Normal Operators

When can an operator be diagonalized by an orthonormal basis? This single question drives a large part of the theory of Inner Product Spaces. The answer turns out to hinge on a deceptively simple algebraic condition: commuting with your own adjoint. An operator TT on a finite-dimensional inner product space is called normal if TT∗=T∗TTT^* = T^*T. Self-adjoint operators (T=T∗T = T^*) are the most familiar examples. But normality is a strictly larger and more flexible class. It encompasses unitary operators, skew-adjoint operators, and rotations in the plane. The importance of normality is that it is exactly the right hypothesis. It makes an operator behave well with respect to lengths and eigenvectors. It does this without forcing the operator to have only real eigenvalues. Normal operators satisfy the striking identity ∥Tv∥=∥T∗v∥\|Tv\| = \|T^*v\| for every vector vv. This single fact cascades into a rich structure theory. Eigenvectors of TT for distinct eigenvalues are automatically orthogonal. Every eigenvector of TT is simultaneously an eigenvector of T∗T^* with the conjugate eigenvalue. These facts are the load-bearing walls on which the Complex Spectral Theorem is built. The big picture is this: Self-Adjoint Operators are the "real numbers" of operator theory. They have real eigenvalues and orthonormal eigenbases. Normal operators are the "complex numbers" of operator theory. Over C\mathbb{C}, they also have orthonormal eigenbases. But their eigenvalues can be any complex numbers. This article isolates the precise commuting condition. It proves the norm-preservation identity that powers everything else. It also establishes the orthogonality-of-eigenspaces and shared-eigenvector results. These results are the direct stepping stones to the spectral theorem. This article delivers a complete, self-contained treatment: the formal definition, geometric intuition with figures, a fully proved central theorem with all its algebraic consequences, two asymmetric worked examples, a counterexample showing why the base field matters, and a map of where normality shows up across mathematics, physics, and computation.

Key takeaways
  • A normal operator satisfies TT∗=T∗TTT^* = T^*T; this is a commutativity condition, not a symmetry condition.
  • Normality is exactly the condition equivalent to ∥Tv∥=∥T∗v∥\|Tv\| = \|T^*v\| for all vectors vv.
  • Self-adjoint, skew-adjoint, unitary, and orthogonal operators are all normal, but normal operators need not be any of these.
  • Eigenvectors of a normal operator for distinct eigenvalues are automatically orthogonal.
  • If Tv=λvTv = \lambda v then T∗v=λ‾vT^*v = \overline{\lambda} v: TT and T∗T^* share eigenvectors with conjugated eigenvalues.
  • Key idea: commuting with the adjoint is precisely what makes an operator's action on lengths and eigenvectors mirror that of its adjoint.

Where this lives today

Normal operators are the natural habitat of the Complex Spectral Theorem, which states that TT on a finite-dimensional complex inner product space is normal if and only if VV has an orthonormal basis of eigenvectors of TT — the orthogonality proved here in part (d) is exactly the ingredient that makes such a basis orthonormal rather than merely a basis. In quantum mechanics, physical observables are represented by self-adjoint operators, but the broader class of normal operators governs the algebra of compatible observables and appears in the study of unitary time-evolution operators, which are normal but not self-adjoint. In numerical linear algebra, normal matrices are exactly the matrices diagonalizable by a unitary similarity, which is the theoretical basis for why algorithms like the symmetric/Hermitian QR algorithm and its normal-matrix generalizations converge so robustly, in contrast to general non-normal matrices where eigenvalue computations can be numerically unstable (the condition number of the eigenvector matrix blows up precisely when normality fails). In signal processing, circulant matrices (which represent convolution operators) are always normal, diagonalized by the discrete Fourier transform matrix — this is the linear-algebraic heart of the Fast Fourier Transform's spectral interpretation. In statistics, normal operators appear via covariance and correlation operators in multivariate analysis, and orthogonality of eigenspaces underlies the validity of principal component analysis. In cryptography and coding theory, normal matrices over finite fields and their spectral properties inform the design of linear codes with good distance properties.

Ask the tutor: Why does the Complex Spectral Theorem require normality exactly, and what fails in the proof of orthonormal diagonalizability if TT merely has nn distinct eigenvalues but is not normal?

DefinitionNormal Operator

Let VV be a finite-dimensional inner product space over F\mathbb{F} (where F=R\mathbb{F} = \mathbb{R} or F=C\mathbb{F} = \mathbb{C}), and let T∈L(V)T \in \mathcal{L}(V) be a linear operator on VV. Recall from The Adjoint of a Linear Map that the adjoint T∗∈L(V)T^* \in \mathcal{L}(V) is the unique operator satisfying

⟨Tv,w⟩=⟨v,T∗w⟩for all v,w∈V.\langle Tv, w \rangle = \langle v, T^*w \rangle \quad \text{for all } v, w \in V.

The operator TT is called normal if

TT∗=T∗T,TT^* = T^*T,

i.e., TT commutes with its own adjoint. Equivalently, (TT∗−T∗T)v=0(TT^* - T^*T)v = 0 for every v∈Vv \in V.

Intuition

Think of T∗T^* as a kind of "mirror image" of TT dictated entirely by the inner product. For a self-adjoint operator, the mirror image is the operator itself: T=T∗T = T^*. Normality is a weaker demand — it does not require TT to equal its mirror image, only that TT and its mirror image commute as transformations. Commuting operators share a great deal of structure (for instance, they often share eigenvectors), and this is exactly what happens here: TT and T∗T^*, while generally different operators, act compatibly enough that they share every eigenvector. A good mental picture is a rotation in the plane. A rotation by angle θ\theta is not self-adjoint (its matrix is not symmetric unless θ\theta is a multiple of π\pi), yet it is normal: rotating and then reflecting (which is what the adjoint of a rotation matrix does, since rotations are orthogonal and T∗=T−1T^* = T^{-1}) commutes with rotating first and then applying that same inverse. The key geometric fact that makes rotations "nice" despite not being self-adjoint is that a rotation never changes lengths — and neither does its adjoint. This length-preservation-of-both-TT-and-T∗T^* phenomenon is the heart of the theorem below: normality is equivalent to the statement ∥Tv∥=∥T∗v∥\|Tv\| = \|T^*v\| for every vv.

Figure
A rotation is normal but not self-adjoint: both TT and T∗T^* preserve the length of vv
v Tv (rotate +40°) T*v (rotate −40°) ‖v‖ = ‖Tv‖ = ‖T*v‖ = 104 (all endpoints lie on one circle)

Core theorem and its consequences

TheoremCharacterization and Consequences of Normal Operators

Let VV be a finite-dimensional inner product space over F\mathbb{F} (F=R\mathbb{F} = \mathbb{R} or C\mathbb{C}), and let T∈L(V)T \in \mathcal{L}(V).

(a) Norm equality. TT is normal if and only if

∥Tv∥=∥T∗v∥for every v∈V.\|Tv\| = \|T^*v\| \quad \text{for every } v \in V.

(b) Kernels agree. If TT is normal, then ker⁡T=ker⁡T∗\ker T = \ker T^*.

(c) Eigenvector transfer. If TT is normal and λ∈F\lambda \in \mathbb{F}, v∈Vv \in V satisfy Tv=λvTv = \lambda v, then

T∗v=λ‾ v.T^*v = \overline{\lambda}\, v.

That is, every eigenvector of TT with eigenvalue λ\lambda is an eigenvector of T∗T^* with eigenvalue λ‾\overline{\lambda}.

(d) Orthogonality of eigenspaces. If TT is normal and λ1≠λ2\lambda_1 \neq \lambda_2 are distinct eigenvalues of TT with eigenvectors v1,v2v_1, v_2 respectively, then v1⊥v2v_1 \perp v_2.

Proof

We prove each part in order; part (a) is the foundation for (b)–(d), and it is proved with no circularity, using only the definition of adjoint and basic inner product axioms.

Proof of (a). For any S∈L(V)S \in \mathcal{L}(V) and any v∈Vv \in V, by definition of the adjoint,

∥Sv∥2=⟨Sv,Sv⟩=⟨v,S∗Sv⟩.\|Sv\|^2 = \langle Sv, Sv \rangle = \langle v, S^*Sv \rangle.

Apply this identity twice, once with S=TS = T and once with S=T∗S = T^* (using (T∗)∗=T(T^*)^* = T, a standard property of adjoints established alongside the definition of the adjoint):

∥Tv∥2=⟨v,T∗Tv⟩,∥T∗v∥2=⟨v,TT∗v⟩.\|Tv\|^2 = \langle v, T^*Tv \rangle, \qquad \|T^*v\|^2 = \langle v, TT^*v \rangle.

Subtracting,

∥Tv∥2−∥T∗v∥2=⟨v,T∗Tv⟩−⟨v,TT∗v⟩=⟨v,(T∗T−TT∗)v⟩.(∗)\|Tv\|^2 - \|T^*v\|^2 = \langle v, T^*Tv \rangle - \langle v, TT^*v \rangle = \langle v, (T^*T - TT^*)v \rangle. \tag{$\ast$}

(⇒\Rightarrow) Suppose TT is normal, so T∗T−TT∗=0T^*T - TT^* = 0. Then the right side of (∗)(\ast) is 00 for every vv, so ∥Tv∥2=∥T∗v∥2\|Tv\|^2 = \|T^*v\|^2 for every vv, hence ∥Tv∥=∥T∗v∥\|Tv\| = \|T^*v\| (norms are nonnegative). (⇐\Leftarrow) Suppose ∥Tv∥=∥T∗v∥\|Tv\| = \|T^*v\| for all v∈Vv \in V. Define S=T∗T−TT∗S = T^*T - TT^*. This operator is self-adjoint: indeed,

S∗=(T∗T)∗−(TT∗)∗=T∗(T∗)∗−(T∗)∗T∗=T∗T−TT∗=S,S^* = (T^*T)^* - (TT^*)^* = T^*(T^*)^* - (T^*)^*T^* = T^*T - TT^* = S,

using (AB)∗=B∗A∗(AB)^* = B^*A^* and (T∗)∗=T(T^*)^* = T. By hypothesis and (∗)(\ast), ⟨v,Sv⟩=0\langle v, Sv \rangle = 0 for every v∈Vv \in V. We now invoke the standard fact (proved for self-adjoint operators, and valid over both R\mathbb{R} and C\mathbb{C} given the self-adjointness of SS, using polarization) that a self-adjoint operator SS satisfying ⟨v,Sv⟩=0\langle v, Sv\rangle = 0 for all vv must be the zero operator. We reprove this here directly to keep the argument self-contained. Since ⟨v,Sv⟩=0\langle v, Sv \rangle = 0 for all vv, replace vv by u+wu + w for arbitrary u,w∈Vu, w \in V:

0=⟨u+w,S(u+w)⟩=⟨u,Su⟩+⟨u,Sw⟩+⟨w,Su⟩+⟨w,Sw⟩.0 = \langle u+w, S(u+w)\rangle = \langle u, Su\rangle + \langle u, Sw\rangle + \langle w, Su\rangle + \langle w, Sw \rangle.

The first and last terms vanish by hypothesis (applied to uu and to ww), leaving

⟨u,Sw⟩+⟨w,Su⟩=0for all u,w∈V.(†)\langle u, Sw \rangle + \langle w, Su \rangle = 0 \quad \text{for all } u, w \in V. \tag{$\dagger$}

Since SS is self-adjoint, ⟨w,Su⟩=⟨Sw,u⟩=⟨u,Sw⟩‾\langle w, Su \rangle = \langle Sw, u \rangle = \overline{\langle u, Sw\rangle}. Substituting into (†)(\dagger):

⟨u,Sw⟩+⟨u,Sw⟩‾=0,i.e.2 Re⟨u,Sw⟩=0 for all u,w.\langle u, Sw \rangle + \overline{\langle u, Sw \rangle} = 0, \quad \text{i.e.} \quad 2\,\mathrm{Re}\langle u, Sw\rangle = 0 \text{ for all } u,w.

So Re⟨u,Sw⟩=0\mathrm{Re}\langle u, Sw \rangle = 0 for all u,w∈Vu, w \in V. If F=R\mathbb{F} = \mathbb{R}, this already says ⟨u,Sw⟩=0\langle u, Sw\rangle = 0 for all u,wu,w; taking u=Swu = Sw gives ∥Sw∥2=0\|Sw\|^2 = 0 so Sw=0Sw = 0 for every ww, i.e. S=0S = 0. If F=C\mathbb{F} = \mathbb{C}, replace uu by iuiu in the same identity: Re⟨iu,Sw⟩=0\mathrm{Re}\langle iu, Sw \rangle = 0. Since ⟨iu,Sw⟩=i‾⟨u,Sw⟩=−i⟨u,Sw⟩\langle iu, Sw\rangle = \overline{i}\langle u, Sw\rangle = -i\langle u, Sw\rangle, and Re(−iz)=Im(z)\mathrm{Re}(-iz) = \mathrm{Im}(z) for z∈Cz \in \mathbb{C}, we get Im⟨u,Sw⟩=0\mathrm{Im}\langle u, Sw\rangle = 0 as well. Combined with Re⟨u,Sw⟩=0\mathrm{Re}\langle u, Sw\rangle = 0, this forces ⟨u,Sw⟩=0\langle u, Sw \rangle = 0 for all u,wu, w; taking u=Swu = Sw again gives S=0S = 0. Either way, S=T∗T−TT∗=0S = T^*T - TT^* = 0, i.e. T∗T=TT∗T^*T = TT^*, so TT is normal. This proves (a).

Proof of (b). By part (a) with v∈ker⁡Tv \in \ker T: ∥T∗v∥=∥Tv∥=∥0∥=0\|T^*v\| = \|Tv\| = \|0\| = 0, so T∗v=0T^*v = 0, i.e. v∈ker⁡T∗v \in \ker T^*. This shows ker⁡T⊆ker⁡T∗\ker T \subseteq \ker T^*. Applying the same argument with TT replaced by T∗T^* — note T∗T^* is also normal, since T∗(T∗)∗=T∗T=TT∗=(T∗)∗T∗T^*(T^*)^* = T^*T = TT^* = (T^*)^*T^* — gives ker⁡T∗⊆ker⁡(T∗)∗=ker⁡T\ker T^* \subseteq \ker (T^*)^* = \ker T. Hence ker⁡T=ker⁡T∗\ker T = \ker T^*.

Proof of (c). Suppose Tv=λvTv = \lambda v. Consider the operator T−λIT - \lambda I. We claim T−λIT - \lambda I is normal: its adjoint is (T−λI)∗=T∗−λ‾I(T-\lambda I)^* = T^* - \overline{\lambda} I (adjoint of a scalar multiple conjugates the scalar, a basic property of adjoints), and

(T−λI)(T−λI)∗=(T−λI)(T∗−λ‾I)=TT∗−λ‾T−λT∗+∣λ∣2I.(T-\lambda I)(T-\lambda I)^* = (T - \lambda I)(T^* - \overline{\lambda}I) = TT^* - \overline{\lambda}T - \lambda T^* + |\lambda|^2 I.
(T−λI)∗(T−λI)=(T∗−λ‾I)(T−λI)=T∗T−λT∗−λ‾T+∣λ∣2I.(T-\lambda I)^*(T-\lambda I) = (T^* - \overline{\lambda}I)(T-\lambda I) = T^*T - \lambda T^* - \overline{\lambda}T + |\lambda|^2 I.

Since TT is normal, TT∗=T∗TTT^* = T^*T, so these two expressions are equal; thus T−λIT - \lambda I is normal. Now apply part (a) to the normal operator T−λIT - \lambda I and the vector vv:

∥(T−λI)v∥=∥(T−λI)∗v∥.\|(T-\lambda I)v\| = \|(T-\lambda I)^* v\|.

The left side is ∥Tv−λv∥=∥0∥=0\|Tv - \lambda v\| = \|0\| = 0 since Tv=λvTv = \lambda v. Hence the right side is also 00:

∥(T∗−λ‾I)v∥=0  ⟹  (T∗−λ‾I)v=0  ⟹  T∗v=λ‾v.\|(T^* - \overline{\lambda}I)v\| = 0 \implies (T^* - \overline{\lambda}I)v = 0 \implies T^*v = \overline{\lambda}v.

Proof of (d). Let Tv1=λ1v1Tv_1 = \lambda_1 v_1 and Tv2=λ2v2Tv_2 = \lambda_2 v_2 with λ1≠λ2\lambda_1 \neq \lambda_2. By part (c) applied to v2v_2 (eigenvalue λ2\lambda_2), we have T∗v2=λ2‾v2T^*v_2 = \overline{\lambda_2}v_2. Compute ⟨Tv1,v2⟩\langle Tv_1, v_2\rangle two ways using the defining property of the adjoint:

⟨Tv1,v2⟩=λ1⟨v1,v2⟩,\langle Tv_1, v_2 \rangle = \lambda_1 \langle v_1, v_2\rangle,
⟨Tv1,v2⟩=⟨v1,T∗v2⟩=⟨v1,λ2‾v2⟩=λ2⟨v1,v2⟩,\langle Tv_1, v_2 \rangle = \langle v_1, T^*v_2 \rangle = \langle v_1, \overline{\lambda_2}v_2 \rangle = \lambda_2\langle v_1, v_2\rangle,

where the last step uses conjugate-linearity of the inner product in the second slot: ⟨v1,λ2‾v2⟩=λ2‾‾⟨v1,v2⟩=λ2⟨v1,v2⟩\langle v_1, \overline{\lambda_2} v_2\rangle = \overline{\overline{\lambda_2}}\langle v_1, v_2\rangle = \lambda_2\langle v_1, v_2\rangle. Equating the two expressions:

λ1⟨v1,v2⟩=λ2⟨v1,v2⟩  ⟹  (λ1−λ2)⟨v1,v2⟩=0.\lambda_1\langle v_1, v_2\rangle = \lambda_2 \langle v_1, v_2 \rangle \implies (\lambda_1 - \lambda_2)\langle v_1, v_2\rangle = 0.

Since λ1≠λ2\lambda_1 \neq \lambda_2, we conclude ⟨v1,v2⟩=0\langle v_1, v_2 \rangle = 0, i.e. v1⊥v2v_1 \perp v_2. ■\blacksquare

∎
CorollarySelf-adjoint operators are normal, with real eigenvalues

If TT is self-adjoint (T=T∗T = T^*), then trivially TT∗=T2=T∗TTT^* = T^2 = T^*T, so TT is normal. If in addition Tv=λvTv = \lambda v for v≠0v \neq 0, part (c) gives T∗v=λ‾vT^*v = \overline{\lambda}v, but T∗v=Tv=λvT^*v = Tv = \lambda v, so λ‾v=λv\overline{\lambda}v = \lambda v, forcing λ‾=λ\overline{\lambda} = \lambda (as v≠0v \neq 0), i.e. λ∈R\lambda \in \mathbb{R}. This recovers the familiar fact that Self-Adjoint Operators have real eigenvalues, as a special case of normality.

Worked examples

Example 1 — a genuinely non-self-adjoint normal operator, with an asymmetric matrix. Let V=C2V = \mathbb{C}^2 with the standard inner product, and let TT be given in the standard basis by

A=(2+i−112+i).A = \begin{pmatrix} 2 + i & -1 \\ 1 & 2+i \end{pmatrix}.

We verify normality directly and then locate the eigenvector-sharing predicted by part (c). The adjoint (conjugate transpose) is

A∗=(2−i1−12−i).A^* = \begin{pmatrix} 2 - i & 1 \\ -1 & 2-i \end{pmatrix}.

Compute AA∗AA^*:

AA∗=(2+i−112+i)(2−i1−12−i)=((2+i)(2−i)+1(2+i)−(2−i)(2−i)−(2+i)1+(2+i)(2−i)).AA^* = \begin{pmatrix} 2+i & -1 \\ 1 & 2+i \end{pmatrix}\begin{pmatrix} 2-i & 1 \\ -1 & 2-i \end{pmatrix} = \begin{pmatrix} (2+i)(2-i) + 1 & (2+i) - (2-i) \\ (2-i) - (2+i) & 1 + (2+i)(2-i) \end{pmatrix}.

Since (2+i)(2−i)=4+1=5(2+i)(2-i) = 4+1 = 5, the diagonal entries are both 5+1=65+1=6. The off-diagonal entries: (2+i)−(2−i)=2i(2+i)-(2-i) = 2i and (2−i)−(2+i)=−2i(2-i)-(2+i) = -2i. So

AA∗=(62i−2i6).AA^* = \begin{pmatrix} 6 & 2i \\ -2i & 6\end{pmatrix}.

Now compute A∗AA^*A:

A∗A=(2−i1−12−i)(2+i−112+i)=((2−i)(2+i)+1−(2−i)+(2+i)−(2+i)+(2−i)1+(2−i)(2+i)).A^*A = \begin{pmatrix} 2-i & 1 \\ -1 & 2-i \end{pmatrix}\begin{pmatrix} 2+i & -1 \\ 1 & 2+i \end{pmatrix} = \begin{pmatrix} (2-i)(2+i)+1 & -(2-i)+(2+i) \\ -(2+i)+(2-i) & 1+(2-i)(2+i) \end{pmatrix}.

Again (2−i)(2+i)=5(2-i)(2+i) = 5, diagonal entries =6= 6. Off-diagonal: −(2−i)+(2+i)=2i-(2-i)+(2+i) = 2i and −(2+i)+(2−i)=−2i-(2+i)+(2-i) = -2i. So

A∗A=(62i−2i6)=AA∗.A^*A = \begin{pmatrix} 6 & 2i \\ -2i & 6\end{pmatrix} = AA^*.

Thus TT is normal. Note A≠A∗A \neq A^* (compare the (1,2)(1,2) entries: −1-1 versus 11), so TT is not self-adjoint, confirming that normality is a strictly weaker condition. Now find the eigenvalues. The characteristic polynomial is

det⁡(A−xI)=(2+i−x)2−(−1)(1)=(2+i−x)2+1.\det(A - xI) = (2+i-x)^2 - (-1)(1) = (2+i-x)^2 + 1.

Set (2+i−x)2=−1(2+i - x)^2 = -1, so 2+i−x=±i2+i - x = \pm i. This gives x=2+i−i=2x = 2+i-i = 2 or x=2+i+i=2+2ix = 2+i+i = 2+2i. So the eigenvalues are λ1=2\lambda_1 = 2 and λ2=2+2i\lambda_2 = 2+2i, matching the corollary's prediction of complex (non-real) eigenvalues for a non-self-adjoint operator. For λ1=2\lambda_1 = 2: solve (A−2I)v=0(A - 2I)v = 0:

(i−11i)(xy)=0  ⟹  ix=y.\begin{pmatrix} i & -1 \\ 1 & i\end{pmatrix}\begin{pmatrix}x\\y\end{pmatrix} = 0 \implies ix = y.

Take v1=(1,i)v_1 = (1, i). For λ2=2+2i\lambda_2 = 2+2i: solve (A−(2+2i)I)v=0(A - (2+2i)I)v = 0:

(−i−11−i)(xy)=0  ⟹  x=iy.\begin{pmatrix} -i & -1 \\ 1 & -i \end{pmatrix}\begin{pmatrix}x\\y\end{pmatrix}=0 \implies x = iy.

Take v2=(i,1)v_2 = (i, 1).

Check orthogonality (part (d)): ⟨v1,v2⟩=1⋅i‾+i⋅1‾=−i+i=0.\langle v_1, v_2\rangle = 1\cdot\overline{i} + i \cdot \overline{1} = -i + i = 0. ✓, confirming eigenspaces for distinct eigenvalues are orthogonal, exactly as guaranteed since TT is normal.

Check eigenvector transfer (part (c)): we predict T∗v1=λ1‾v1=2v1T^*v_1 = \overline{\lambda_1}v_1 = 2v_1. Compute directly:

A∗v1=(2−i1−12−i)(1i)=((2−i)+i−1+i(2−i))=(2−1+2i+1)=(22i)=2v1.A^*v_1 = \begin{pmatrix}2-i & 1\\ -1 & 2-i\end{pmatrix}\begin{pmatrix}1\\i\end{pmatrix} = \begin{pmatrix}(2-i)+i \\ -1 + i(2-i)\end{pmatrix} = \begin{pmatrix}2 \\ -1+2i+1\end{pmatrix} = \begin{pmatrix}2\\2i\end{pmatrix} = 2v_1.

Indeed A∗v1=2v1=λ1‾v1A^*v_1 = 2v_1 = \overline{\lambda_1}v_1, confirming part (c) numerically.

Example 2 — using the norm identity to prove normality fails, with an asymmetric 2×22\times2 real matrix. Let TT on R2\mathbb{R}^2 (standard inner product) have matrix

B=(1201).B = \begin{pmatrix} 1 & 2 \\ 0 & 1 \end{pmatrix}.

Is TT normal? By part (a), it suffices to find one vector vv with ∥Tv∥≠∥T∗v∥\|Tv\| \neq \|T^*v\|. We have B∗=BT=(1021)B^* = B^T = \begin{pmatrix}1 & 0\\ 2& 1\end{pmatrix}. Take v=(0,1)v = (0,1). Then

Tv=Bv=(21),∥Tv∥=4+1=5.Tv = Bv = \begin{pmatrix}2\\1\end{pmatrix}, \qquad \|Tv\| = \sqrt{4+1} = \sqrt5.
T∗v=BTv=(01),∥T∗v∥=1.T^*v = B^Tv = \begin{pmatrix}0\\1\end{pmatrix}, \qquad \|T^*v\| = 1.

Since 5≠1\sqrt5 \neq 1, part (a) tells us TT is not normal — without ever computing BB∗BB^* or B∗BB^*B directly, though for confirmation:

BB∗=(1201)(1021)=(5221),B∗B=(1021)(1201)=(1225),BB^* = \begin{pmatrix}1&2\\0&1\end{pmatrix}\begin{pmatrix}1&0\\2&1\end{pmatrix} = \begin{pmatrix}5&2\\2&1\end{pmatrix}, \qquad B^*B = \begin{pmatrix}1&0\\2&1\end{pmatrix}\begin{pmatrix}1&2\\0&1\end{pmatrix} = \begin{pmatrix}1&2\\2&5\end{pmatrix},

and indeed BB∗≠B∗BBB^* \neq B^*B. This example illustrates the practical power of part (a): a single failed norm comparison certifies non-normality without a full matrix multiplication, and it shows that a generic (asymmetric, real, non-orthogonal) matrix is normal only under real algebraic constraints — here the off-diagonal asymmetry (22 versus 00) is exactly what breaks the commuting condition.

Counterexamplewhy the ground field / conjugate-linearity structure matters: an operator satisfying a "half" version of the theorem

The proof of part (a) used conjugate-linearity of the inner product in one slot in an essential way (specifically, in isolating both real and imaginary parts to force S=0S=0 over C\mathbb{C}). One might ask: is the converse-style implication in part (a) — "norm equality forces normality" — perhaps not needed at all, and could we instead try to get eigenvector-orthogonality (part (d)) from a weaker hypothesis than full normality, such as merely TT having an orthogonal eigenbasis to begin with, assumed rather than derived? Here is why the normality hypothesis itself cannot be dropped from parts (c) and (d): consider V=C2V = \mathbb{C}^2 and the operator TT with matrix

C=(1102),C = \begin{pmatrix} 1 & 1 \\ 0 & 2 \end{pmatrix},

which is not normal (a direct computation, as in Example 2's style, shows CC∗≠C∗CCC^* \ne C^*C: with C∗=(1012)C^* = \begin{pmatrix}1&0\\1&2\end{pmatrix},

CC∗=(2224),C∗C=(1115),CC^* = \begin{pmatrix}2&2\\2&4\end{pmatrix}, \qquad C^*C = \begin{pmatrix}1&1\\1&5\end{pmatrix},

and these disagree, e.g. in the (1,1)(1,1) entry: 2≠12 \ne 1). The eigenvalues of CC are λ1=1\lambda_1 = 1 (eigenvector v1=(1,0)v_1 = (1,0)) and λ2=2\lambda_2 = 2 (eigenvector: solve (C−2I)v=0(C - 2I)v=0, i.e. (−1100)v=0\begin{pmatrix}-1&1\\0&0\end{pmatrix}v = 0, giving v2=(1,1)v_2 = (1,1)). Check orthogonality: ⟨v1,v2⟩=1⋅1‾+0⋅1‾=1≠0\langle v_1, v_2 \rangle = 1\cdot\overline{1} + 0\cdot\overline{1} = 1 \neq 0. So v1v_1 and v2v_2, eigenvectors for the distinct eigenvalues 1≠21 \ne 2, are not orthogonal — part (d) fails completely once normality is dropped. Likewise, part (c) fails: C∗v1=(1012)(10)=(11)C^*v_1 = \begin{pmatrix}1&0\\1&2\end{pmatrix}\begin{pmatrix}1\\0\end{pmatrix} = \begin{pmatrix}1\\1\end{pmatrix}, which is not a scalar multiple of v1=(1,0)v_1 = (1,0) at all (let alone λ1‾v1=v1\overline{\lambda_1}v_1 = v_1), so T∗T^* does not even share this eigenvector, never mind with the conjugated eigenvalue. This shows the normality hypothesis in the theorem is not a technical convenience — it is exactly the condition without which both conclusions (c) and (d) can fail outright.

Scope, edge cases & common traps

SituationStatusNote
TT self-adjointNormalSpecial case: T=T∗T = T^* commutes with itself trivially
TT unitary/orthogonalNormalTT∗=I=T∗TTT^* = I = T^*T
TT skew-adjoint (T∗=−TT^* = -T)NormalTT∗=−T2=T∗TTT^* = -T^2 = T^*T
TT nilpotent, T≠0T \neq 0Generally not normale.g. (0100)\begin{pmatrix}0&1\\0&0\end{pmatrix} fails norm test
Infinite-dimensional VVDefinition extends, but proof of (a) needs careRequires bounded operators, adjoint existence assumed
F=R\mathbb{F} = \mathbb{R}, TT normalEigenvalues may be non-real as roots of char. poly, but TT may have no real eigenvectorsReal spectral theory needs orthogonal block structure instead
Distinct eigenvalues assumed in (d)RequiredEqual eigenvalues give no orthogonality constraint from this argument

The mathematician behind it

Origins · 1 minDavid Hilbert and how this was discovered
DH
David Hilbert
1862–1943 · Göttingen, foundational work on inner product spaces and operator theory

Hilbert's work on integral equations and quadratic forms in infinitely many variables, developed in the first decade of the 1900s, laid the groundwork for the abstract theory of what are now called Hilbert spaces. The classification of operators by their relationship to their adjoints — self-adjoint, normal, unitary — emerged from this tradition as mathematicians and physicists (notably John von Neumann, working to formalize quantum mechanics) sought the right algebraic conditions under which an operator admits a complete orthogonal eigenbasis or an analogous spectral decomposition.

History
- 1904–1910: Hilbert develops spectral theory for integral equations, introducing what becomes Hilbert space theory.
- 1929–1932: von Neumann gives the rigorous Hilbert-space formulation of quantum mechanics, formalizing self-adjoint and unitary operators as observables and symmetries.
- 1930s: The term "normal operator" and the identity ∥Tv∥=∥T∗v∥\|Tv\|=\|T^*v\| become standard tools in operator theory, unifying self-adjoint and unitary cases.
- 1940s–1950s: The spectral theorem for normal operators on finite-dimensional and Hilbert spaces is codified in the modern textbook form.
- Present: Normality remains the standard hypothesis for diagonalizability-by-orthonormal-basis results across pure and applied linear algebra.
Is every normal operator diagonalizable?

Over C\mathbb{C}, yes: this is precisely the content of the Complex Spectral Theorem, which builds directly on the results proved here (orthogonality of eigenspaces for distinct eigenvalues, and the shared-eigenvector fact) to show a normal operator has an orthonormal basis of eigenvectors. Over R\mathbb{R}, the situation is more subtle: a normal real operator need not be diagonalizable over the reals (a rotation by a nonzero, non-π\pi angle is normal but has no real eigenvectors at all), though it does admit a block-diagonal decomposition into 1×11\times1 blocks and 2×22\times2 rotation-scaling blocks. This article establishes the algebraic core; the diagonalization statement itself is the spectral theorem, treated separately.

Why is the condition called "normal" — is there a geometric meaning to the word?

The terminology is largely historical, tracing to the German "normal" in the sense of "standard" or "canonical form," reflecting the fact that these are exactly the operators reducible to a canonical (diagonal, in the complex case) form via an orthonormal change of basis. There is no direct connection to "normal vectors" in the sense of perpendicularity, although the eigenvectors of a normal operator do end up mutually orthogonal (part (d)) — a coincidence of terminology that is worth noting but is not the origin of the name.

Does TT∗=T∗TTT^* = T^*T need to be checked on a basis, or does it really need to hold on every vector?

Since TT∗TT^* and T∗TT^*T are both linear operators, they are equal as operators if and only if they agree on every vector of any one basis (spanning set), by linearity. So in practice, checking TT∗ei=T∗TeiTT^*e_i = T^*Te_i for each vector eie_i in a basis is sufficient, or equivalently (and usually easiest) simply multiplying the matrices AA∗AA^* and A∗AA^*A in some orthonormal basis and comparing the two resulting matrices entrywise, as done in the worked examples.

Is the sum or product of two normal operators always normal?

No, in general neither is normal — this is a common trap. The product TSTS of two normal operators need not be normal unless TT and SS commute with each other; even the sum T+ST+S need not be normal without further hypotheses. However, if TT and SS are both normal and commute with each other and with each other's adjoints in the right combination, more can be said, but the general statement requires care and is not automatic from normality of the individual factors alone.

How does normality relate to self-adjointness and unitarity — is there a Venn diagram?

Yes: self-adjoint operators and unitary (orthogonal) operators are both special cases of normal operators, but they are generally disjoint classes from each other outside trivial overlaps (an operator that is both self-adjoint and unitary must satisfy T2=IT^2 = I with T=T∗T = T^*, so it is an orthogonal projection reflection, i.e. T2=IT^2=I). Normal operators form the larger umbrella class: TT is normal iff it commutes with T∗T^*, a strictly weaker condition than T=T∗T = T^* (self-adjoint) or T∗T=IT^*T = I (unitary), each of which forces commuting with T∗T^* trivially but not conversely.

Why does the proof of part (a) need the self-adjointness of S=T∗T−TT∗S = T^*T - TT^*, rather than working with TT directly?

The trick of reducing "⟨v,Sv⟩=0\langle v, Sv\rangle = 0 for all vv implies S=0S=0" to a clean polarization argument genuinely requires SS to be self-adjoint (or at least normal) — otherwise the identity ⟨w,Su⟩=⟨Sw,u⟩\langle w, Su\rangle = \langle Sw, u\rangle used in the proof would fail, since that step used S=S∗S = S^* explicitly. Fortunately, S=T∗T−TT∗S = T^*T - TT^* is automatically self-adjoint for any operator TT whatsoever (as shown directly in the proof), which is precisely why this auxiliary operator, rather than TT itself, is the right object to apply the "zero quadratic form implies zero operator" lemma to.

Can normality be checked just by looking at a matrix's entries, without computing two matrix products?

There is a useful necessary condition: the sum of squared absolute values of all matrix entries of AA equals the sum of the squares of the absolute values of the eigenvalues of AA (counted with multiplicity) if and only if AA is normal — this is a consequence of the spectral theorem together with the trace identity tr(A∗A)=∑i,j∣aij∣2\mathrm{tr}(A^*A) = \sum_{i,j}|a_{ij}|^2. This gives a quick numerical test one can apply after finding eigenvalues, though the direct definitional check (AA∗AA^* vs A∗AA^*A) or the norm-equality check from part (a) remain the standard hands-on methods for a specific matrix.

Before this

Where this leads