ implies $U$ is invertible with two-sided inverse $U^{*}$ , so $U U^{*} = I$ as well. $□$

Proposition (isometries carry orthonormal bases to orthonormal bases; determinant of modulus one). If $U$ is an isometry and ${e_{1}, \dots, e_{n}}$ is an orthonormal basis, then ${U e_{1}, \dots, U e_{n}}$ is an orthonormal basis, and $∣ det U ∣ = 1$ .

Proof. $⟨ U e_{i}, U e_{j} ⟩ = ⟨ e_{i}, e_{j} ⟩ = δ_{ij}$ by inner-product preservation, so ${U e_{i}}$ is orthonormal, hence linearly independent (01.01.09) and a basis of the $n$ -dimensional space. For the determinant, $U^{*} U = I$ gives $det (U^{*}) det (U) = 1$ . Since $det (U^{*}) = \overline{det U}$ (the conjugate transpose conjugates the determinant: $det \overline{A}^{⊤} = det \overline{A} = \overline{det A}$ ), this reads $∣ det U ∣^{2} = 1$ , so $∣ det U ∣ = 1$ . Over $R$ the conjugation is the identity and $(det U)^{2} = 1$ , giving $det U = \pm 1$ . $□$

Proposition (a hyperplane reflection is self-adjoint and isometric). For nonzero $u \in V$ , the reflection $s_{u} (v) = v - 2 \frac{⟨ v , u ⟩}{⟨ u , u ⟩} u$ satisfies $s_u = s_u^ $,$ s_u^* s_u = I $,$ s_u^2 = I $, an d$ \det s_u = -1$.*

Proof. Write $c = 2/ ⟨ u, u ⟩$ , so $s_{u} (v) = v - c ⟨ v, u ⟩ u$ . Self-adjointness: for all $v, w$ , $$ \langle s_u v, w \rangle = \langle v, w \rangle - c \langle v, u \rangle \langle u, w \rangle, \qquad \langle v, s_u w \rangle = \langle v, w \rangle - c \overline{\langle w, u \rangle} \langle v, u \rangle = \langle v, w \rangle - c \langle u, w \rangle \langle v, u \rangle, $$ using conjugate symmetry $\overline{⟨ w, u ⟩} = ⟨ u, w ⟩$ ; the two expressions agree, so $s_{u} = s_{u}^{*}$ . Idempotence-to-involution: $s_{u} (s_{u} (v)) = s_{u} (v) - c ⟨ s_{u} (v), u ⟩ u$ , and $⟨ s_{u} (v), u ⟩ = ⟨ v, u ⟩ - c ⟨ v, u ⟩ ⟨ u, u ⟩ = ⟨ v, u ⟩ - 2 ⟨ v, u ⟩ = - ⟨ v, u ⟩$ , so $s_{u} (s_{u} (v)) = v - c ⟨ v, u ⟩ u - c (- ⟨ v, u ⟩) u = v$ , i.e., $s_{u}^{2} = I$ . Combining $s_{u} = s_{u}^{*}$ and $s_{u}^{2} = I$ gives $s_{u}^{*} s_{u} = s_{u}^{2} = I$ , so $s_{u}$ is an isometry. Finally $s_{u}$ fixes the hyperplane $u^{⊥}$ pointwise (eigenvalue $1$ , multiplicity $n - 1$ ) and negates $u$ (eigenvalue $- 1$ ), so $det s_{u} = (- 1) (1)^{n - 1} = - 1$ . $□$

Connections Master

The adjoint is the hypothesis carrier of the spectral theorem 01.01.13: the normality condition $T T^{*} = T^{*} T$ and the self-adjoint case $T = T^{*}$ are stated entirely in adjoint language, and the reality of self-adjoint eigenvalues proved here is the first step of that theorem. The isometries built here are the diagonalising operators that conjugate a normal operator to diagonal form.

The four-fundamental-subspaces orthogonality $ker T^{*} = (im T)^{⊥}$ and $im T^{*} = (ker T)^{⊥}$ refines the rank-nullity theorem 01.01.05 from a count of dimensions to a statement about orthogonal complements, and it is the algebraic backbone of least-squares projection and the normal equations.

The polar decomposition $T = U ∣ T ∣$ proved here is the bridge to the singular value decomposition 01.01.12: applying the spectral theorem to the positive self-adjoint $∣ T ∣ = (T^{*} T)^{1/2}$ and absorbing the isometric factor produces $T = W Σ V^{*}$ , so the singular values are the eigenvalues of $∣ T ∣$ and the singular vectors are its orthonormal eigenbasis.

The conjugate-transpose matrix of the adjoint specialises the transpose-and-bilinear-form theory of 01.01.15: a self-adjoint operator corresponds to a Hermitian (real symmetric) matrix, the matrix object whose congruence classes Sylvester's law of inertia classifies, and the orthogonal diagonalisation of such matrices is the principal-axes theorem.

Historical & philosophical context Master

The adjoint entered analysis through Hilbert's 1906 study of symmetric integral kernels on $L^{2}$ , where the symmetry condition $K (x, y) = \overline{K (y, x)}$ on the kernel is precisely self-adjointness of the associated integral operator, and the reality of the spectrum and orthogonality of eigenfunctions were the payoff ^{[Hilbert 1906]}. The existence mechanism — representing the functional $w \mapsto ⟨ T v, w ⟩$ by a vector — was isolated by Riesz in 1907 as the representation theorem for continuous linear functionals, the statement that the dual of a Hilbert space is the space itself ^{[Riesz 1907]}. Schmidt's 1908 paper recast the whole apparatus geometrically, introducing the norm, orthogonal projection, and the orthogonality language now standard, and it is Schmidt's framing that makes the finite-dimensional theory of this unit read as elementary geometry rather than as a chapter of integral-equation theory ^{[Schmidt 1908]}.

The finite-dimensional packaging that the present unit follows is due to Halmos, whose Finite-Dimensional Vector Spaces (1958) deliberately wrote the linear algebra of the adjoint, self-adjoint operators, and isometries in the notation of Hilbert-space operator theory, so that the finite-dimensional case would serve as a transparent model of the infinite-dimensional one ^{[Halmos 1958]}. Shilov's Linear Algebra gives the parallel Russian-school treatment, defining the adjoint by $⟨ T x, y ⟩ = ⟨ x, T^{*} y ⟩$ and developing the orthogonal and unitary operators as the isometries of the form ^{[Shilov 1971]}. The reflection-generation theorem for the orthogonal group was proved in the modern form by Cartan in 1937 and extended to general fields by Dieudonné in 1948, founding the structural theory of the classical groups and, through it, the reflection-group description of Weyl groups ^{[Cartan 1937]}.

Bibliography Master

@article{Hilbert1906,
  author  = {Hilbert, David},
  title   = {Grundz{\"u}ge einer allgemeinen Theorie der linearen Integralgleichungen. Vierte Mitteilung},
  journal = {Nachrichten von der K{\"o}niglichen Gesellschaft der Wissenschaften zu G{\"o}ttingen, Mathematisch-Physikalische Klasse},
  year    = {1906},
  pages   = {157--227}
}

@article{Riesz1907,
  author  = {Riesz, Frigyes},
  title   = {Sur une esp{\`e}ce de g{\'e}om{\'e}trie analytique des syst{\`e}mes de fonctions sommables},
  journal = {Comptes Rendus de l'Acad{\'e}mie des Sciences},
  volume  = {144},
  year    = {1907},
  pages   = {1409--1411}
}

@article{Schmidt1908,
  author  = {Schmidt, Erhard},
  title   = {{\"U}ber die Aufl{\"o}sung linearer Gleichungen mit unendlich vielen Unbekannten},
  journal = {Rendiconti del Circolo Matematico di Palermo},
  volume  = {25},
  year    = {1908},
  pages   = {53--77}
}

@book{Halmos1958,
  author    = {Halmos, Paul R.},
  title     = {Finite-Dimensional Vector Spaces},
  edition   = {2nd},
  publisher = {Van Nostrand},
  address   = {Princeton},
  year      = {1958}
}

@book{Shilov1977,
  author    = {Shilov, Georgi E.},
  title     = {Linear Algebra},
  publisher = {Dover Publications},
  address   = {New York},
  year      = {1977},
  note      = {Translation of the 1971 Russian edition, transl. R. A. Silverman}
}

@book{Axler2015,
  author    = {Axler, Sheldon},
  title     = {Linear Algebra Done Right},
  edition   = {3rd},
  publisher = {Springer},
  series    = {Undergraduate Texts in Mathematics},
  year      = {2015}
}

@book{Dieudonne1948,
  author    = {Dieudonn{\'e}, Jean},
  title     = {Sur les groupes classiques},
  publisher = {Hermann},
  address   = {Paris},
  year      = {1948}
}

Prerequisites

01.01.05
01.01.09

Tier anchors

beginner: The mirror-image map T* that moves an operator across the dot product; isometries as the rigid motions of space that fix the origin — Strang *Introduction to Linear Algebra* Ch. 4; 3Blue1Brown *Essence of Linear Algebra* Ch. 9 (dot products and duality)
intermediate: Shilov *Linear Algebra* §9.1–§9.2 (adjoint, orthogonal and unitary operators); Axler *Linear Algebra Done Right* §7.A (adjoint) + §7.D (isometries)
master: Shilov *Linear Algebra* Ch. 9 §9.1–§9.2; Halmos *Finite-Dimensional Vector Spaces* §44–§74; Axler *Linear Algebra Done Right* §7.A + §7.D; Lang *Algebra* Ch. XV; Reed-Simon *Methods of Modern Mathematical Physics* Vol. I §VI for the Hilbert-space adjoint

References

images/Shilov-Linear-Algebra__4cbdee00cc.jpg · Shilov *Linear Algebra* — Fast Track archive cover; Ch. 9 §9.1–§9.2 adjoint operator, isometries, orthogonal and unitary operators
linear-algebra_compress.pdf · linear-algebra (compressed reference) — adjoint / transpose, orthogonal and unitary matrices, conjugate-transpose chapters
Hilbert, D. — Grundzüge einer allgemeinen Theorie der linearen Integralgleichungen, Vierte Mitteilung · Nachrichten der K. Gesellschaft der Wissenschaften zu Göttingen (1906), 157–227 — symmetric (self-adjoint) integral operators on $L^2$, the infinite-dim antecedent of the finite-dim adjoint
Riesz, F. — Sur une espèce de géométrie analytique des systèmes de fonctions sommables · Comptes Rendus de l'Académie des Sciences 144 (1907), 1409–1411 — the Riesz representation theorem for continuous linear functionals, the existence mechanism for the adjoint
Schmidt, E. — Über die Auflösung linearer Gleichungen mit unendlich vielen Unbekannten · Rendiconti del Circolo Matematico di Palermo 25 (1908), 53–77 — the geometric inner-product / orthogonality framework (norm, orthogonal projection, Bessel) underlying the finite-dim adjoint and isometry theory
Halmos, P. R. — Finite-Dimensional Vector Spaces (2nd ed.) · §44–§74 — adjoint of a transformation, self-adjoint and normal operators, isometries (orthogonal and unitary), and the spectral theory built on them
Shilov, G. E. — Linear Algebra (Dover, 1977 transl. of the 1971 Russian ed.) · Ch. 9 §9.1–§9.2 — the adjoint operator $T^*$ defined by $\langle T x, y \rangle = \langle x, T^* y \rangle$; isometric (orthogonal / unitary) operators and their matrix characterisation
Axler, S. — Linear Algebra Done Right (3rd ed.) · §7.A (adjoint, self-adjoint and normal operators) and §7.D (isometries, polar decomposition, singular values)
Cartan, É. — La théorie des groupes finis et continus et la géométrie différentielle traitées par la méthode du repère mobile · Gauthier-Villars (1937); the Cartan-Dieudonné theorem that every isometry of a non-degenerate quadratic space is a product of at most $n$ reflections, with Dieudonné's *Sur les groupes classiques* (Hermann, 1948) for the general-field statement

Estimated time

beginner: 18m
intermediate: 45m
master: 85m