Adjoint Operators and Hermitian Operators

Self-adjointness is the daily working assumption of several live fields. In quantum computing every measurement reads out a Hermitian observable, and the flagship near-term algorithm — the variational quantum eigensolver — estimates the lowest eigenvalue of a Hermitian Hamiltonian, leaning on exactly the real-spectrum guarantee proved above. In data science the covariance matrix is real symmetric, so principal-component analysis is the spectral decomposition of Example 1 scaled to thousands of dimensions: orthogonal eigenvectors are the principal directions, real eigenvalues the variances. Engineers run the same computation as modal analysis, extracting real natural frequencies and orthogonal normal modes from a stiffness matrix. At the frontier it gets subtler: mathematical physicists still prove essential self-adjointness theorems (Kato-Rellich) to certify that a Schrödinger Hamiltonian with a singular $-1/r$ Coulomb potential is a legitimate observable — precisely the symmetric-versus-self-adjoint gap of the warning box. And a growing community deliberately breaks Hermiticity: PT-symmetric and open-system Hamiltonians are non-Hermitian yet can have real spectra, and their exceptional points now drive experiments in photonics and sensing.

Picture the inner product as a two-pan balance. The identity $\langle A^\dagger x,y\rangle=\langle x,Ay\rangle$ says that whatever $A$ does on the right vector, the adjoint $A^\dagger$ does the equivalent on the left, so the scalar reading never moves — $A^\dagger$ keeps the books balanced when you carry $A$ across the bracket. In an orthonormal basis this has a mechanical face, transpose and conjugate, but the bracket definition is primary because it survives into infinite dimensions where matrices no longer make sense.

A Hermitian operator looks the same from either pan: $\langle Ax,y\rangle=\langle x,Ay\rangle$. That symmetry is rigid — it pins every eigenvalue onto the real line, since a complex eigenvalue would unbalance $\langle Ax,x\rangle=\overline{\langle Ax,x\rangle}$, and it forces distinct eigenspaces perpendicular. Geometrically the spectral theorem then says a Hermitian operator does nothing but stretch space along mutually orthogonal axes by real factors — no twist, no shear. Choose those axes as your basis and the operator is a diagonal real matrix. The second figure makes this concrete with a $2\times2$ real symmetric operator, eigenvalues $3$ and $1$.

Everything begins with one identity. The adjoint is not built from a formula; it is defined by the requirement that it reproduce $A$ on the opposite side of the inner product, for every pair of vectors. The physics convention makes the inner product antilinear in its first slot, threading the conjugations through the results below.

where: $A$ — a linear operator on an inner-product space $V$; $A^\dagger$ — its adjoint, the unique operator satisfying this identity; $\langle\cdot,\cdot\rangle$ — the inner product, antilinear in the first slot; $x,y$ — arbitrary vectors in $V$. Choose an orthonormal basis and the abstract adjoint collapses to something computable by hand: the matrix of $A^\dagger$ is that of $A$ transposed and conjugated — the conjugate transpose, also written $A^\ast$ or $A^H$.

where: $A_{ij}=\langle e_i,Ae_j\rangle$ — matrix elements in an orthonormal basis $\{e_i\}$; $\overline{(\cdot)}$ — complex conjugation; ${\mathsf T}$ — transpose. Transpose swaps the indices; conjugation handles the antilinearity of the first slot. The adjoint respects the operator algebra, with two twists: it reverses product order and conjugates scalars. These four rules are all you need to manipulate daggers fluently.

where: $\alpha$ — a complex scalar; $A,B$ — linear operators. The order reversal $(AB)^\dagger=B^\dagger A^\dagger$ mirrors $(AB)^{\mathsf T}=B^{\mathsf T}A^{\mathsf T}$ and causes many order slips. The central special case is the operator that equals its own adjoint — the symmetric condition below, in which $A$ moves freely from slot to slot at no cost.

where: $A$ — the Hermitian (self-adjoint) operator. In an orthonormal basis $A=\bar A^{\mathsf T}$: the matrix is invariant under conjugate transpose, so diagonal entries are real and off-diagonals are conjugate mirror images. Self-adjointness immediately constrains the spectrum: the eigenvalues of a Hermitian operator cannot be complex — they must be real, which is what makes them fit to be measured physical quantities.

where: $\lambda$ — an eigenvalue of $A$; $x$ — the corresponding nonzero eigenvector. The proof is a two-line consequence of (1) and (4); see Derivation A. Collecting the eigenvalues and their eigenspaces gives the spectral decomposition: a Hermitian operator is a real-weighted sum of orthogonal projectors that resolve the identity.

where: $\lambda_k$ — the distinct real eigenvalues; $P_k$ — the orthogonal projector onto the eigenspace of $\lambda_k$, with $P_k=P_k^\dagger=P_k^2$; $\mathbb 1$ — the identity, so $\sum_k P_k=\mathbb 1$ resolves the identity. Equivalently, stacking the orthonormal eigenvectors as columns of a unitary matrix diagonalizes $A$ — the matrix face of the spectral theorem, and the form used to compute functions of $A$.

where: $U$ — the unitary whose columns are the orthonormal eigenvectors; $\Lambda$ — the diagonal matrix of real eigenvalues. Since $U^\dagger=U^{-1}$, this similarity transformation also preserves the inner product. Finally, the Hermitian and unitary families are welded by the exponential map: a Hermitian generator exponentiates to a unitary, its real eigenvalues becoming phases on the unit circle.

where: $A$ — a Hermitian generator; $U$ — the resulting unitary, eigenvalues $e^{i\lambda_k}$ of unit modulus. In quantum mechanics $A=-Ht/\hbar$ turns the Hamiltonian into the unitary time-evolution operator.

The single relation "compare $A^\dagger$ with $A$" specializes to every named operator class in linear algebra — each row a different answer, each with its own spectral fingerprint.

1 · Diagonalizing a real symmetric operator (fully worked)

Take the real symmetric matrix with rows $(2,1)$ and $(1,2)$. Find its eigenvalues and orthonormal eigenbasis, build the diagonalizing unitary $U$, and write its spectral decomposition $A=\sum_k\lambda_k P_k$.

Step 1. The characteristic equation is $\det(A-\lambda\mathbb 1)=(2-\lambda)^2-1=\lambda^2-4\lambda+3=0$, so $(\lambda-3)(\lambda-1)=0$ and the eigenvalues are $\boldsymbol{\lambda_1=3}$, $\boldsymbol{\lambda_2=1}$ — both real, as Derivation A guarantees for a symmetric matrix.

Step 2. For $\lambda_1=3$, $(A-3\mathbb 1)v=0$ has row $(-1,1)$, giving $v_1=v_2$, so $\boldsymbol{u_1=(1,1)/\sqrt2}$. For $\lambda_2=1$, the row $(1,1)$ gives $v_1=-v_2$, so $\boldsymbol{u_2=(1,-1)/\sqrt2}$. Their overlap $\langle u_1,u_2\rangle=\tfrac12(1-1)=0$ — orthogonal, as required for distinct eigenvalues.

Step 3. Stack the eigenvectors as columns to form the (orthogonal, hence unitary) matrix $U=\tfrac1{\sqrt2}\begin{pmatrix}1&1\\1&-1\end{pmatrix}$. Then $U^{\mathsf T}U=\mathbb 1$ and $U^{\mathsf T}AU=\operatorname{diag}(3,1)=\Lambda$, so $\boldsymbol{A=U\Lambda U^{\mathsf T}}$.

Step 4. The spectral projectors are the outer products $P_1=u_1u_1^{\mathsf T}=\tfrac12\begin{pmatrix}1&1\\1&1\end{pmatrix}$ and $P_2=u_2u_2^{\mathsf T}=\tfrac12\begin{pmatrix}1&-1\\-1&1\end{pmatrix}$. Check the identity resolution: $P_1+P_2=\mathbb 1$ and $3P_1+P_2=\begin{pmatrix}2&1\\1&2\end{pmatrix}=A$. The decomposition $A=3P_1+P_2$ reproduces $A$ exactly — geometrically a stretch by $3$ along $(1,1)$ with the perpendicular $(1,-1)$ direction left at scale $1$, the ellipse of the second figure.

2 · A genuinely complex Hermitian operator (fully worked)

Let $A$ have rows $(1,\,i)$ and $(-i,\,1)$. Verify it is Hermitian and find its real eigenvalues and orthogonal eigenvectors.

Step 1. Form the conjugate transpose: conjugating gives rows $(1,-i)$ and $(i,1)$, and transposing returns rows $(1,i)$ and $(-i,1)$ — identical to $A$, so $\boldsymbol{A^\dagger=A}$. The diagonal is real and the off-diagonal pair are complex conjugates, the signature of a Hermitian matrix.

Step 2. The characteristic equation is $\det(A-\lambda\mathbb 1)=(1-\lambda)^2-(i)(-i)=(1-\lambda)^2-1=0$, since $(i)(-i)=1$. So $(1-\lambda)^2=1$, giving $\boldsymbol{\lambda=2}$ and $\boldsymbol{\lambda=0}$ — real, despite the complex entries.

Step 3. For $\lambda=2$: the row $(-1,i)$ gives $v_1=iv_2$, so $\boldsymbol{u_1=(i,1)/\sqrt2}$. For $\lambda=0$: the row $(1,i)$ gives $v_1=-iv_2$, so $\boldsymbol{u_2=(-i,1)/\sqrt2}$.

Step 4. Test orthogonality with the antilinear first slot: $\langle u_1,u_2\rangle=\tfrac12\big(\overline{i}\cdot(-i)+\overline{1}\cdot1\big)=\tfrac12\big((-i)(-i)+1\big)=\tfrac12(i^2+1)=0$. The eigenvectors are orthogonal and the spectrum $\{0,2\}$ is real — the two Hermitian guarantees, on an operator with honestly complex entries. (Sanity check: $A=\mathbb 1+\sigma_y$, and $\sigma_y$ has eigenvalues $\pm1$, so $A$ has $1\pm1=0,2$.)

3 · Symmetric but not self-adjoint (a cautionary example)

Why is the momentum operator $p=-i\hbar\,\mathrm d/\mathrm dx$ on the interval $[0,L]$ a warning rather than a worked number?

Step 1. Integrate by parts: $\langle p\psi,\phi\rangle-\langle\psi,p\phi\rangle=-i\hbar\big[\overline{\psi}\phi\big]_0^L$, so the two sides of the bracket agree only when the boundary term vanishes. For a bare differential expression there is no matrix to conjugate-transpose — self-adjointness is decided at the boundary.

Step 2. On functions with $\psi(0)=\psi(L)=0$ (the naive box condition) the boundary term vanishes for all $\phi$, so $p$ is symmetric. But its adjoint $p^\dagger$ lives on a larger domain with no boundary condition, so $\mathcal D(p^\dagger)\neq\mathcal D(p)$ and $p$ is not self-adjoint — it has no real eigenvalues here.

Step 3. Impose instead $\psi(L)=e^{i\theta}\psi(0)$ for a fixed phase $\theta$. Now the boundary term cancels and the adjoint inherits the same condition, so $p$ is genuinely self-adjoint, with real spectrum $p_n=(2\pi n+\theta)\hbar/L$, $n\in\mathbb Z$. The lesson: the same expression is symmetric, non-self-adjoint, or self-adjoint depending entirely on the domain — a distinction that does not exist in finite dimensions.

The self-adjoint operator is where nineteenth-century algebra of forms meets twentieth-century functional analysis: a real eigenvalue Hermite found in a matrix became, in von Neumann's hands, the guarantee that a quantum measurement returns a real number.

Texts. Stone and Goldbart, Mathematics for Physics, linear-algebra chapter (adjoints, Hermitian operators, the spectral theorem) · Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed., ch. 5-6 (Hermitian matrices and eigenvalue problems) · Hassani, Mathematical Physics, ch. 4-6 (operators on inner-product spaces, spectral decomposition) · Byron and Fuller, Mathematics of Classical and Quantum Physics, ch. 3-5 · Reed and Simon, Methods of Modern Mathematical Physics I: Functional Analysis (self-adjointness and domains) · Sakurai and Napolitano, Modern Quantum Mechanics, ch. 1 (observables as Hermitian operators). Key checkpoints for the material in this article: Derivation A turns on the antilinearity of the first slot; Derivation B rests on the eigenvector over $\mathbb C$, invariance of the orthogonal complement, and $\sum_k P_k=\mathbb 1$; the non-orthonormal adjoint $A^\dagger=G^{-1}\bar A^{\mathsf T}G$ reduces to $\bar A^{\mathsf T}$ at $G=\mathbb 1$; Example 1 yields eigenvalues $3,1$ with $3P_1+P_2=A$; Example 2 is Hermitian with eigenvalues $0,2$ and $\langle u_1,u_2\rangle=0$.

Prerequisites. Inner-Product Spaces and Norms · Linear Operators and Matrix Representations · Eigenvalues, Eigenvectors, and the Characteristic Equation · Orthogonality and the Gram-Schmidt Process. Leads to. Diagonalization and the Spectral Decomposition · Unitary and Orthogonal Operators · Change of Basis and Similarity Transformations.

Ask the tutor: Ask me why the momentum operator on an interval is symmetric but not self-adjoint, or to diagonalize any Hermitian matrix you name step by step.