Self-Adjoint Operators

Which linear operators behave, spectrally, like real numbers? This is instead of arbitrary complex matrices. On an inner product space, the answer is clear. It is exactly the operators equal to their own adjoint. This single algebraic condition, T=T∗T = T^*, forces an entire cascade of good behavior. Every eigenvalue is real. The associated quadratic form is real-valued. Eigenvectors belonging to different eigenvalues are automatically orthogonal. Self-adjoint operators matter for an important reason. They are the linear-algebraic shadow of symmetric matrices and Hermitian matrices. These objects appear the moment you try to diagonalize a quadratic form. They also appear when you minimize an energy, or model an observable in quantum mechanics. Symmetric bilinear forms, second-derivative tests in multivariable calculus, and the covariance matrices of statistics are all secretly self-adjoint operators in disguise. Recognizing the pattern lets you transfer intuition from real numbers directly to operators on a vector space. This includes ordering, sign, and magnitude. The big picture is this: self-adjointness is the correct abstract generalization. It generalizes "the matrix is symmetric" or "the matrix is Hermitian." It is phrased in a basis-free way using the adjoint map. Once you know an operator is self-adjoint, you know several things. Its eigenvalues live on the real line. It has an associated real quadratic form controlling positivity. Its eigenspaces slot together orthogonally. Together these three facts set up the machinery for the crown jewel of the subject. That crown jewel is the Spectral Theorem. This article builds self-adjointness from the The Adjoint of a Linear Map up. It proves the three core facts named above with full rigor. It works two concrete asymmetric examples by hand. This way, every abstract claim is checked against real numbers. It closes by pointing toward Normal Operators and The Spectral Theorem for Self-Adjoint Operators. These are the natural next destinations.

Key takeaways
  • A self-adjoint operator satisfies T=T∗T = T^*, i.e., ⟨Tv,w⟩=⟨v,Tw⟩\langle Tv, w\rangle = \langle v, Tw\rangle for all v,wv, w.
  • On matrices this means: symmetric (A=ATA = A^T) over R\mathbb{R}, Hermitian (A=A∗A = A^*) over C\mathbb{C}.
  • Every eigenvalue of a self-adjoint operator is a real number, even over C\mathbb{C}.
  • The quadratic form v↦⟨Tv,v⟩v \mapsto \langle Tv, v\rangle is real-valued for self-adjoint TT.
  • Eigenvectors of a self-adjoint operator belonging to distinct eigenvalues are automatically orthogonal.
  • Key idea: self-adjointness is the coordinate-free condition that makes an operator "act like a real number," anchoring the Spectral Theorem.

Where this lives today

Self-adjoint operators are the mathematical backbone of quantum mechanics, where physical observables (position, momentum, energy) are modeled as self-adjoint operators precisely so that measured values — the eigenvalues — come out real, matching physical experience. In multivariable calculus and optimization, the Hessian matrix of a twice-differentiable real function is symmetric, and its eigenvalues (guaranteed real by this theorem) determine whether a critical point is a local minimum, maximum, or saddle. In statistics and data science, covariance matrices are symmetric and positive semidefinite, and principal component analysis relies on the orthogonal eigenspace decomposition guaranteed here to find uncorrelated directions of maximal variance. In numerical linear algebra and engineering, symmetric matrices arising from finite-element discretizations of physical systems (vibration modes, stress tensors) are diagonalized using exactly the orthogonality of eigenspaces proved above, and specialized fast algorithms (Jacobi eigenvalue algorithm, Lanczos iteration) exploit self-adjointness for efficiency and numerical stability. In graph theory, the (symmetric) adjacency and Laplacian matrices of undirected graphs are self-adjoint, and their real eigenvalues encode connectivity, expansion, and clustering properties studied in spectral graph theory.

Ask the tutor: How does the Spectral Theorem use these three facts to build a full orthonormal eigenbasis for a self-adjoint operator?

DefinitionSelf-Adjoint 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 linear operator satisfying

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

The operator TT is called self-adjoint (or, when F=R\mathbb{F} = \mathbb{R}, symmetric, and when F=C\mathbb{F} = \mathbb{C}, Hermitian) if

T=T∗,T = T^*,

equivalently, if

⟨Tv,w⟩=⟨v,Tw⟩for all v,w∈V.\langle Tv, w \rangle = \langle v, Tw \rangle \qquad \text{for all } v, w \in V.

In matrix terms, relative to an orthonormal basis, TT is self-adjoint if and only if its matrix AA satisfies A=A∗A = A^*, where A∗=A‾TA^* = \overline{A}^T is the conjugate transpose (so A=ATA = A^T when F=R\mathbb{F} = \mathbb{R}).

Intuition

Think of the adjoint as the linear-algebra analogue of complex conjugation. A complex number zz satisfies z=zˉz = \bar z exactly when zz is real. The condition T=T∗T = T^* says TT is "fixed by conjugation" in the operator sense, and this is precisely what forces its eigenvalues to be real numbers rather than genuinely complex ones — even when the ambient scalar field is C\mathbb{C}. A second picture: the sesquilinear form B(v,w)=⟨Tv,w⟩B(v,w) = \langle Tv, w\rangle associated to any operator TT is, in general, not symmetric (B(v,w)≠B(w,v)‾B(v,w) \ne \overline{B(w,v)}). Self-adjointness is exactly the condition B(v,w)=B(w,v)‾B(v,w) = \overline{B(w,v)}, which makes the diagonal values B(v,v)=⟨Tv,v⟩B(v,v) = \langle Tv,v\rangle real. This is the seed of the "real quadratic form" fact proved below, and it is the reason symmetric matrices are the natural objects underlying quadratic forms in multivariable calculus (Hessians, conic sections, ellipsoids). Geometrically, distinct eigenvalues of a self-adjoint operator correspond to eigenspaces meeting at right angles. Picture a symmetric 2×22\times 2 matrix stretching the plane along two perpendicular axes by different factors — this orthogonal splitting is not a coincidence of the example, it is forced by the algebra, as the theorem below shows.

Figure
Eigenvectors of a self-adjoint operator are forced apart at 90°
v₁ (λ₁ = 3) v₂ (λ₂ = −1) 90° T self-adjoint on ℝ²: eigenvectors for λ₁ ≠ λ₂ meet orthogonally
TheoremSelf-Adjoint Operators: Real Eigenvalues, Real Quadratic Form, Orthogonal Eigenspaces

Let VV be a finite-dimensional inner product space over F\mathbb{F} (F=R\mathbb{F} = \mathbb{R} or C\mathbb{C}), with dim⁡V≥1\dim V \geq 1, and let T∈L(V)T \in \mathcal{L}(V) be self-adjoint, i.e., T=T∗T = T^*. Then: 1. (Real quadratic form.) For every v∈Vv \in V, the scalar ⟨Tv,v⟩\langle Tv, v\rangle is a real number. 2. (Real eigenvalues.) Every eigenvalue λ\lambda of TT is real. (This is automatic and non-vacuous when F=C\mathbb{F}=\mathbb{C}; when F=R\mathbb{F} = \mathbb{R} it says every eigenvalue that exists is real, which is trivial, but the same computation is the engine of part 3.) 3. (Orthogonality of eigenspaces.) If λ1≠λ2\lambda_1 \ne \lambda_2 are distinct eigenvalues of TT with eigenvectors v1,v2v_1, v_2 respectively (i.e. Tv1=λ1v1Tv_1 = \lambda_1 v_1, Tv2=λ2v2Tv_2 = \lambda_2 v_2, v1,v2≠0v_1, v_2 \ne 0), then ⟨v1,v2⟩=0\langle v_1, v_2 \rangle = 0.

Proof

We use only the defining property of the adjoint and the axioms of the inner product (conjugate symmetry ⟨x,y⟩=⟨y,x⟩‾\langle x,y\rangle = \overline{\langle y,x\rangle}, and linearity in the first slot, conjugate-linearity in the second, following the The Adjoint of a Linear Map convention). No result from later theory (spectral theorem, diagonalizability) is used, so there is no circularity.

Part 1: the quadratic form is real. Fix v∈Vv \in V. By conjugate symmetry of the inner product,

⟨Tv,v⟩=⟨v,Tv⟩‾.\langle Tv, v \rangle = \overline{\langle v, Tv \rangle}.

By self-adjointness, ⟨v,Tv⟩=⟨T∗v,v⟩\langle v, Tv\rangle = \langle T^*v, v\rangle... let us be careful and use the defining identity directly. The adjoint satisfies ⟨Tx,y⟩=⟨x,T∗y⟩\langle Tx, y\rangle = \langle x, T^*y\rangle for all x,yx,y. Apply this with x=vx = v, y=vy=v:

⟨Tv,v⟩=⟨v,T∗v⟩=⟨v,Tv⟩,\langle Tv, v \rangle = \langle v, T^*v \rangle = \langle v, Tv \rangle,

using T∗=TT^* = T in the last step. Combining this with the conjugate-symmetry identity above,

⟨Tv,v⟩=⟨v,Tv⟩‾=⟨Tv,v⟩‾.\langle Tv, v \rangle = \overline{\langle v, Tv \rangle} = \overline{\langle Tv, v\rangle}.

A complex number equal to its own conjugate is real (if z=zˉz = \bar z, write z=a+biz = a+bi; then a+bi=a−bia+bi = a-bi forces 2bi=02bi=0, so b=0b=0). Hence ⟨Tv,v⟩∈R\langle Tv, v\rangle \in \mathbb{R}. Since v∈Vv \in V was arbitrary, this holds for every v∈Vv \in V. ■\blacksquare (Part 1)

Part 2: eigenvalues are real. Let λ\lambda be an eigenvalue of TT, with eigenvector v≠0v \ne 0, so Tv=λvTv = \lambda v. Consider ⟨Tv,v⟩\langle Tv, v \rangle. On one hand, by Part 1, this quantity is real. On the other hand, substituting Tv=λvTv = \lambda v and using linearity of the inner product in the first argument,

⟨Tv,v⟩=⟨λv,v⟩=λ⟨v,v⟩.\langle Tv, v \rangle = \langle \lambda v, v \rangle = \lambda \langle v, v \rangle.

Since v≠0v \ne 0, we have ⟨v,v⟩=∥v∥2>0\langle v, v\rangle = \|v\|^2 > 0, a positive real number. Therefore

λ=⟨Tv,v⟩⟨v,v⟩\lambda = \frac{\langle Tv, v\rangle}{\langle v, v\rangle}

is a ratio of a real number to a positive real number, hence real. ■\blacksquare (Part 2)

Part 3: eigenspaces for distinct eigenvalues are orthogonal. Let λ1≠λ2\lambda_1 \ne \lambda_2 be eigenvalues of TT with eigenvectors v1,v2v_1, v_2 (both nonzero), Tv1=λ1v1Tv_1 = \lambda_1 v_1, Tv2=λ2v2Tv_2 = \lambda_2 v_2. By Part 2, λ1,λ2∈R\lambda_1, \lambda_2 \in \mathbb{R}; this will be used to discard a conjugation below. We compute ⟨Tv1,v2⟩\langle Tv_1, v_2\rangle two ways. First way, using Tv1=λ1v1Tv_1 = \lambda_1 v_1 and linearity in the first slot:

⟨Tv1,v2⟩=⟨λ1v1,v2⟩=λ1⟨v1,v2⟩.\langle Tv_1, v_2 \rangle = \langle \lambda_1 v_1, v_2 \rangle = \lambda_1 \langle v_1, v_2 \rangle.

Second way, using the defining property of the adjoint and T∗=TT^*=T:

⟨Tv1,v2⟩=⟨v1,T∗v2⟩=⟨v1,Tv2⟩=⟨v1,λ2v2⟩=λ2‾ ⟨v1,v2⟩,\langle Tv_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,

where the last equality uses conjugate-linearity of the inner product in the second slot. Since λ2\lambda_2 is real by Part 2, λ2‾=λ2\overline{\lambda_2} = \lambda_2, so this simplifies to

⟨Tv1,v2⟩=λ2⟨v1,v2⟩.\langle Tv_1, v_2\rangle = \lambda_2 \langle v_1, v_2 \rangle.

Equating the two computations of ⟨Tv1,v2⟩\langle Tv_1, v_2\rangle:

λ1⟨v1,v2⟩=λ2⟨v1,v2⟩,\lambda_1 \langle v_1, v_2 \rangle = \lambda_2 \langle v_1, v_2 \rangle,

so

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

Since λ1≠λ2\lambda_1 \ne \lambda_2, we have λ1−λ2≠0\lambda_1 - \lambda_2 \ne 0, and since we are in a field with no zero divisors, we may divide, giving

⟨v1,v2⟩=0.\langle v_1, v_2 \rangle = 0.

This proves v1⊥v2v_1 \perp v_2, as required. ■\blacksquare (Part 3) All three parts follow from a single computational technique — moving TT across the inner product via its defining adjoint identity and comparing the result to the same expression computed via the eigenvalue equation — applied with the hypothesis T=T∗T=T^* supplying exactly the extra equality needed to close each argument. ■\blacksquare

∎
CorollaryOrthogonal Decomposition of Eigenspaces

Let T∈L(V)T \in \mathcal{L}(V) be self-adjoint and let λ1,…,λk\lambda_1, \dots, \lambda_k be the distinct eigenvalues of TT, with eigenspaces Eλ1,…,EλkE_{\lambda_1}, \dots, E_{\lambda_k} (where Eλ={v∈V:Tv=λv}E_\lambda = \{v \in V : Tv = \lambda v\}). Then the eigenspaces are pairwise orthogonal:

Eλi⊥Eλjfor i≠j,E_{\lambda_i} \perp E_{\lambda_j} \quad \text{for } i \ne j,

meaning ⟨u,w⟩=0\langle u, w\rangle = 0 for all u∈Eλiu \in E_{\lambda_i}, w∈Eλjw \in E_{\lambda_j}.

Proof

Take any nonzero u∈Eλiu \in E_{\lambda_i} and nonzero w∈Eλjw \in E_{\lambda_j} with i≠ji \ne j (if either is 00 the inner product is trivially 00). Then Tu=λiuTu = \lambda_i u and Tw=λjwTw = \lambda_j w with λi≠λj\lambda_i \ne \lambda_j, so Self-Adjoint Operators Part 3 (proved above) gives ⟨u,w⟩=0\langle u, w \rangle = 0 directly, applied to this pair of eigenvectors. ■\blacksquare

∎

Worked examples

Example 1: a genuinely asymmetric-looking 2×22\times 2 symmetric matrix. Let V=R2V = \mathbb{R}^2 with the standard inner product, and let TT have matrix

A=(5222)A = \begin{pmatrix} 5 & 2 \\ 2 & 2 \end{pmatrix}

relative to the standard orthonormal basis. Since AT=AA^T = A, TT is self-adjoint (symmetric). Eigenvalues. The characteristic polynomial is

det⁡(A−λI)=(5−λ)(2−λ)−4=λ2−7λ+10−4=λ2−7λ+6=(λ−6)(λ−1).\det(A - \lambda I) = (5-\lambda)(2-\lambda) - 4 = \lambda^2 - 7\lambda + 10 - 4 = \lambda^2 - 7\lambda + 6 = (\lambda - 6)(\lambda - 1).

So λ1=6\lambda_1 = 6, λ2=1\lambda_2 = 1: both real, consistent with Part 2 of the theorem (and here F=R\mathbb{F}=\mathbb{R} so this direction is not surprising by itself, but the numbers set up the orthogonality check, which is the substantive fact). Eigenvectors. For λ1=6\lambda_1 = 6: solve (A−6I)v=0(A - 6I)v = 0,

(−122−4)v=0  ⟹  −v1+2v2=0  ⟹  v1=2v2.\begin{pmatrix} -1 & 2 \\ 2 & -4 \end{pmatrix} v = 0 \implies -v_1 + 2v_2 = 0 \implies v_1 = 2v_2.

Take v1vec=(2,1)v_1^{\text{vec}} = (2,1). For λ2=1\lambda_2 = 1: solve (A−I)v=0(A-I)v=0,

(4221)v=0  ⟹  4v1+2v2=0  ⟹  v2=−2v1.\begin{pmatrix} 4 & 2 \\ 2 & 1 \end{pmatrix} v = 0 \implies 4v_1 + 2v_2 = 0 \implies v_2 = -2v_1.

Take v2vec=(1,−2)v_2^{\text{vec}} = (1,-2). Check orthogonality (Part 3).

⟨(2,1),(1,−2)⟩=2⋅1+1⋅(−2)=2−2=0.\langle (2,1), (1,-2)\rangle = 2\cdot 1 + 1\cdot(-2) = 2 - 2 = 0.

Indeed orthogonal, exactly as guaranteed since λ1≠λ2\lambda_1 \ne \lambda_2. Check the quadratic form is real (trivial over R\mathbb{R}, but instructive numerically). Take v=(1,3)v = (1,3):

Tv=A(13)=(5+62+6)=(118),⟨Tv,v⟩=11(1)+8(3)=11+24=35∈R.Tv = A\begin{pmatrix}1\\3\end{pmatrix} = \begin{pmatrix}5+6\\2+6\end{pmatrix} = \begin{pmatrix}11\\8\end{pmatrix}, \qquad \langle Tv, v\rangle = 11(1) + 8(3) = 11+24 = 35 \in \mathbb{R}.

Example 2: a genuinely complex Hermitian matrix, where reality of eigenvalues is the substantive fact. Let V=C2V = \mathbb{C}^2 with the standard Hermitian inner product ⟨x,y⟩=x1y1‾+x2y2‾\langle x,y\rangle = x_1\overline{y_1} + x_2\overline{y_2}, and let TT have matrix

A=(32−i2+i4).A = \begin{pmatrix} 3 & 2-i \\ 2+i & 4 \end{pmatrix}.

Check self-adjointness: A∗=A‾TA^* = \overline{A}^T. Transposing AA gives (32+i2−i4)\begin{pmatrix}3 & 2+i \\ 2-i & 4\end{pmatrix}, and conjugating entrywise gives (32−i2+i4)=A\begin{pmatrix}3 & 2-i\\ 2+i & 4\end{pmatrix} = A. So A∗=AA^* = A: TT is Hermitian, hence self-adjoint. Eigenvalues. The characteristic polynomial is

det⁡(A−λI)=(3−λ)(4−λ)−(2−i)(2+i).\det(A-\lambda I) = (3-\lambda)(4-\lambda) - (2-i)(2+i).

Compute (2−i)(2+i)=4−i2=4+1=5(2-i)(2+i) = 4 - i^2 = 4+1 = 5. So

det⁡(A−λI)=(3−λ)(4−λ)−5=λ2−7λ+12−5=λ2−7λ+7.\det(A-\lambda I) = (3-\lambda)(4-\lambda) - 5 = \lambda^2 - 7\lambda + 12 - 5 = \lambda^2 - 7\lambda + 7.

By the quadratic formula,

λ=7±49−282=7±212.\lambda = \frac{7 \pm \sqrt{49 - 28}}{2} = \frac{7\pm\sqrt{21}}{2}.

Both roots are real numbers (21≈4.583\sqrt{21} \approx 4.583, giving λ1≈5.7915\lambda_1 \approx 5.7915 and λ2≈1.2085\lambda_2 \approx 1.2085), even though the matrix has genuinely nonreal off-diagonal entries. This is exactly Part 2 of the theorem doing real work: a matrix with complex entries is forced to have a real spectrum because it is Hermitian. Orthogonality check, symbolically. Let v1=(2−i,λ1−3)v_1 = (2-i, \lambda_1 - 3) be an eigenvector for λ1\lambda_1 (solving the first row of (A−λ1I)v=0(A-\lambda_1 I)v=0: (3−λ1)v1(1)+(2−i)v1(2)=0⇒v1(1)=−2−i3−λ1v1(2)(3-\lambda_1)v_1^{(1)} + (2-i)v_1^{(2)} = 0 \Rightarrow v_1^{(1)} = -\frac{2-i}{3-\lambda_1}v_1^{(2)}; scaling by −(3−λ1)-(3-\lambda_1) gives the coordinate form above up to sign, but let's instead verify orthogonality using the trace/determinant relations, which is cleaner). Since AA is Hermitian with distinct eigenvalues λ1≠λ2\lambda_1 \ne \lambda_2 (as 21≠0\sqrt{21}\ne 0), Self-Adjoint Operators Part 3 guarantees the corresponding eigenvectors v1,v2v_1, v_2 satisfy ⟨v1,v2⟩=0\langle v_1, v_2\rangle = 0 automatically — this is the payoff of the theorem: we do not need to grind through the messy complex arithmetic to know the eigenvectors will be orthogonal; the self-adjointness of AA guarantees it in advance. (A direct computation, substituting the exact surd values of λ1,λ2\lambda_1,\lambda_2 and simplifying using λ1+λ2=7\lambda_1+\lambda_2 = 7, λ1λ2=7\lambda_1\lambda_2 = 7, confirms ⟨v1,v2⟩=0\langle v_1,v_2\rangle = 0, but the theorem renders this computation unnecessary.)

CounterexampleDropping self-adjointness breaks all three conclusions

Consider V=R2V = \mathbb{R}^2 with the standard inner product, and let TT have matrix

A=(01−10),A = \begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix},

a rotation by 90°90°. This TT is not self-adjoint: AT=(0−110)≠AA^T = \begin{pmatrix}0 & -1\\1 & 0\end{pmatrix} \ne A. Eigenvalues fail to be real. The characteristic polynomial is λ2−(tr A)λ+det⁡A=λ2+1\lambda^2 - (\mathrm{tr}\,A)\lambda + \det A = \lambda^2 + 1, with roots λ=±i\lambda = \pm i — not real at all, contradicting what Part 2 would have given had TT been self-adjoint. (Since F=R\mathbb{F}=\mathbb{R} here, TT has no real eigenvalues whatsoever: geometrically, a rotation by 90°90° fixes no line through the origin, and the algebra reflects this.) The quadratic form need not be real-valued in the complexified sense, and more importantly the symmetric identity fails outright. Consider the associated bilinear form B(v,w)=⟨Tv,w⟩B(v,w) = \langle Tv,w\rangle on R2\mathbb{R}^2; take v=(1,0)v=(1,0), w=(0,1)w=(0,1):

B(v,w)=⟨T(1,0),(0,1)⟩=⟨(0,−1),(0,1)⟩=−1,B(w,v)=⟨T(0,1),(1,0)⟩=⟨(1,0),(1,0)⟩=1.B(v,w) = \langle T(1,0),(0,1)\rangle = \langle(0,-1),(0,1)\rangle = -1, \qquad B(w,v) = \langle T(0,1),(1,0)\rangle = \langle(1,0),(1,0)\rangle = 1.

Since B(v,w)≠B(w,v)B(v,w) \ne B(w,v), the form is not symmetric — exactly the failure mode that self-adjointness was designed to rule out. (Over R\mathbb{R} "real-valued" is automatic, so the substantive breakage is the loss of symmetry B(v,w)=B(w,v)B(v,w)=B(w,v), which is what powered Part 1's proof in the Hermitian case.) Orthogonality of eigenvectors fails / is not even meaningful over R\mathbb{R}. Since TT has no real eigenvectors at all, the orthogonality conclusion of Part 3 cannot even be stated over R\mathbb{R}; if we complexify and use eigenvectors v±i=(1,±i)v_{\pm i} = (1, \pm i) for eigenvalues ±i\pm i in C2\mathbb{C}^2, one computes with the Hermitian inner product ⟨x,y⟩=x1yˉ1+x2yˉ2\langle x,y\rangle = x_1\bar y_1+x_2\bar y_2:

⟨(1,i),(1,−i)⟩=1⋅1‾+i⋅(−i)‾=1+i⋅i=1+i2=1−1=0.\langle (1,i), (1,-i)\rangle = 1\cdot\overline{1} + i\cdot\overline{(-i)} = 1 + i\cdot i = 1 + i^2 = 1 - 1 = 0.

Interesting — these happen to be orthogonal here because TT is a normal (not self-adjoint) operator, a preview of Normal Operators where a strictly weaker hypothesis than self-adjointness (TT∗=T∗TTT^*=T^*T) still yields orthogonal eigenspaces, but real eigenvalues are genuinely lost. The essential failure exhibited by this example is the loss of real eigenvalues and of the symmetric real quadratic form, both direct consequences of removing T=T∗T=T^*.

Scope, edge cases & common traps

SituationWhat holdsWhat can fail
F=R\mathbb{F}=\mathbb{R}, T=TTT=T^T (symmetric)Real quadratic form (trivial), eigenspaces orthogonal for distinct λ\lambdaExistence of eigenvalues at all requires the full Spectral Theorem, not proved here
F=C\mathbb{F}=\mathbb{C}, T=T∗T=T^* (Hermitian)All eigenvalues real, quadratic form real, eigenspaces orthogonalIf you forget conjugate-linearity in the second slot, the orthogonality proof breaks
TT normal but not self-adjointEigenspaces for distinct eigenvalues still orthogonal (see Normal Operators)Eigenvalues need not be real (Counterexample above)
dim⁡V=∞\dim V = \inftyStatement requires care (need TT densely defined, bounded/self-adjoint distinctions)This article's proofs assume finite dimension implicitly via existence of eigenvectors; infinite-dimensional self-adjointness is a separate, subtler theory
Repeated eigenvaluesTheorem says nothing about eigenvectors within one eigenspace — they need not be orthogonal to each other automatically, only across distinct eigenspacesA common mistake is assuming self-adjointness alone orthogonalizes a chosen basis of a single eigenspace; you must apply Gram–Schmidt within the eigenspace separately

The mathematician behind it

Origins · 1 minDavid Hilbert and how this was discovered
DH
David Hilbert
1862–1943 · Göttingen, foundational work on integral equations and spectral theory

Hilbert's early-1900s study of integral equations led him to consider symmetric bilinear forms in infinitely many variables, the seed of what became Hilbert space theory. His student circle and successors (Schmidt, von Neumann) turned "self-adjoint operator" into the central object of spectral theory, later providing the rigorous mathematical language for quantum mechanics's observables.

History
- 1904 — Hilbert publishes foundational work on integral equations, introducing spectral ideas for symmetric kernels.
- 1908 — Erhard Schmidt develops geometric (orthogonality-based) methods for symmetric operators, sharpening Hilbert's analysis.
- 1927 — Hilbert and collaborators connect operator spectra to physical spectra, anticipating quantum-mechanical observables.
- 1929 — John von Neumann gives the rigorous Hilbert-space definition of self-adjoint operator, resolving subtleties invisible in finite dimensions.
- 1932 — Von Neumann's Mathematical Foundations of Quantum Mechanics cements self-adjoint operators as the mathematical model for physical observables.
Is every symmetric real matrix diagonalizable?

Yes, but that stronger fact — full diagonalizability by an orthonormal eigenbasis — is the content of The Spectral Theorem for Self-Adjoint Operators, a separate and deeper theorem built on top of what is proved here. This article establishes three ingredients used inside that proof: eigenvalues are real, the associated quadratic form is real, and eigenspaces for distinct eigenvalues are orthogonal. The Spectral Theorem additionally needs an existence argument (that enough eigenvectors exist to span the whole space), typically via induction on dimension using the existence of at least one eigenvalue plus the orthogonal-complement structure. So yes, but the full diagonalizability claim is not proved in this article.

Does T = T* require choosing an orthonormal basis to check?

No — the definition T=T∗T = T^* is basis-free, since the adjoint itself is defined intrinsically via ⟨Tv,w⟩=⟨v,T∗w⟩\langle Tv,w\rangle = \langle v, T^*w\rangle for all v,wv, w, independent of any coordinate choice (see The Adjoint of a Linear Map). However, to compute T∗T^* concretely as a matrix, it is most convenient to work in an orthonormal basis, where the adjoint's matrix is simply the conjugate transpose of TT's matrix. In a non-orthonormal basis, the matrix of T∗T^* is more complicated to write down, though the operator T∗T^* itself, and the truth of T=T∗T=T^*, does not depend on this choice.

Why does the proof need conjugate-linearity in the second slot of the inner product?

Over C\mathbb{C}, the standard convention (as in Axler) is that ⟨⋅,⋅⟩\langle \cdot,\cdot\rangle is linear in the first argument and conjugate-linear in the second: ⟨v,cw⟩=cˉ⟨v,w⟩\langle v, cw\rangle = \bar c\langle v,w\rangle. This convention is exactly what produces the crucial conjugate λ2‾\overline{\lambda_2} in the proof of Part 3, which then gets converted to λ2\lambda_2 using Part 2 (reality of eigenvalues). Without this conjugate-linearity convention (or its mirror image with the roles swapped), the algebra rearranges slightly but the same conjugate always appears somewhere, and reality of the eigenvalue is always what discharges it — this is the precise mechanism linking Parts 2 and 3.

Is the converse true: if all eigenvalues of T are real, must T be self-adjoint?

No, this converse is false, and it is a classic trap. A real 2×22\times 2 matrix with two distinct real eigenvalues need not be symmetric — for instance (2103)\begin{pmatrix}2&1\\0&3\end{pmatrix} has real eigenvalues 2,32,3 but is not symmetric, and its eigenvectors are not orthogonal. Real eigenvalues are a necessary consequence of self-adjointness proved in this article, not a sufficient condition for it. The correct sufficient-and-necessary characterization of self-adjointness in terms of spectral data requires the full Spectral Theorem machinery (real eigenvalues and an orthonormal eigenbasis spanning the whole space, for the operator itself, not merely its restriction).

How does this relate to normal operators?

Every self-adjoint operator is normal, meaning it commutes with its own adjoint (TT∗=T∗TTT^* = T^*T, which holds trivially when T=T∗T=T^* since then TT∗=T2=T∗TTT^*=T^2=T^*T). Normal operators form a strictly larger class that still enjoys orthogonality of eigenspaces for distinct eigenvalues (proved by a nearly identical argument using ∥Tv∥=∥T∗v∥\|Tv\|=\|T^*v\|), but normal operators can have non-real eigenvalues, as the rotation-matrix counterexample above illustrates once complexified. Normal Operators is the natural next topic generalizing the orthogonality half of this theorem while explicitly giving up the real-eigenvalue guarantee.

What goes wrong in infinite dimensions?

In infinite-dimensional Hilbert spaces, unbounded self-adjoint operators (like the position or momentum operator in quantum mechanics) require a much more careful definition distinguishing "symmetric" (⟨Tv,w⟩=⟨v,Tw⟩\langle Tv,w\rangle=\langle v,Tw\rangle on a dense domain) from the strictly stronger "self-adjoint" (domain of TT equals domain of T∗T^* exactly). This article's proofs implicitly use finite-dimensionality when they assume eigenvectors exist and that all algebraic manipulations are unconditioned by domain issues. The three conclusions (real eigenvalues, real quadratic form, orthogonal eigenspaces) do generalize to the honestly-defined infinite-dimensional self-adjoint case, but proving existence of eigenvalues at all becomes a genuinely different, harder problem (spectral theory replaces finite spectra with more general spectra).

Before this

Where this leads